Following the Power

The Case for Microsoft at Soluna’s Project Kati 2

The bottleneck in the AI buildout is becoming increasingly clear: power.

The largest technology companies can spend billions on GPUs, land and new data centers, but several hundred megawatts of electricity cannot be ordered on the same schedule. Transmission takes time. Interconnection takes time. Substations and other high-voltage infrastructure take time. For the largest AI projects, access to power can determine not only where a campus is built, but how quickly it can be brought online.

In South Texas, a different problem already exists.

Large wind projects produce substantial amounts of electricity, yet generation and demand are not always located where the grid can connect them efficiently. Congestion, transmission constraints and periods of weak local pricing can reduce the value of power that is already being produced.

Soluna has spent years building around this problem. Its model places computing close to renewable generation, allowing electricity to be consumed behind the meter rather than relying entirely on the grid to transport that power somewhere else.

That history matters when looking at Project Kati 2.

Kati 2 is now being developed in Willacy County as a large AI and high-performance-computing campus. But it did not begin with a search for hundreds of megawatts for an AI tenant. The power asset was already there. Soluna’s relationship with the site had already been established. The commercial structure that made large-scale computing possible had already begun to take shape.

The project now being developed grew out of that earlier work.

To understand Kati 2, it is necessary to go back to Soluna’s original renewable-computing model, to Las Majadas, to EDF and Masdar, and to the process that brought Soluna to this particular site. From there, the project can be followed as it changes in scale and purpose, eventually becoming something far larger than what was originally contemplated there.

And somewhere along that path, Microsoft begins to appear.

Then it appears again.

And again.

What first looks like a loose connection starts to become a pattern.

Microsoft surfaces around the energy side of the story. It surfaces around the infrastructure side. It appears in the technical research. It appears through major energy relationships. It appears through people with direct experience building some of the world’s largest AI infrastructure. And as Kati 2 grows into a hyperscale-class project, several of those threads begin to move closer together.

That is the reason for this research.

The Microsoft thesis is not built around one clue. It is built around accumulation.

Different sources, different people, different organizations and different parts of the infrastructure stack repeatedly point in the same direction. Some of those connections begin years apart. Others develop almost on top of the Kati 2 timeline. Taken together, they create a much more interesting question than whether one isolated link can be explained away.

The question is how many independent overlaps can accumulate around the same project before coincidence stops being the most useful explanation.

To answer that, the story has to be rebuilt from the beginning.

Before Kati 2 became a 350 MW AI campus.

Before the current race for AI power.

Before Microsoft started appearing around the project from multiple directions.

The first question is simpler:

How did Soluna end up at this particular wind asset in South Texas, and why was it there in the first place?

1. Before Kati – The Model Came First

To understand Project Kati, it helps to begin with something that did not happen in Texas at all.

The first version of the idea was built around a wind project in Morocco.

A.M Wind came first

In 2009, Moroccan engineer Larbi Loudiyi partnered with German wind developer ALTUS AG to create A.M Wind SARL. A.M Wind spent the following years developing wind projects in Morocco, including what would later become Soluna’s flagship site in the country’s south. Soluna’s own 2018 white paper described the project as having been under development for nine years by the time Soluna arrived.

The site was unusually large. The original documents describe roughly 37,000 acres — about 15,000 hectares — with close to one gigawatt of potential wind capacity. Later SEC material identified the development as the Harmattan Project in Dakhla and described a planned 900 MW wind farm. A.M Wind had already commissioned wind-resource studies from Garrad Hassan and CUBE Engineering and had completed parts of the environmental and permitting work before Soluna took control.

Larbi is important to this story because he is not simply someone who joined Soluna years later.

He was there before Soluna.

He co-founded A.M Wind, helped develop the renewable asset that became the basis of the original Soluna concept, and then moved directly into Soluna’s energy organization after the acquisition. SEC material from 2021 describes him as the person who led design and engineering for the 900 MW Harmattan project and who was responsible for the architecture of Soluna’s flexible electrical designs for modular data centers in both Morocco and the United States.

That continuity is worth remembering. One of the people involved in building the renewable side of the concept in 2009 remained inside Soluna as the company later moved into U.S. data centers, behind-the-meter power and eventually Kati.

2018: Soluna joins the wind farm and the data center

The decisive change came in June 2018, when Soluna acquired A.M Wind from ALTUS. Soluna’s July 2018 white paper says the transaction gave it the exclusive rights to continue developing the Moroccan wind site. The same document describes Soluna as a Brookstone Partners portfolio company formed by combining renewable-energy assets with a new computing business.

That origin is important because the first Soluna was not conceived as a conventional data-center developer searching the market for inexpensive electricity.

The plan was much more integrated.

Soluna intended to develop the renewable generation and place high-density computing directly alongside it. Its 2018 white paper called the approach vertical integration: renewable power plants dedicated to on-site computing, with the energy system and computing system designed together rather than as separate businesses connected by an ordinary utility bill.

The language was written for the blockchain era, because cryptocurrency mining was the immediate commercial opportunity. But the engineering concept underneath it is striking when read from 2026.

The Moroccan site was to be developed in modular “Pods.” Each Pod would contain 12 MW of wind generation, an associated storage system and 6 MW of computing capacity. The full property was expected to support roughly 75 such Pods. The initial phase contemplated 36 MW of generation and 18 MW of computing.

More important than the exact numbers was how the system was supposed to behave.

Soluna designed the project to operate even without a conventional grid connection while retaining the ability to integrate with the grid later. Because wind production varies, the plan included energy storage — but storage was not expected to solve everything. The computing facilities themselves were to be designed to consume variable power.

That distinction sits close to the core of what Soluna still does today.

The load was not supposed to sit passively at the end of the electrical system and demand the same amount of electricity every second. The computing side was expected to respond to the characteristics of the generation.

Flexible load was part of the architecture from the beginning.

And the 2018 plan was not as narrowly tied to Bitcoin as it can appear in hindsight.

Soluna’s white paper explicitly said that cryptocurrency mining would only be the starting point. It described a future in which the same computing infrastructure could support distributed graphics rendering, file storage, machine learning, AI and other computing services.

That does not mean Soluna had an AI data-center business in 2018. It means something more useful for understanding the company’s development: the original concept was always broader than the first workload used to monetize it.

The asset was renewable power.

The infrastructure was modular computing.

The workload could change.

That distinction becomes central years later when the company begins moving from Bitcoin hosting toward AI and HPC.

The original architecture already looks familiar

Read without the cryptocurrency language, the 2018 design contained several ideas that remain recognizable in Soluna’s current infrastructure strategy:

renewable generation located beside compute;

modular data-center construction;

storage as part of the electrical design;

a computing load capable of responding to variable power;

large-scale development made possible by locating the computing infrastructure at the energy resource;

and the ability to change the type of computing performed as economics and technology evolved.

The technology and commercial structure would change substantially over the following eight years, but the central idea survived:

build the computing system around the characteristics of the power source rather than treating power as an unlimited input delivered from somewhere else.

There is also an important difference between the original Moroccan concept and what Soluna eventually built in the United States.

In Morocco, Soluna planned to own and develop the renewable generation itself. The company described itself as owning its own renewable-energy resources and vertically integrating the wind farm with its computing centers.

The U.S. model would eventually become more capital-efficient and, for the Kati story, much more consequential.

Instead of always having to build the renewable plant itself, Soluna could work with an existing renewable asset owner that already had generation, land and grid infrastructure — then bring a flexible computing load directly to that asset.

That shift did not happen overnight.

2020: the model moves into the United States

A second corporate story was developing in Albany, New York.

On January 8, 2020, Mechanical Technology Incorporated created a subsidiary called EcoChain to pursue cryptocurrency and blockchain computing. Five days later, on January 13, EcoChain entered into an operating and management agreement with Soluna Technologies. At the same time, Mechanical Technology invested $500,000 in Soluna.

The agreement was practical rather than theoretical.

Soluna was hired to help EcoChain create and operate a pilot cryptocurrency mining facility in North America. During the first months of 2020, Soluna prepared budgets, technical plans, operating models and a detailed business plan. In May, EcoChain acquired assets and intellectual property from the bankrupt GigaWatt mining operation in Washington State, and Soluna helped operate the resulting facility.

More U.S. site work followed.

In November 2020, EcoChain and Soluna signed another operating agreement for a potential mining location in the southeastern United States. A further agreement followed in December for another U.S. target; that site was eventually rejected because it did not satisfy the business requirements. Another southeastern target was added in February 2021. The SEC filings describe Soluna’s role as including project sourcing, acquisition negotiations, operating-model development, financing timelines and project-development paths.

This period is easy to overlook because it sits between the ambitious Moroccan plan and the later Soluna Holdings that investors recognize today.

But it was the bridge between them.

Soluna was learning how to take the energy-and-compute expertise developed around the Moroccan project and apply it to sites owned or financed by other parties in the United States.

The company no longer needed to wait for its own 900 MW wind farm to be completed before proving that the computing side worked.

It could source sites, integrate power and compute, operate facilities and build a development pipeline around existing energy infrastructure.

2021: Soluna becomes the public company’s identity

The relationship with EcoChain eventually turned into an acquisition.

On August 11, 2021, Mechanical Technology signed an agreement to acquire Soluna Computing. The transaction closed on October 29, 2021, and Mechanical Technology changed its corporate name to Soluna Holdings, Inc. effective November 2.

The legal structure was more complicated than the name change makes it sound.

The public company did not simply acquire the entire original Soluna organization including the Moroccan wind asset. SEC filings state that the transaction was designed to acquire substantially all of Soluna’s relevant assets other than the assets physically located in Morocco. What moved into the public company was primarily the U.S. development pipeline and the people and capabilities required to develop those projects.

That separation marks a major step in the evolution of the business model.

The first Soluna concept had been:

own renewable generation + build compute beside it.

The U.S. platform was increasingly becoming:

partner with renewable owners + build flexible compute beside their existing generation.

The acquisition announcement makes the transition unusually clear. Mechanical Technology described the combined company as a platform that would monetize wasted renewable-energy production from wind and solar farm owners. It said the acquired pipeline included approximately 300 MW of green-powered computing projects under letters of intent, with more than 200 MW of additional opportunities in the broader pipeline. John Belizaire framed the value proposition around helping renewable producers sell more of the electricity their assets could generate.

That is much closer to the company that would eventually sit across the table from EDF.

The renewable asset no longer had to belong to Soluna.

The asset owner had the wind farm.

Soluna could bring the load.

By 2022, the curtailment thesis was explicit

By the spring of 2022, Soluna had put a much more precise economic framework around the U.S. strategy.

On May 2, 2022, a white paper prepared by Isaac Maze-Rothstein in collaboration with Soluna described renewable curtailment as a growing economic problem and divided the solutions available to renewable developers into three practical categories:

transmission, battery storage and flexible computing.

The argument was straightforward.

Transmission can move more electricity away from a constrained renewable site, but expanding transmission is expensive and slow.

Batteries can shift electricity across time, but they add capital cost and have physical limits.

Flexible computing offers another possibility: place a controllable load at or near the renewable asset and consume electricity when it is abundant.

Soluna’s numbers in that paper estimated 14.9 TWh of U.S. wind and solar generation had been curtailed in 2021, representing roughly $610 million of lost revenue under its methodology. The paper also noted that curtailment was increasingly appearing in Texas and the Southwest rather than remaining concentrated in only a few historical markets.

The company’s response to the White House’s digital-assets inquiry a week later made the architecture even clearer. Soluna described modular data centers co-located with utility-scale renewable assets in remote locations, with computing demand adjusting to intermittent generation in real time. And by this point, it was explicitly describing those facilities as capable of performing scientific, blockchain and AI-related computing.

There is one detail in the May 2022 white paper that becomes particularly interesting later in this research.

While describing flexible computing as a curtailment tool, Soluna pointed to projects undertaken by Google and Microsoft alongside Bitcoin mining as examples of the broader category of flexible computing.

This was years before the current Kati 2 tenant speculation and before Microsoft’s later wind-co-located AI research began citing Soluna itself.

In other words, the awareness eventually runs in both directions.

Soluna was already citing Microsoft when explaining flexible computing in 2022. Microsoft researchers would later cite Soluna when examining AI computing at renewable-energy farms.

That connection belongs later in the Microsoft section, where its significance can be examined properly. But chronologically, its first appearance is here.

By 2022, Soluna had moved a long way from the original Moroccan proposal.

The company had started with a plan to build its own wind farm and computing facilities together. It had carried the model into the United States, tested the operating side, built a project-sourcing function and become part of a public company. It had also refined the commercial proposition from “renewable-powered cryptocurrency mining” into something much more useful to an independent power producer:

a flexible data-center load capable of sitting beside an existing renewable asset and turning constrained electricity into revenue.

That is the Soluna that matters for what happens next.

Because in Texas, a large wind asset was already operating.

Its owners were dealing with exactly the kind of problem Soluna had spent years learning how to solve.

And they were about to start looking for a data-center partner.

2. Las Majadas Before Soluna — The Asset That Made Kati Possible

By the time Soluna entered the picture, Las Majadas was already a large, operating wind asset with years of development work behind it.

It had moved through the full process required to turn a South Texas wind resource into institutional-scale energy infrastructure: local agreements, commercial origination, turbine procurement, construction, interconnection, financing and ultimately a new international co-owner.

The wind farm came first.

That matters because Kati did not need to create the underlying energy resource from scratch. It would eventually be built alongside an existing utility-scale generation asset with hundreds of megawatts of capacity, an established connection to ERCOT, a substation and two major renewable-energy companies behind it.

The later data-center opportunity was built on top of that foundation.

The Las Majadas story begins years before Kati

Las Majadas was already visible in Texas public records by 2016.

The Texas Comptroller’s Chapter 313 archive shows that Las Majadas Wind Farm, LLC filed its application with Lyford Consolidated Independent School District in September 2016. An amended application followed later that year, the state certification packet was posted in September 2017, and the final agreement and findings followed in January 2018.

The record places the project firmly inside EDF’s development pipeline. Las Majadas Wind Farm, LLC served as the project entity, while EDF Renewable Energy was identified as the developer.

This is important for the chronology.

Las Majadas was not created around a later computing opportunity. It was a large renewable project with its own development history, its own commercial structure and its own place in the South Texas energy market years before Soluna appeared.

Willacy County was already becoming a significant wind region, and Las Majadas added another large-scale asset to that landscape.

At roughly 273 MW, the project was substantial enough that even a portion of its generation could eventually support a very large computing load.

That possibility would become important later.

At this stage, however, the objective was straightforward: build a utility-scale wind farm, connect it to ERCOT and create a long-term commercial energy asset.

Jake Susman and the early commercial history

One of the clearest people in the early Las Majadas record is Jacob “Jake” Susman.

From August 2015 through June 2019, Susman served as Vice President and Head of Origination at EDF Renewables. A later SEC biography describes him as the executive responsible for customer-facing activities within EDF’s U.S. Grid Scale Power business, with a focus on originating long-term power contracts with corporate and utility customers.

Las Majadas appears directly in his own public project history.

Susman’s LinkedIn project record lists:

Las Majadas Wind Farm — March 2019 — 272 MW Wind Farm in South Texas — Offtaker undisclosed.

The same profile states that approximately 600 MW of wind and 600 MW of solar projects at EDF were enabled by contracts originated by the team he led.

That places Las Majadas directly within the commercial and origination work being handled by EDF’s senior customer-facing organization as the project moved toward construction.

The distinction matters.

Susman’s role was not simply technical or operational. Origination sits at the point where a renewable project becomes commercially viable: identifying counterparties, structuring long-term contracts and creating the revenue framework that allows large energy assets to move forward.

That makes him one of the earliest identifiable commercial nodes in the Las Majadas history.

Years later, different EDF teams would return to the same underlying question from another angle: how could the asset support new forms of demand and create additional value from the infrastructure already in place?

2019: Las Majadas moves into construction

By 2019, the project was moving decisively from development into physical execution.

On May 20, 2019, Vestas announced that EDF Renewables had ordered 249 MW of V120-2.2 MW turbines for the 273 MW Las Majadas project. The remaining capacity came from previously purchased 2 MW components.

EDF’s Art Del Rio described the turbine selection in economic terms, emphasizing the goal of achieving a competitive levelized cost of energy for the project.

A few weeks later, Infrastructure and Energy Alternatives announced the construction award for Las Majadas.

The scope included turbine installation, access roads, public-road work and the project’s medium-voltage collection system. Construction was scheduled to begin in July 2019. The project was described at the time as approximately 272 MW with 125 Vestas turbines.

By then, the future foundation for Kati was physically taking shape.

The turbines were being installed.

The internal electrical collection system was being built.

The site was being connected into the Texas power market.

The infrastructure that would later make behind-the-meter computing possible was being created years before anyone publicly associated the site with Soluna.

Masdar enters the ownership structure

The next major development came while Las Majadas was still being built.

On August 13, 2020, EDF Renewables North America and Abu Dhabi Future Energy Company — Masdar — announced a major U.S. partnership.

Masdar agreed to acquire a 50% interest in a 1.6 GW portfolio of eight EDF renewable-energy projects across Texas, Nebraska and California.

Las Majadas was one of the three wind projects specifically named in the transaction.

This is one of the most important ownership events in the Kati history because it establishes the EDF–Masdar relationship around Las Majadas well before Soluna enters the asset.

Masdar was not investing in a data-center project.

It was acquiring a long-term interest in a large U.S. renewable portfolio.

Las Majadas happened to be one of those assets.

That distinction becomes important later, because the eventual Soluna relationship would develop on top of an ownership structure that had already been established independently.

The Masdar transaction also significantly expanded the company’s footprint in the United States and deepened an existing global relationship with EDF.

One person was already at the top of Masdar during this period:

Mohamed Jameel Al Ramahi.

Al Ramahi was Masdar’s Chief Executive Officer when the EDF portfolio transaction was announced. In the August 2020 release, he described the deal as both an expansion of Masdar’s U.S. presence and a strengthening of the company’s partnership with EDF Renewables.

His presence here is worth establishing early.

At this point in the story, the relevant fact is simple:

Al Ramahi was already leading Masdar when the company entered the ownership structure containing Las Majadas.

That creates an important point of continuity that will become much more interesting later.

EDF and Masdar were already long-term partners

Las Majadas was also not an isolated transaction between two energy companies.

EDF and Masdar already had a significant international relationship before the U.S. portfolio deal.

When the transaction was announced, EDF Renewables North America CEO Tristan Grimbert described the companies’ existing cooperation in the Middle East and North Africa and positioned the U.S. investment as another step in a much broader partnership.

That background matters.

The later Kati agreement would not bring together unfamiliar counterparties. It would sit inside an existing institutional relationship between two large renewable-energy organizations with experience investing and developing together across multiple markets.

EDF brought the U.S. development and operating platform.

Masdar became a long-term economic participant in the assets.

The ownership structure therefore preceded Soluna by years.

In February 2021, Masdar announced the first closing on its acquisition of 50% of the 1.6 GW EDF portfolio. Las Majadas, Coyote and Milligan were then in the final stages of construction.

Al Ramahi again appeared as Masdar CEO in that announcement.

Masdar’s later financing disclosures provide additional detail on how the assets were held. Las Majadas and several other EDF projects were grouped within the Blue Palm structure, with Masdar holding its effective interest through Blue Palm Class B Wind HoldCo.

The legal and tax-equity structure is more complex than the shorthand “50/50 EDF–Masdar,” but the commercial picture is straightforward:

EDF and Masdar both had long-term economic exposure to Las Majadas before Soluna became part of the site.

Years later, when the Soluna agreement was publicly announced, EDF and Masdar would describe themselves directly as the co-owners of Las Majadas.

February 26, 2021: Las Majadas becomes operational

Las Majadas reached commercial operation on February 26, 2021.

Masdar’s later investor documentation identifies the project as a 273 MW U.S. onshore wind asset with commercial operation beginning on that date. EDF’s own reporting likewise records Las Majadas as a roughly 272.6 MW Texas wind farm commissioned in February 2021.

Masdar currently describes the project as 273 MW, consisting of 113 2.2 MW Vestas turbines and 12 2.0 MW Vestas turbines.

By early 2021, the future Kati site therefore already had the core pieces of a major energy platform:

• roughly 273 MW of operating wind generation;

• a completed electrical collection system;

• an established ERCOT interconnection;

• a substation;

• institutional ownership involving EDF and Masdar;

• and a long-term commercial structure around the electricity being produced.

This is one of the most important points in the entire Kati history.

Kati did not need to create the energy asset. It could be built around an asset that was already operating at scale.

That changed the development equation completely.

A later computing project would not have to wait for a new wind farm to be permitted, financed, built and interconnected before it could begin thinking about power.

The generation was already there.

The interconnection was already there.

The institutional owners were already there.

The next opportunity was to find a new way to use more of that infrastructure.

A commercial structure with room for optimization

Masdar’s financing disclosures also provide useful insight into the commercial profile of Las Majadas.

The project was not operating under a simple full-output physical PPA in which every megawatt-hour was sold at the same fixed price to a utility.

Masdar classifies Las Majadas as having a merchant offtake profile and describes energy hedges extending through December 2032. Las Majadas and Coyote shared an investment-grade hedge counterparty, with the projects required to provide specified monthly quantities of power under fixed-price arrangements.

The identity of that counterparty is not disclosed in the document.

What matters for the Kati story is the structure.

Las Majadas remained connected to the economics of the ERCOT market.

That created room for optimization.

A large renewable asset can already be successful as a generation project while still becoming more valuable if additional demand can be created close to the point of production.

Behind-the-meter load offers exactly that possibility.

Instead of relying solely on selling electricity through the existing market structure, a portion of the generation can be paired directly with a large customer located beside the asset.

That can create a new revenue stream for the wind farm while giving the customer access to large-scale power without beginning from an undeveloped site.

The two sides complement each other.

Las Majadas had the generation. A large computing customer could bring additional demand directly to it.

That is the commercial logic that begins to matter after the wind farm reaches operation.

The opportunity after COD

By the end of 2021, Las Majadas had completed its transition from development project to institutional-scale operating asset.

EDF had developed it.

A commercial origination team had helped establish the contracts required to advance it.

Vestas had supplied the turbine fleet.

The project had been built and connected to ERCOT.

Masdar had entered the ownership structure.

And hundreds of megawatts of wind generation were now operating in South Texas.

The foundation was already there.

The next opportunity was to use that foundation more efficiently.

A large, controllable customer located close to the wind farm could create additional demand without requiring EDF and Masdar to build another renewable project from scratch. It could make use of existing generation, existing interconnection and existing site infrastructure while opening a new commercial path for the asset.

That possibility was particularly well suited to data-center load.

And EDF was about to begin looking for exactly that kind of partner.

Las Majadas already had the power. The next opportunity was to bring the right load to it.

In 2022, EDF formalized that search.

It had a name:

Project Tumbleweed.

3. Project Tumbleweed — EDF Goes Looking for Load

By 2022, the pieces on both sides of the future Kati relationship were already beginning to align.

Soluna had spent years refining a model built around flexible computing placed close to renewable generation. Its U.S. strategy was increasingly centered on working with existing power producers rather than developing every renewable asset itself. EDF and Masdar, meanwhile, already owned a 273 MW operating wind farm in South Texas, with generation, land, interconnection and a substation in place.

The opportunity was no longer theoretical. The energy infrastructure already existed. What remained was to find the right kind of load and the right counterparty to place beside it.

EDF did not leave that process to chance. It began looking for exactly that kind of partner.

Project Tumbleweed

The clearest evidence comes from EDF itself.

In later documentation for its 2024 behind-the-meter initiative, Project Arete, EDF looked back at the programs that had preceded it. Among them was Project Tumbleweed, launched in 2022 and focused specifically on behind-the-meter data centers in ERCOT. EDF also identified a 2023 follow-on initiative called Project Champagne and said the earlier programs had generated “tangible commercial opportunities” that had moved into various stages of deal flow.

That makes Tumbleweed one of the most important early markers in the Kati timeline. It shows that EDF was not simply waiting for a computing company to approach one of its wind projects with an unsolicited proposal. The company had begun a structured search for counterparties capable of placing meaningful electrical demand directly alongside renewable generation.

The fit with Soluna was unusually close. By that point, Soluna had already spent years developing a business around behind-the-meter computing, variable renewable power and large controllable loads. Tumbleweed was effectively asking the market a question that Soluna had already organized itself around: who could bring data-center demand directly to renewable generation and make that relationship work commercially and technically?

There is also an important feature of EDF’s RFP process that helps explain why Tumbleweed does not leave behind the kind of public paper trail one might normally expect from a competitive selection.

EDF handled this class of process under strict confidentiality.

The surviving documentation for Project Arete shows how EDF structured the same continuing family of behind-the-meter RFPs. Before receiving detailed project information, participants were required to execute EDF’s standard two-year confidentiality and non-disclosure agreement. RFP documents and project information were treated as proprietary and confidential, detailed materials were released only to participants with active NDAs through a secure data room, contact with project counterparties required EDF approval, and public announcements relating to the RFP could not be made without EDF’s prior written consent.

The same document explicitly identifies Arete as following the earlier Tumbleweed and Champagne processes. It also shows the progression EDF used: participant outreach and NDA execution, release of detailed materials, evaluation, a commercial shortlist, finalists, and finally term sheets leading toward definitive agreements.

That framework matters when reconstructing Tumbleweed.

A process built this way would not be expected to produce a public bidder list, detailed scorecards or an announcement naming every finalist. The commercially meaningful information would emerge later, as selected counterparties moved into term sheets and definitive transactions.

And that is almost exactly what we see with Soluna.

Bringing demand to existing generation

EDF’s later account of the strategy describes a broader effort to increase the value of its existing renewable infrastructure by bringing flexible demand closer to generation. Rather than viewing transmission expansion as the only path for additional energy delivery, EDF’s Asset Optimization team began exploring whether large customers could be located directly at renewable sites.

The logic was attractive on both sides.

For the renewable owner, a large behind-the-meter customer could create additional demand using generation and infrastructure that were already in place. For the computing company, the same arrangement could provide access to large blocks of power without starting from an undeveloped location and waiting for an entirely new energy system to be built around it.

Paces, which later worked with EDF on its broader data-center strategy, describes the approach directly. EDF’s Asset Optimization team ran RFP processes seeking flexible counterparties willing to co-locate at EDF substations and purchase electricity behind the meter. The Soluna agreement at Las Majadas is identified as one of the early commercial agreements to emerge from that work.

That retrospective account gives Tumbleweed much more weight. It was not simply an innovation exercise inside EDF. It formed part of a wider effort to turn existing renewable infrastructure into the foundation for large new loads.

Las Majadas was particularly well suited to that idea. It was already operating at utility scale, already interconnected and already backed by EDF and Masdar. The energy side of the equation existed. The next step was to identify a computing partner capable of building and operating beside it.

What EDF expected from a partner

Later EDF materials also provide useful insight into the kind of counterparty the company wanted.

Project Arete came after Tumbleweed, so its requirements should not be treated as the exact 2022 Tumbleweed bid package. But EDF explicitly presented Arete as a continuation of the same line of work, and its requirements show the level of capability expected from a behind-the-meter participant.

Participants were expected to demonstrate the ability to design a complete BTM load system, handle planning and permitting, construct the facility, understand local utilities, develop project financial analysis, maintain and monitor the system and operate under established safety requirements. EDF also made clear that the participant would carry the capital and operating requirements of its own facility, while EDF’s role centered on supplying energy through a PPA rather than taking an equity position in the load facility.

In other words, EDF was not simply looking for a company willing to consume a lot of electricity. It was looking for a counterparty capable of developing real infrastructure beside an institutional renewable asset.

That required more than appetite for megawatts. The partner needed development capability, engineering knowledge, power-market experience, financing capacity and enough operational credibility for EDF to place a major new facility directly beside one of its existing generation assets.

By 2022, Soluna had been building toward exactly that role.

It had already developed modular computing infrastructure in the United States. It had experience sourcing sites, negotiating energy relationships and operating around variable power. Its commercial proposition was increasingly centered on becoming a flexible load partner to renewable owners rather than simply a buyer of inexpensive electricity.

On paper, the match was unusually strong.

The people behind the strategy

One of the central people in EDF’s Asset Optimization work was Gabe Messercola.

Messercola had moved into Asset Optimization after working in solar asset management. His work expanded into opportunities for renewable projects to support energy-intensive flexible loads, including computing and data-center applications.

His place in the Las Majadas timeline represents a different stage in the life of the asset from the earlier origination work associated with Jake Susman.

Susman’s role had been connected to bringing renewable projects and their commercial arrangements into existence. Asset Optimization was looking at what more an already operating renewable asset could support.

Once Las Majadas was built, interconnected and producing power, the opportunity expanded. The existing generation and electrical infrastructure could now become the foundation for a new class of customer located directly beside the asset.

Messercola was not working alone. EDF’s later Arete documentation names Peter Hoegler and Chase Rebeil alongside him within Capital Improvements, Portfolio Management and Asset Optimization. Messercola was listed as Associate Director, Hoegler as Transaction Manager and Rebeil as Transaction Analyst.

This gives a clearer picture of what sat behind Tumbleweed and the later co-location strategy. EDF had a real transaction and asset-optimization function working through how large new loads could be integrated with renewable generation.

That organizational structure becomes particularly important when Soluna begins appearing publicly in 2023.

Soluna emerges from the selection process

By 2023, Soluna had begun speaking publicly about a large new development called Project Kati.

The identity of the renewable partner was still undisclosed, but John Belizaire described it as a serious independent power producer with major assets in the United States and internationally.

Then he revealed how Soluna had reached that position.

In the Water Tower Research interview published in August 2023, Belizaire said the power producer had selected Soluna from roughly 25 companies. He described a genuine competitive process in which Soluna and other potential partners had been put through extensive evaluation before the IPP chose Soluna.

Water Tower’s own summary was even more direct. It referred to “the win at Project Kati,” said Soluna had been selected from more than twenty other players, and highlighted the importance of the wind-farm owners visiting Project Dorothy during the selection.

That site visit adds an important dimension to the story.

The potential partner did not have to rely solely on presentations about what Soluna claimed its model could eventually do. It could visit a functioning Soluna site and see the operating system in practice.

Belizaire described how the visitors came to Project Dorothy and how Soluna’s existing operations became part of the evaluation. Water Tower specifically highlighted their interest in MaestroOS and Soluna’s modular data-center design. Belizaire said the companies were already collaborating to move agreements forward and begin the interconnection process.

Project Dorothy therefore appears to have served as a practical proof point during the Kati selection.

That changes the way the early relationship should be understood.

Soluna was not simply another buyer looking for renewable power. It competed for the opportunity, demonstrated an operating version of its model, and was chosen as the computing counterparty.

Tumbleweed, the NDA and the “25 companies”

This is where the chronology becomes particularly compelling.

EDF tells us that Project Tumbleweed was launched in 2022 as an RFP focused specifically on BTM data centers in ERCOT. EDF later says the process generated tangible commercial opportunities that moved into deal flow.

Soluna tells us that its major IPP partner subsequently evaluated approximately 25 companies and selected Soluna. Water Tower independently summarizes the same event as a Project Kati win from a field of more than twenty other players.

The power partner visited Project Dorothy during that evaluation.

A Kati term sheet followed.

ERCOT work followed.

And EDF later identified Soluna at Las Majadas as one of the early agreements resulting from its effort to find flexible counterparties willing to co-locate at renewable substations.

The resulting project was exactly the kind of opportunity Tumbleweed had been created to find:

a behind-the-meter data center at an EDF renewable asset in ERCOT.

The confidentiality structure helps explain why there is no public document laying out the process as:

“Tumbleweed bidder list → finalists → Soluna wins.”

EDF’s documented RFP framework keeps precisely that type of commercially sensitive information behind NDAs while successful participants move from evaluation to finalist status, term sheets and eventually definitive agreements.

What becomes visible publicly is the outcome of that process.

And the surrounding evidence points strongly in one direction.

Project Tumbleweed appears to have produced Soluna as the winning computing partner for what became Project Kati, whether through the original Tumbleweed award itself or through the direct commercial selection funnel that followed from it.

We do not need a public winner’s certificate to see the sequence. The process is visible on both sides of the missing confidential layer.

Before it, EDF is searching for BTM data-center partners in ERCOT.

After it, Soluna says it beat roughly 25 candidates, hosts the power partner at Project Dorothy, signs a Kati term sheet and begins ERCOT work on a 166 MW BTM project.

The NDA sits between those two public views.

The commercial outcome sits on the other side.

2023: the Kati term sheet

Soluna’s SEC filings confirm that a term sheet for Project Kati was signed in 2023.

The project was already large. At roughly 166 MW, it represented a significant expansion in the scale of what Soluna was attempting to develop around renewable generation and was described as being integrated with a large South Texas wind facility.

Soluna then spent the remainder of 2023 advancing the project through ERCOT’s planning process.

The progression is now much easier to see.

EDF launches a formal search for BTM data-center load in ERCOT. The process is handled through a confidential institutional RFP structure. Soluna later says the IPP selected it from roughly 25 companies after a serious evaluation. The partner visits Project Dorothy and sees Soluna’s operating model firsthand. Kati receives a term sheet. ERCOT and development work begin.

What started as a competitive search was turning into a specific commercial project.

Kati was large from the beginning

The original scale of Kati also deserves attention.

At approximately 166 MW, this was never intended to be a small demonstration project. For Soluna, it represented a major expansion in the scale of what the company was attempting to develop around renewable generation.

The site was tied to a wind asset of roughly 300 MW, closely aligning with Las Majadas’ approximately 273 MW nameplate capacity.

Later disclosures would make the two-phase structure clearer, but even at this stage the project was being discussed in a way that left room for more than one kind of computing workload. Subsequent filings would explicitly describe Kati as capable of supporting high-performance computing applications, including AI.

That is an important part of the longer Kati 2 history.

The potential for higher-value computing was already present in the broader Kati concept. What would change later was the scale, design and commercial ambition around AI.

From selection to execution

Being selected was the beginning of a much deeper development process.

A 166 MW facility integrated with an operating utility-scale wind farm required the parties to coordinate electrical planning, interconnection, land, engineering and the commercial mechanics of using power behind the meter. The structure also had to work across three institutional parties: EDF, Masdar and Soluna.

That helps explain the work taking place between the 2023 term sheet and the definitive agreement that followed.

When Gabe Messercola later reflected publicly on the transaction, he highlighted the breadth of EDF functions involved in bringing it together: Power Marketing, Interconnection, Engineering, Asset Management and Legal, alongside Masdar Americas and Soluna.

The range of teams involved says a great deal about what had been accomplished. This was not simply the sale of electricity to a new customer. EDF and Masdar were integrating a major computing facility with an existing renewable asset, while Soluna was building a new kind of data-center relationship around power infrastructure that was already operating.

By the spring of 2024, that work was ready to become definitive.

On May 22, 2024, Soluna announced the commercial structure that had emerged from the process:

a 166 MW behind-the-meter Power Purchase Agreement with EDF Renewables and Masdar.

A confidential search had produced a selected partner.

The selected partner had become a development counterparty.

And the development relationship had become a signed power agreement.

Project Kati was ready to move into its next phase.

4. From Agreement to Operating Infrastructure — Kati Proves the Model

The selection process established that EDF had found the kind of computing partner it was looking for. The next phase was more important: turning that selection into infrastructure.

On May 22, 2024, Soluna announced that it had signed a definitive Power Purchase Agreement for Project Kati with EDF Renewables and Masdar. The agreement covered up to 166 MW of renewable power and placed the planned data center alongside the wind facility jointly owned by the two renewable-energy companies. From the beginning, the project was structured in two equal phases of 83 MW each, with the power intended for high-performance computing applications including AI.

For Soluna, this was considerably more than another addition to its development pipeline. The agreement took the model it had spent years building — flexible computing located directly beside renewable generation — and attached it to a 273 MW institutional wind asset owned by EDF and Masdar.

For EDF and Masdar, the arrangement created a large new source of demand directly beside Las Majadas. The wind farm did not have to change its fundamental purpose as a renewable-energy asset. Instead, the infrastructure around it could support another route for its electricity: a major computing facility capable of purchasing power behind the meter.

The commercial idea that began with Tumbleweed had now become a signed 166 MW power agreement.

The structure was designed around flexibility

The significance of the PPA becomes clearer when the physical arrangement is considered.

Project Kati was not simply going to purchase renewable-energy certificates or sign a conventional contract for wind power generated somewhere else on the grid. The data center was to be constructed close to the Las Majadas substation and receive power generated by the wind farm behind the meter.

That physical proximity was central to the model.

When EDF and Masdar later described the agreement publicly, they explained that Project Kati would purchase up to 166 MW from Las Majadas while retaining the ability to reduce its operations during certain market conditions when the grid placed greater value on the electricity elsewhere. EDF described the structure as an alternative route for utilizing electricity from the wind project, while Gabe Messercola characterized behind-the-meter offtake as a way to physically deliver part of a renewable plant’s output directly to a co-located buyer.

This was the flexible-load concept Soluna had been developing for years, now operating at utility scale.

The computing facility could become a substantial local customer when the economics favored direct consumption, while preserving the ability for the wind farm and grid to respond to broader market conditions. That flexibility gave the same megawatts more than one potential path to value.

EDF and Masdar themselves presented the result positively. Messercola called the arrangement a “win-win situation,” while Masdar Americas’ Dustin Priemer emphasized that the structure could improve the utilization of electricity generated at Las Majadas while supplying renewable energy to a new data center.

The structure also reflected something Soluna had been saying since its earliest years: the computing load should not simply sit at the end of the power system and demand electricity regardless of conditions. It should be designed to work with the characteristics of the generation beside it.

Kati was the largest expression of that idea Soluna had attempted to build.

The PPA was followed by the work that makes a site real

Signing the power agreement did not mark the end of development. It allowed the project to move deeper into it.

Soluna had already been working through ERCOT planning during 2023. In 2024, the company resubmitted a key study to reflect a change in the project’s distance from the substation, showing that the electrical layout was being refined around the actual configuration of the site. The process continued after the PPA as Soluna worked through the planning, land and infrastructure steps necessary for construction.

The next major milestone came on February 18, 2025, when Soluna announced that Project Kati had successfully exited the ERCOT planning phase. The company described the milestone as clearing the path toward construction and expansion of the full 166 MW development.

This is an important point in the Kati timeline because it separates a power agreement from a development that is actually progressing through the electrical system.

The sequence was becoming increasingly concrete. Soluna had been selected. A term sheet had been signed. The definitive PPA followed. ERCOT planning advanced. The electrical relationship to the site was being engineered. Land and construction requirements were moving alongside it.

By early 2025, Kati was no longer simply a commercial concept attached to a wind farm. The pieces required to build the facility were being put into place.

EDF and Masdar put their own names behind the model

Another important moment arrived only weeks after the ERCOT milestone.

On March 4, 2025, EDF Renewables North America and Masdar jointly announced the Las Majadas agreement from their side.

The date can easily create confusion because Soluna had already announced the definitive PPA in May 2024. The March 2025 release was therefore not the beginning of the relationship. Its importance is that EDF and Masdar themselves publicly identified Las Majadas, explained the economics of the structure and placed Project Kati inside their own strategy for renewable power and data-center demand.

Their announcement removed the remaining ambiguity around the underlying renewable asset. EDF and Masdar identified themselves as the co-owners of the 273 MW Las Majadas Wind Project in Willacy County and confirmed that up to 166 MW of its generation would supply a Soluna data center constructed close to the wind project’s substation.

The language used by the energy partners is particularly important for understanding how they viewed Kati.

EDF did not present Soluna as a simple purchaser of commodity electricity. The company described the PPA as an innovative behind-the-meter structure capable of creating a direct route between renewable generation and an energy-intensive computing operation. It explicitly connected the facility to advanced computing applications, including artificial intelligence.

Masdar framed the agreement in similarly strategic terms. Dustin Priemer pointed directly to the rapid expansion of U.S. data centers and the rising electricity demand associated with them, presenting the Las Majadas arrangement as a way to combine renewable generation with that new class of demand.

That public endorsement matters.

Soluna was no longer the only company explaining why the model made sense. The owners of the power asset were now describing the same logic themselves.

Las Majadas was becoming more than the source of electricity behind Project Kati. It was becoming a working example of how renewable generation and large-scale computing could be developed together.

The first 83 MW becomes Kati 1

During 2025, the original two-phase structure began to take a much more visible form.

The first 83 MW became Project Kati 1. The second 83 MW would become Project Kati 2.

This split is important because the two phases began serving different strategic purposes. Kati 1 was developed around Bitcoin hosting, giving Soluna a workload it already knew how to build and operate at scale. Kati 2 was preserved for AI and high-performance computing, where the potential economics and infrastructure requirements were considerably larger.

By September 2025, Soluna was describing the structure explicitly: Kati 1 would provide 83 MW of Bitcoin hosting, while Kati 2 would follow with another 83 MW focused on AI and HPC.

This created a useful development sequence.

Rather than waiting for a large AI tenant before proving that the Las Majadas relationship could support a real computing facility, Soluna could build the first half of the original power allocation using a familiar workload.

Kati 1 could prove the site.

It could prove the electrical configuration.

It could prove that the behind-the-meter relationship with Las Majadas worked outside a contract document.

And it could establish operating infrastructure beside the same wind asset intended to support the future AI development.

That made Kati 1 much more significant than its workload alone might suggest.

Kati 1 was becoming the physical proof point for the larger Kati platform.

Land, substation equipment and capital begin locking into place

The development milestones accelerated through the spring and summer of 2025.

By May, Soluna reported that land for the 166 MW Project Kati had been secured. The Phase 1 substation upgrade for the first 83 MW had also been completed, and orders for important long-lead electrical equipment — including medium-voltage switchgear, transformers and busway — had been secured.

These details matter because they show the project moving across several development layers at the same time.

Power had been contracted through the PPA. ERCOT planning had been completed. The land was being controlled. The substation work was advancing. Critical electrical equipment was being ordered.

The project was steadily becoming physical.

Capital followed.

In June 2025, Soluna announced that longtime infrastructure investor Spring Lane Capital had agreed to lead financing for the first 35 MW of Kati 1, with at least $20 million committed and a broader agreement that could provide up to $100 million of additional project-level capital across Soluna’s pipeline. Spring Lane was already familiar with the company through investments in Projects Dorothy 1A and Dorothy 2.

The financing closed on July 22, 2025, allowing construction of Kati 1 to move forward. SEC filings show that Spring Lane’s Kati investment vehicle was ultimately structured with a capital contribution cap of up to approximately $48.98 million for the project.

That brought another institutional partner into the Kati structure.

EDF and Masdar supplied the renewable-energy foundation.

Soluna developed and operated the computing infrastructure.

Spring Lane supplied project capital.

The model was starting to look less like an individual power contract and more like a repeatable project-development platform.

Galaxy fills out the first phase

The next major piece arrived in August 2025.

Soluna announced an expanded relationship with Galaxy Digital, which agreed to deploy 48 MW of its proprietary Bitcoin mining operations at Kati 1. Combined with the 35 MW Soluna had already been preparing, the Galaxy agreement brought the first phase to its full 83 MW capacity.

The scale was meaningful for Soluna. The Galaxy deployment was the company’s largest customer deployment announced to that point and gave Kati 1 a major operating counterparty before construction began in earnest.

It also completed the commercial structure of the first 83 MW.

The power was there.

The land was there.

The substation work had advanced.

The project capital was being assembled.

The customer capacity was now committed.

Construction could begin.

On September 18, 2025, Soluna formally broke ground on Project Kati in Willacy County. The company described the development as its largest site to date and again emphasized the two-track nature of the full project: Kati 1 for Bitcoin hosting and Kati 2 for AI and HPC.

The original 166 MW PPA was now becoming visible on the ground.

From contract to energized megawatts

The most important proof arrived in early 2026.

On February 9, 2026, Soluna received ERCOT approval to begin energizing and commissioning Project Kati 1. The company announced the milestone the following day. Kati 1 was structured as 48 MW of Kati 1A and 35 MW of Kati 1B, for the full 83 MW first phase.

Soluna stated that Kati 1 was powered entirely by the Las Majadas wind energy project and described the site as a demonstration of its Renewable Computing blueprint at scale.

That is the point where the Kati history changes character.

Years earlier, EDF had gone searching for a behind-the-meter data-center counterparty in ERCOT. Soluna had emerged from a competitive field and entered into a term sheet. EDF, Masdar and Soluna had then spent the following period building the commercial and electrical structure around Las Majadas.

Now electricity from that same wind asset was actually energizing the computing facility.

The chain could finally be followed all the way through:

Tumbleweed and the partner search led to Soluna.

The selection led to a term sheet.

The term sheet led to the definitive PPA.

The PPA led through ERCOT, land, substation work, procurement, financing and construction.

And construction led to energized computing capacity beside Las Majadas.

The model was no longer only contractual. It was operating.

By the first quarter of 2026, all three Kati 1A phases totaling 48 MW had reached completion ahead of schedule, supported by additional project financing from Spring Lane. By June, those Galaxy operations were running steadily while construction and commissioning continued across the remaining 35 MW of Kati 1B.

For the future of Kati 2, that operating history is difficult to overstate.

The second phase would ultimately become much larger and considerably more sophisticated than the original 83 MW plan. But it would not be starting beside an untested renewable asset with an untested commercial relationship.

It would be expanding beside a site where Soluna, EDF and Masdar had already taken the core behind-the-meter concept from selection to power contract to physical operation.

Las Majadas becomes a proof point beyond Soluna

The importance of Kati was not confined to Soluna.

EDF’s subsequent development strategy shows that the Las Majadas experience had become valuable to the renewable owner as well.

A later case study from Paces describes the Soluna co-location agreement at Las Majadas as one of EDF’s early behind-the-meter transactions and says the success of those early agreements validated the model. EDF then moved beyond using co-location only as an optimization tool for existing operating assets and began considering how large flexible loads could be incorporated into new energy developments from the outset.

The organizational consequence was significant.

Peter Hoegler moved from Asset Optimization into a newly created Data Center Solutions team, whose mandate became identifying locations where large flexible loads and power generation could be developed together. EDF later worked through more than 90 potential sites, ranking them against data-center-specific criteria such as infrastructure, siting, regulatory conditions, fiber and firm gas access.

That is a remarkable evolution from the original Tumbleweed search.

In 2022, EDF was asking which companies might be willing and able to place data-center load beside its renewable assets.

A few years later, it had a dedicated organization evaluating its portfolio specifically for large-scale data-center co-location.

And Las Majadas was part of the proof that helped move the strategy forward.

Paces describes the shift plainly: early co-location agreements, including Soluna at Las Majadas, validated the model and contributed to the creation of EDF’s dedicated Data Center Solutions effort.

That makes Kati important in two directions.

For Soluna, it demonstrated that the company could win a competitive process with a major renewable owner and take a utility-scale BTM project into operation.

For EDF, it demonstrated that renewable assets could become platforms for large computing loads rather than remaining solely generation projects.

Soluna had become more than a customer at Las Majadas. The relationship had helped prove a model that EDF itself would begin looking to reproduce.

And the second 83 MW was still waiting

By the time Kati 1 began energizing in February 2026, one half of the original 166 MW structure had therefore completed most of the journey from paper to operation.

The second half had been deliberately reserved for something different.

Throughout 2025, Soluna continued identifying Kati 2 as the AI and HPC side of the site. At that stage, it was still generally presented within the original 83 MW framework.

But while Kati 1 was moving through financing, construction and energization, the ambitions for Kati 2 were beginning to change.

The future AI project would not remain an 83 MW mirror image of Kati 1.

It would grow far beyond the original PPA phase.

The land footprint would expand.

The power architecture would expand.

The substation plan would expand.

New partners would enter.

And the project would eventually be redesigned around hundreds of megawatts of critical IT capacity.

Before following that transformation, however, another story has to be introduced.

During the same years that Soluna and EDF were proving that large-scale computing could be placed directly beside renewable generation, Microsoft had been studying the same fundamental idea from the computing side.

That research had started years earlier.

And by the time Kati 1 was becoming real, Microsoft Research had already begun putting Soluna’s name into its own work.

5. Microsoft Research — Moving Compute Toward the Power

By the time Kati 1 was moving from contract to construction, Microsoft had already spent years exploring a remarkably similar question from the computing side.

Soluna had started with renewable generation and asked how computing could be brought closer to it.

Microsoft Research was increasingly asking what happens when the location and behavior of computing itself become flexible — when workloads can move in response to where power is available, when modular data centers can be placed beside renewable generation, and when software becomes part of the electrical architecture rather than merely consuming whatever power the grid delivers.

The overlap did not appear suddenly in 2026. It developed over several years, and as the research became more specific, it moved steadily closer to the physical model Soluna had already been building.

2021: Virtual Battery changes the direction of the problem

One of the clearest starting points is Microsoft’s 2021 paper “Redesigning Data Centers for Renewable Energy.”

The paper was published at ACM HotNets by researchers affiliated with Microsoft, Carnegie Mellon and the University of Illinois, including Shadi Noghabi, Srinivasan Iyengar, Anirudh Badam, Ranveer Chandra and Shivkumar Kalyanaraman on the Microsoft side. Microsoft Research continues to host the work as part of its publication archive.

The central idea was called Virtual Battery.

Instead of treating renewable variability as something that had to be solved entirely by batteries, transmission or additional generation, the researchers proposed changing the behavior of the computing load itself.

Their formulation was unusually direct: rather than always adapting the availability of power to match computational demand, computational demand could be shifted to match the availability of power. Applications could be made more flexible, delayed when appropriate, or migrated toward locations where renewable energy was available.

The architecture went beyond workload scheduling inside an ordinary centralized data center.

The paper proposed placing smaller data centers directly alongside renewable-energy farms, creating a geographically distributed group of computing sites. Each renewable source would be paired with compute sized around the local generation profile, and the computing capacity at each location could scale up or down according to the electricity being produced there.

Microsoft was not merely researching how to buy more renewable electricity for existing cloud regions. It was exploring a different physical arrangement for cloud infrastructure: put computing beside the energy source, then make the workload respond to the power.

The paper also anticipated that these renewable-edge sites would not need to replace Microsoft’s large conventional data centers. The researchers envisioned a mixed architecture in which large centralized facilities continued operating while smaller renewable-linked sites complemented them.

That same principle becomes increasingly important in Microsoft’s later work: renewable co-located compute is treated as additional infrastructure capacity that can work alongside the existing cloud, not as an all-or-nothing replacement.

Virtual Battery also introduced another idea that would survive through the later research. A single renewable location has its own production profile, while several geographically separated renewable sites can behave differently at the same moment. When generation falls at one location, another can have more available power.

Microsoft’s researchers therefore explored multi-site operation, replication and workload migration across renewable-powered computing sites. They found that complementary generation patterns across locations could reduce aggregate variability, while a network- and power-aware scheduler could control where applications were placed.

The significance for this research is that by 2021 Microsoft researchers had already articulated the same underlying inversion that had been central to Soluna’s model for years:

instead of always moving electricity toward a fixed data center, computing itself can move toward the electricity.

The idea was not isolated inside Microsoft

Virtual Battery also sits inside a broader Microsoft effort to treat power as an actively managed computing resource.

Microsoft’s Power Efficiency and Sustainability research program dates back to 2016 and brings together Azure, Cloud Operations + Innovation and research teams around power capping, oversubscription and the recovery of otherwise underutilized data-center power capacity. Microsoft says versions of this work have been deployed across millions of servers and have released hundreds of megawatts of power capacity for additional computing.

That operational history matters because it shows that the underlying philosophy extends beyond academic papers.

Microsoft has spent years treating computing demand as something that can be controlled around electrical limits. Power can be measured, workloads adjusted, server density changed, reserved capacity reclaimed, and software and hardware controls used to make the computing system respond more intelligently to the power system.

Virtual Battery took that thinking outside the walls of the conventional data center and applied it to geographically distributed renewable generation.

The following research would push that idea further.

2024: SkyBox moves modular data centers directly to energy farms

In June 2024, a new paper appeared:

“Exploring the Efficiency of Renewable Energy-based Modular Data Centers at Scale.”

The system proposed in the paper was called SkyBox.

Among the authors were Shadi Noghabi and Ranveer Chandra of Microsoft Research, both of whom had also appeared in the 2021 Virtual Battery work. The paper was later presented at ACM’s Symposium on Cloud Computing in November 2024.

That author continuity matters. SkyBox was not a separate research direction that happened to resemble Virtual Battery. It was a further development of the same architectural idea.

The paper starts from modular data centers that can be placed “right at the energy farms” and powered primarily by renewable energy. It then asks how those deployments can work efficiently across multiple geographic regions when renewable production differs from site to site.

The researchers used real-world power traces from renewable-energy farms and found that generation patterns could be both predictable and geographically complementary. SkyBox used those characteristics to identify suitable renewable sites, group locations whose production profiles complemented one another, and use smart workload placement and migration to move computing around the available renewable power.

Virtual Battery had established the principle. SkyBox was beginning to turn that principle into a deployment framework.

The question was becoming more concrete: which renewable farms should host computing, how much compute should be placed at each one, and how should workloads move between them as generation changes?

Those are infrastructure questions, and they were increasingly relevant to the kind of development Soluna was pursuing.

By the time SkyBox appeared in June 2024, Soluna had already announced its definitive 166 MW EDF/Masdar PPA for Project Kati. The two organizations were approaching the same broader opportunity from different directions. Soluna was turning renewable co-location into a commercial energy-and-data-center project, while Microsoft researchers were developing increasingly sophisticated methods for deploying and operating compute directly at renewable-energy farms.

Microsoft was also learning to manage AI power inside the data center

Another Microsoft research branch was developing in parallel.

In April 2024, Microsoft researchers published “Characterizing Power Management Opportunities for LLMs in the Cloud.”

The work examined real power characteristics of large-language-model training and inference and introduced POLCA, a framework designed to make more efficient use of the electrical capacity already available inside GPU data centers. Microsoft reported that the approach could support roughly 30% more servers in existing inference clusters with limited throttling under the evaluated conditions.

The important idea is broader than the specific percentage. Microsoft was treating GPU power consumption as something that could be actively orchestrated.

The research distinguished between training and inference because the two workloads have different electrical behavior. Large synchronized training jobs can produce highly coordinated power peaks, while inference fleets can contain more exploitable headroom at cluster level.

This thread continued into 2025.

DynamoLLM dynamically reconfigured inference clusters around workload conditions, performance requirements, energy consumption and cost. TAPAS, published in April 2025, went further by combining thermal and power constraints with workload placement, inference routing and GPU configuration.

Together, these papers establish a wider pattern in Microsoft’s AI infrastructure thinking: the workload does not have to be passive. It can be routed, reconfigured and moved between resources, while power itself becomes a scheduling variable alongside latency, performance and cost.

That broader capability becomes essential once Microsoft begins applying the same philosophy directly to wind-powered AI infrastructure.

May 2025: Heron brings AI inference to wind farms

On May 15, 2025, Microsoft researchers submitted a paper that moves much closer to the architecture at the center of this investigation:

“AI Greenferencing: Routing AI Inferencing to Green Modular Data Centers with Heron.”

The paper focused explicitly on modular AI computing clusters co-located at wind farms.

This was no longer generic renewable-aware cloud scheduling. The physical location of the computing infrastructure was now central to the proposal.

The researchers argued that AI inference could be brought directly to wind generation, allowing GPUs to consume lower-cost renewable electricity at the source while creating valuable local demand for wind assets. They called the model AI Greenferencing.

Using Global Energy Monitor wind data and Azure data-center locations, the paper estimated that more than 640 GW of large wind capacity globally — including more than 150 GW in the United States — sat within 50 milliseconds of fiber round-trip latency from Azure data centers. Approximately 77% of the identified capacity was within 20 milliseconds.

The researchers then asked how much AI hardware could realistically be deployed at those wind farms without sizing the compute fleet to the absolute peak output of the turbines. Their answer was a right-sizing strategy, with residual variation handled through software and the complementarity of multiple wind locations. Under the paper’s assumptions, more than six million H100-class GPUs could theoretically be deployed at large wind farms.

Then came Heron.

Heron was a cross-site software router that directed AI inference requests among wind-powered computing clusters. It considered power availability, hardware constraints, network latency and workload characteristics when deciding where inference should run.

Using real wind-power traces together with one week of Azure production traces, Microsoft’s researchers reported that Heron improved aggregate AI-compute goodput by up to roughly 80% relative to the comparison system used in the study.

The relevance to Soluna becomes much more direct at this point.

The Heron paper explicitly cited Soluna.

Microsoft Research names Soluna

In the introduction to the Heron paper, Microsoft researchers point to windCORES, Soluna and WestfalenWIND as examples of companies already deploying computing around wind generation.

The researchers later describe co-located compute startups such as Soluna as potential beneficiaries of AI Greenferencing because AI inference represents a higher-value workload than the workloads these facilities had historically supported.

They also discuss two commercial paths. An AI provider could deploy its own computing equipment directly at renewable sites, or it could use third-party compute operated by companies already co-locating infrastructure at wind farms, with Soluna again included among the examples.

That is a significant distinction for this research.

Microsoft’s researchers were not merely aware that Soluna existed. They had placed the company directly inside a technical and commercial discussion about how AI workloads could be served from computing infrastructure co-located at wind generation.

And the timing is notable.

The paper appeared in May 2025. Two months earlier, EDF and Masdar had publicly identified Las Majadas and Project Kati as their 166 MW behind-the-meter computing project with Soluna. During the same period, the first Kati phase was moving toward construction and the second phase was increasingly being positioned toward AI and HPC.

Microsoft Research was therefore discussing Soluna as an example of a compute-at-wind operator while Soluna was actively developing its largest wind-co-located project toward AI use.

The connection was no longer simply architectural similarity. There was now direct awareness.

A second Microsoft branch: stabilizing hyperscale GPU power

Later in 2025, another Microsoft paper addressed a different piece of the same infrastructure problem.

In August, Microsoft Research published “Power Stabilization for AI Training Datacenters.”

The paper focused on AI training jobs spanning tens of thousands of GPUs and the unusual electrical behavior created when large synchronized workloads move repeatedly between compute-heavy and communication-heavy phases. Those shifts can produce substantial and coordinated changes in power consumption across a large training cluster.

Microsoft treated this as a cross-stack engineering problem, evaluating techniques across software, GPU-level controls and data-center infrastructure. The researchers tested approaches using real hardware together with Microsoft’s own cloud power simulator and advocated a combined strategy in which software and GPU controls work together with energy storage to deliver a smoother electrical profile to the utility.

This research sits on a parallel track to Heron.

Heron asked how AI inference could operate effectively when the available renewable power changes. Power Stabilization asked how the AI workload itself could be managed around rapid changes in electrical demand.

The two directions meet at the same interface between computing and power: the compute system adapts to the energy source, while the energy infrastructure is supported by increasingly intelligent control of the compute load.

That becomes particularly relevant later when we return to Soluna’s work with Siemens on AI power behavior.

By 2025, power had clearly become part of Microsoft’s AI systems design.

2026: XWind becomes CWind

The next major step arrived in May 2026.

On May 22, 2026, Microsoft researchers submitted a new version of the AI Greenferencing architecture under the title:

“XWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy Farms.”

XWind and CWind are not two separate Microsoft papers. They are two versions of the same work.

The original arXiv submission on May 22 used the name XWind. On July 21, 2026, the authors submitted version 2, changing the title and system name to:

“CWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy Farms.”

Microsoft Research’s publication page still preserves the earlier XWind title, while the current arXiv version is CWind.

For the chronology, the progression is therefore XWind v1 in May 2026 → CWind v2 in July 2026, with the same underlying research line.

CWind turns the idea into a hyperscale opportunity

CWind begins with the same core concept introduced by Heron but pushes it considerably further.

The paper proposes a regionally distributed AI infrastructure model in which modular AI compute is co-located behind the meter at wind farms. Microsoft frames the architecture as a way to expand available AI capacity by creating local computing demand at renewable sites and making use of energy where it is generated.

The feasibility analysis also grows substantially from the 2025 Heron paper. CWind identifies more than 890 GW of operating and under-construction wind capacity at large 100+ MW wind farms within 50 milliseconds of Azure data centers, with 73% within 20 milliseconds. The researchers estimate that the available resource could support the equivalent of more than ten million H100 GPUs under their right-sizing assumptions.

This is not a niche edge-computing proposal. Microsoft researchers are modeling wind-farm co-location as infrastructure capable of supporting AI at enormous aggregate scale.

CWind was evaluated on a real 64-GPU NVIDIA A100 testbed emulating three wind-powered sites, using Azure production traces for inference workloads. The system combines local control over active compute capacity and GPU operating frequency with cross-site routing based on real-time telemetry such as queue depth, latency and KV-cache utilization.

The result is a computing fleet designed to respond directly to the electrical state of geographically separated renewable assets.

CWind is not a “wind-only” architecture

One of the most important details in CWind is easy to miss if the paper is reduced to the phrase “AI at wind farms.”

Microsoft’s architecture allows renewable sites to be supported by batteries, complementary renewable resources and opportunistic grid draw, while software handles the remaining variability through site-level controls and cross-site routing.

Geographically separated wind farms can also complement one another because their production profiles do not move in perfect synchronization. Conventional hyperscale data centers remain part of the wider architecture as well, operating alongside Greeninferencing sites rather than being replaced by them.

The model is therefore much broader than:

wind turbine → GPU.

It looks more like renewable generation, right-sized local compute, energy storage or complementary supply, grid interaction, GPU-level reconfiguration and cross-site workload orchestration operating together.

The similarity to the direction Soluna would later describe for large-scale AI infrastructure is increasingly difficult to miss.

Soluna appears again

The second explicit Soluna reference makes the timeline even more interesting.

In the current CWind paper, Microsoft again names Soluna, alongside windCORES and WestfalenWIND, as a company already deploying computing at wind farms.

Microsoft’s renewable-AI research line had therefore cited Soluna in Heron in 2025 and then cited the company again as the work evolved into XWind/CWind in 2026.

The commercial transition described by Microsoft is also remarkably close to Soluna’s own direction.

Existing co-located operators had historically focused on workloads such as cryptocurrency and other flexible computing. Microsoft’s proposed next step was high-value LLM inference.

Soluna was making the same transition at Kati: from a first phase centered on Bitcoin hosting toward a second phase built specifically around AI and HPC.

Microsoft is already working with a large renewable-energy company

CWind then adds one of the most intriguing disclosures in the entire research line.

In its summary of contributions, the Microsoft researchers state that they are “working closely with a large renewable energy company that sees significant value in this strategy.”

That sentence moves Greeninferencing closer to the commercial energy industry. Microsoft is not only modeling the concept internally; its researchers are already discussing the strategy directly with a large renewable company that sees value in the same behind-the-meter AI architecture.

And the surrounding details make EDF — potentially through the broader EDF–Masdar relationship — an especially compelling candidate for that unnamed partner.

CWind’s hardware evaluation is built around power profiles derived from real-world wind data from three wind-farm sites in the Central United States. Microsoft selected the three sites specifically because their generation profiles provide geographic complementarity, then scaled those real wind traces to the 64-GPU testbed.

That three-site structure stands out when placed beside the renewable portfolio already surrounding Project Kati.

When Masdar entered EDF Renewables’ 1.6 GW U.S. portfolio, the wind portion consisted of exactly three utility-scale projects in Texas and Nebraska totaling 815 MW: Coyote Wind, 243 MW in Scurry County, Texas; Las Majadas, 273 MW in Willacy County, Texas; and Milligan 1, 300 MW in Saline County, Nebraska.

The geographic fit is striking. CWind uses three real Central U.S. wind sites. EDF and Masdar share a three-project wind portfolio concentrated in Texas and Nebraska.

And one of those three assets is Las Majadas.

Las Majadas is already the wind farm where EDF and Masdar brought Soluna into a behind-the-meter computing structure, where Project Kati was developed, and where Kati 1 moved into actual operation.

The same CWind paper that describes an unnamed large renewable-energy collaborator also explicitly points to Soluna as a real-world compute-at-wind operator.

That creates an unusually tight chain of overlapping facts.

Microsoft Research is developing behind-the-meter AI infrastructure at wind farms. It is working directly with a large renewable-energy company. Its experimental work draws on three real Central U.S. wind sites. EDF and Masdar jointly hold a highly relevant three-wind-project portfolio in the same broad geography. One of those projects is Las Majadas, already home to Soluna’s Kati platform. And Soluna itself is repeatedly cited inside Microsoft’s Greeninferencing research.

EDF therefore emerges as one of the strongest candidates for CWind’s unnamed renewable-energy relationship, with the broader EDF–Masdar ecosystem making the connection even more compelling.

The fit is not based only on the number of wind farms. EDF had already spent years commercializing essentially the type of energy relationship Microsoft was now studying. Project Tumbleweed had sought behind-the-meter data-center partners in ERCOT. Soluna had emerged from that process and developed Kati at Las Majadas. EDF’s early co-location work then expanded into a dedicated Data Center Solutions strategy.

Masdar strengthens the same connection from another direction. It co-owns Las Majadas with EDF and, as the next stages of this research will show, has built its own increasingly direct relationship with Microsoft around renewable energy, AI and data-center infrastructure.

The result is a chain that deserves to be read together:

Microsoft Research → AI compute behind the meter at wind farms → unnamed large renewable-energy collaborator → three Central U.S. wind sites → EDF/Masdar’s three-project wind portfolio → Las Majadas → Soluna → Project Kati.

The individual CWind traces do not need to be conclusively matched to Coyote, Las Majadas and Milligan for that broader alignment to remain important. The portfolio provides a highly specific physical match to the kind of multi-site wind ecosystem CWind is modeling, while EDF’s commercial history provides an equally strong match to the kind of renewable partner Microsoft says it is already working with.

By July 2026, Microsoft Research had therefore reached an unusually interesting position: it had cited Soluna repeatedly, built and tested a behind-the-meter AI-at-wind architecture, and disclosed active work with a large renewable-energy company whose strongest surrounding candidate sits inside the same EDF–Masdar ecosystem already connected to Kati.

Another 2026 paper pushes toward the hybrid microgrid model

A separate Microsoft Research publication from May 2026 broadens the architecture in another direction.

“Carbon-Aware Compute–Power Scheduling for AI Data Centers with Microgrid Prosumer Operations” studies AI data centers as integrated compute-and-energy systems rather than treating the electrical side and workload side independently.

The model jointly considers training workloads, inference routing, local generation, battery storage, electricity procurement and bidirectional interaction with the grid.

The researchers find particular value in flexible inference routing, battery storage and sites with substantial local generation.

Its importance is that it points in the same broader direction inside Microsoft Research.

The AI data center increasingly becomes a prosumer energy system: it can consume grid power, use local generation, operate storage, move flexible workloads and coordinate those resources through software.

That is very close to the kind of integrated energy-and-compute architecture that large behind-the-meter AI projects are now trying to build commercially.

By 2026, the research line is no longer subtle

Taken chronologically, the development is striking.

In 2021, Virtual Battery proposed putting data centers beside renewable-energy farms and shifting computational demand toward available power.

In 2024, SkyBox studied how modular renewable-powered data centers could be selected and coordinated across geographic regions using real generation traces and workload migration.

In 2024 and 2025, Microsoft’s broader AI systems work increasingly treated GPU power as something that could be actively managed through software, scheduling and hardware controls.

In May 2025, Heron applied the renewable-co-location concept specifically to AI inference at wind farms, used Azure production traces, estimated millions of GPUs could be supported under the model and explicitly cited Soluna.

In August 2025, Power Stabilization addressed the other side of the interface: how massive AI workloads themselves can be shaped so that their power behavior works more smoothly with electrical infrastructure.

In May 2026, XWind expanded the wind-farm AI architecture and tested it on real GPU hardware.

In July 2026, the revised CWind paper increased the identified opportunity to more than 890 GW of wind capacity, modeled more than ten million H100-equivalent GPUs, incorporated batteries, complementary generation and grid interaction into the architecture, cited Soluna again and disclosed that the researchers were already working closely with a large renewable-energy company.

And behind that final disclosure sits another potentially important alignment: three real Central U.S. wind sites, an EDF–Masdar portfolio containing three major Texas/Nebraska wind farms, Las Majadas among them, and Soluna already operating inside that exact asset ecosystem.

That progression matters far more than any one paper viewed alone.

Microsoft did not publish one isolated academic experiment that happened to resemble Soluna.

It developed a multi-year research line in which compute progressively moves closer to renewable generation, becomes more responsive to power availability, becomes specifically AI-focused, and eventually reaches a behind-the-meter wind architecture in which Soluna itself is repeatedly cited as a real-world example.

At the same time, Soluna was moving in the opposite direction.

It had started with renewable generation and flexible computing. Then it secured Las Majadas. Kati moved from contract to construction. Kati 1 demonstrated the behind-the-meter model at utility scale. And the second half of the site was beginning its transformation into a much larger AI project.

By 2026, the two paths were converging around the same type of infrastructure — and increasingly around the same energy ecosystem.

The next part of the story takes that convergence out of the research lab.

Because while Microsoft researchers were studying how AI could move closer to renewable power, Microsoft itself was building an enormous real-world organization around AI data centers, power procurement, interconnection and energy infrastructure.

6. Microsoft Leaves the Lab — Power Becomes the Infrastructure

The Microsoft research trail matters because it shows how the company has been thinking about the relationship between computing and energy. But Microsoft is not approaching that problem only through academic work.

Outside Microsoft Research sits one of the largest physical infrastructure organizations in the technology industry. Cloud Operations + Innovation — CO+I — designs, builds and operates Microsoft’s global data-center infrastructure, coordinating the physical systems required to turn cloud demand into operating capacity. Microsoft describes the organization as responsible for the physical infrastructure behind its cloud, including supply planning, construction, operations, hardware, security and power. By 2026, Noelle Walsh, President of Cloud Operations + Innovation, was overseeing the organization that Microsoft says “powers the global Microsoft Cloud.”

That distinction is important when looking at the Microsoft–Soluna thesis. The Greeninferencing papers show what Microsoft’s researchers believe can be done with flexible AI computing and renewable generation. CO+I represents the organization that must solve the same problem in the physical world: find sites, secure enormous quantities of electricity, build electrical infrastructure, construct data centers, commission them and turn that power into usable AI capacity.

And Microsoft has been working directly with renewable-energy developers for much longer than the current AI boom.

Microsoft and EDF go back to 2014

One of the earliest connections is surprisingly direct.

On July 15, 2014, Microsoft announced a 20-year PPA for the Pilot Hill Wind Project in Illinois. The project was then approximately 175 MW and sat on the same electrical grid that supplied Microsoft’s Chicago-area data center. The renewable developer on the other side of that agreement was EDF Renewable Energy.

EDF described the project in equally direct terms. It had acquired a 96% interest in Pilot Hill, and the project benefited from a 20-year PPA with Microsoft Corporation. Construction began that year, with commercial operations following in 2015. EDF later said Microsoft’s long-term commitment had helped provide the revenue certainty needed to advance the wind project into construction.

The relationship is notable because it predates Project Kati by almost a decade.

Long before EDF’s Asset Optimization team began searching for behind-the-meter data-center partners through Tumbleweed, EDF was already supplying wind power into Microsoft’s data-center energy strategy.

Microsoft’s own description of Pilot Hill was also revealing. Paul Scanlan, then part of Microsoft’s Energy Strategy organization, said the company was focused on transforming the energy supply chain for cloud services “from the power plant to the computer chip.”

That phrase was written in 2014.

Years later, Microsoft’s researchers would study modular compute located beside renewable-energy farms. EDF would search for data centers capable of co-locating beside its generation. Soluna would win its way into Las Majadas. And Microsoft Research would begin citing Soluna as an example of the resulting compute-at-wind model.

The companies involved in the later Kati story were therefore not entering completely separate worlds.

Microsoft and EDF already had a history of connecting renewable generation to data-center demand.

EDF and Masdar were already building power solutions for data centers

The next institutional bridge appeared before Kati as well.

In October 2022, Emerge — the energy-services joint venture between Masdar and EDF — signed an agreement with Khazna Data Centers to develop a ground-mounted solar facility for a new hyperscale data center in Masdar City.

The project was relatively small compared with Kati, at 7 MWp of solar, but its structure is revealing. Emerge was responsible for the entire energy project: financing, design, procurement, construction, operation and maintenance under a 25-year arrangement. The power would support Khazna’s AUH6 data center.

AUH6 subsequently emerged as a 31.8 MW IT-capacity, AI-ready hyperscale data center, with Khazna describing it as infrastructure designed for AI-related data development and hosting.

This adds another layer to the institutional history.

By 2022, EDF and Masdar were not only joint owners of renewable assets such as Las Majadas. Through Emerge, they were already combining energy development with hyperscale data-center infrastructure.

Then EDF launched Tumbleweed.

Then Soluna was selected.

Then Kati followed.

The progression shows that renewable power for digital infrastructure was already an established area of interest inside the EDF–Masdar relationship before Project Kati became public.

AI changes the scale of Microsoft’s power requirement

What changed dramatically after 2023 was the size of the computing infrastructure Microsoft needed to power.

The clearest example is Fairwater.

In September 2025, Microsoft unveiled its Fairwater AI data center in Mount Pleasant, Wisconsin, describing it as the most sophisticated AI facility it had built. The campus covers approximately 315 acres, with three large buildings totaling roughly 1.2 million square feet. Its electrical build includes around 120 miles of medium-voltage underground cable, while the computing architecture is designed to connect hundreds of thousands of advanced NVIDIA GPUs into a single massive AI system.

Fairwater was not designed as an isolated facility. Microsoft said multiple facilities using the same architecture were under construction across the United States, with Fairwater sites eventually interconnected as part of a distributed AI “superfactory.” The company later described the Wisconsin and Atlanta sites as parts of a dedicated network capable of working together on large AI workloads across geographic locations.

The similarity to the direction of Microsoft’s research is notable. At the research level, workloads can move and coordinate across geographically separated power-aware compute sites. At the hyperscale infrastructure level, Microsoft is simultaneously constructing geographically distributed AI facilities designed to function together as one computing system.

But the most important figure for the Kati story is the power requirement.

During Microsoft’s fiscal 2026 first-quarter earnings call, Satya Nadella said the company expected to increase total AI capacity by more than 80% during the year and roughly double its overall data-center footprint over two years.

Then he gave Fairwater’s scale:

the Wisconsin campus alone was expected to scale to 2 GW.

A 2 GW AI campus changes the way energy has to be treated.

Power is no longer a utility contract negotiated after the data-center design is finished. It becomes one of the central constraints determining where, when and how AI capacity can be built.

By April 2026, Fairwater had already begun operating ahead of schedule. Microsoft told investors during its fiscal third-quarter call that the company had brought another gigawatt of capacity online during that quarter alone. In June 2026, Microsoft confirmed that the first Mount Pleasant facility was fully operational, while construction continued on a second facility beside it.

This is the scale of infrastructure against which Kati 2 has to be understood.

A 100 MW first AI phase is substantial for Soluna.

For Microsoft, hundreds of megawatts increasingly represent building blocks inside a multi-gigawatt expansion.

Microsoft now treats electricity as part of the development pipeline

Microsoft’s own statements show how deeply energy procurement has moved into the data-center development process.

In January 2026, Brad Smith described a policy in which Microsoft works with utilities early, provides visibility into future power requirements, contracts for the electricity its data centers will require and helps fund supporting infrastructure. When Microsoft projects require additional transmission or substation capability, Smith said the company follows an existing practice of paying for those improvements.

Microsoft also disclosed an extraordinary number for the Midcontinent Independent System Operator region.

It had contracted to add 7.9 GW of new electricity generation in MISO, which Microsoft said was more than twice its existing consumption there.

A month later, Microsoft provided an even broader view of the portfolio. By February 2026, the company said it had contracted 40 GW of new renewable energy globally over more than a decade of procurement. CO+I President Noelle Walsh joined Microsoft’s Chief Sustainability Officer in announcing the milestone.

Those figures place Microsoft’s energy organization in a different category from a conventional corporate procurement function.

It is effectively part of the infrastructure-development machine.

Microsoft needs land, transmission, substations, generation, long-duration supply agreements and grid interconnection. Increasingly, it needs those elements to move on timelines dictated by AI demand rather than traditional utility development cycles.

That makes architectures capable of accessing power differently particularly valuable.

Behind-the-meter power does not eliminate the need for the grid. Kati 2 itself is evolving toward a much broader hybrid power architecture. But the ability to locate a large computing load beside existing generation creates another route to megawatts — exactly the opportunity Soluna and EDF had been developing at Las Majadas.

Microsoft is directly involved in the large-load power debate

The depth of Microsoft’s involvement is also visible in the power markets themselves.

In August 2025, Microsoft submitted formal comments to PJM Interconnection regarding the rapidly growing problem of connecting large new electrical loads.

The letter is notable for how Microsoft describes its own priorities. It says the company constructs, owns and operates data centers globally and that obtaining reliable, firm, sustainable and cost-effective energy is paramount to those operations as cloud and AI demand grows. Microsoft also said it was working with utilities on grid planning and helping bring additional generation into the system.

The document was signed by Jeff Riles, Director of Energy Markets, Americas at Microsoft.

That is a useful window into the real organization behind Microsoft’s AI expansion.

The company does not simply send a power bill to an accounting department after a data center is completed. It has people working directly on energy markets, interconnection structures, generation procurement and the rules governing how massive new computing loads can connect to the electrical system.

The PJM filing also points to Microsoft’s 20-year agreement with Constellation for the 835 MW Crane Clean Energy Center as one example of its effort to bring new firm generation capacity into the system.

This gives the Microsoft Research work a different context.

Greeninferencing is being developed inside a company whose operating organization is simultaneously attempting to secure tens of gigawatts of power and solve large-load interconnection at real data-center sites.

The research question and the commercial infrastructure problem are moving together.

Then Microsoft and Masdar connect directly

Against that backdrop, the Masdar relationship becomes considerably more significant.

On October 31, 2024, ADNOC, Masdar and Microsoft jointly published “Powering Possible: AI and Energy for a Sustainable Future.” The report dealt directly with the growing relationship between AI, data-center electricity demand and energy infrastructure. The three organizations argued for much deeper collaboration between the technology and energy sectors as AI expands.

Five days later, the relationship moved beyond a joint report.

On November 5, 2024, ADNOC and Masdar announced a Strategic Collaboration Agreement with Microsoft covering AI and low-carbon initiatives in the UAE and globally.

One part of the agreement was particularly direct:

the companies would evaluate opportunities to power Microsoft’s data centers with renewable energy through Masdar.

That sentence creates one of the strongest institutional links in the entire Kati research.

Masdar was not merely participating in the same conference as Microsoft.

It was not simply providing renewable certificates into a broad market.

Microsoft and Masdar had formally agreed to evaluate Masdar renewable-energy projects for Microsoft data centers.

And this happened while Masdar already owned 50% of Las Majadas with EDF.

The same CEO sits on both sides of the timeline

The continuity becomes even stronger when the people are added.

Mohamed Jameel Al Ramahi has served as Masdar’s CEO since 2016.

He was therefore already leading Masdar when the company agreed in 2020 to acquire interests in EDF’s U.S. renewable portfolio and when that transaction advanced in 2021. The portfolio included Coyote, Milligan 1 and Las Majadas. When the first closing was announced, Al Ramahi personally described the transaction as a significant milestone in Masdar’s collaboration with EDF Renewables North America.

That same executive was still leading Masdar when its relationship with Microsoft became explicit in 2024.

In the November 2024 Microsoft collaboration announcement, Al Ramahi spoke directly about delivering clean energy to the data centers that would power the AI future. The agreement itself contemplated renewable-energy supply to Microsoft data centers through Masdar.

Then, only months later, Al Ramahi publicly turned his attention to Las Majadas and Soluna.

Following the March 2025 public announcement of the EDF–Masdar–Soluna agreement, Al Ramahi described rising demand for advanced computing and AI and said renewable energy had become increasingly important to powering data centers. He highlighted the Las Majadas arrangement specifically: Masdar and EDF supplying renewable power to Soluna’s co-located Texas data center, while directly connecting the wind farm to the computing facility.

His framing is worth paying attention to.

He did not present Kati simply as a Bitcoin-power agreement.

He placed it inside the growth of advanced computing, AI and data-center energy demand.

The executive chain is therefore unusually clean:

2020–2021: Al Ramahi leads Masdar as it enters the EDF U.S. portfolio containing Las Majadas.

2024: Al Ramahi leads Masdar as it enters a strategic collaboration with Microsoft that explicitly evaluates renewable power for Microsoft data centers.

March 2025: Al Ramahi personally highlights the Soluna/Las Majadas co-location model in the context of advanced computing, AI and data-center power.

The same person sits across the entire sequence.

Microsoft and Masdar deepen the relationship again

The Microsoft–Masdar relationship did not stop with the 2024 agreement.

On November 2, 2025, ADNOC, Masdar, XRG and Microsoft announced a further strategic agreement built around what they called “AI for Energy and Energy for AI.”

This time, the energy objective became even broader.

The agreement brought Masdar and XRG into a collaboration to develop energy projects and infrastructure in support of Microsoft’s global AI and data-center expansion.

That language is materially stronger than a general sustainability partnership.

The objective was infrastructure for Microsoft’s global AI growth.

Masdar CEO Mohamed Jameel Al Ramahi later described himself as one of the signatories. In his account of the event, he said he signed the strategic collaboration agreement between ADNOC, Masdar, XRG and Microsoft in the presence of Microsoft Vice Chair and President Brad Smith, and described the agreement as a way to deliver energy solutions for Microsoft’s global AI and data-center growth.

The progression from 2024 to 2025 is therefore clear.

In 2024:

evaluate opportunities to power Microsoft data centers through Masdar.

In 2025:

develop energy projects and infrastructure supporting Microsoft’s global AI and data-center expansion.

And throughout that period, Masdar remained directly connected to Las Majadas and Project Kati.

The Masdar bridge runs through Las Majadas

This is why the Masdar relationship deserves more weight than a conventional corporate partnership.

Masdar is not merely a renewable-energy company that happens to know Microsoft.

It is already embedded in the physical asset at the center of Project Kati.

Through its EDF transaction, Masdar became a 50% economic participant in Las Majadas. Its executives remained involved as the project moved into operations. When EDF and Soluna developed the behind-the-meter structure, Masdar became a counterparty to the eventual power agreement.

At the same time, Masdar was building a direct strategic relationship with Microsoft around exactly the issue Kati 2 is designed to address:

energy for AI data centers.

The overlap operates in both directions.

From the Kati side:

Las Majadas → EDF + Masdar → Soluna → behind-the-meter computing → Kati 2 AI/HPC.

From the Microsoft side:

Microsoft AI expansion → enormous new power requirements → renewable-energy partnerships → Masdar → energy infrastructure for Microsoft data centers.

Masdar sits in the middle of both chains.

And Mohamed Jameel Al Ramahi provides executive continuity across them.

EDF creates a second route into the same Microsoft ecosystem

Masdar is not the only route.

EDF provides another.

Microsoft has had a direct renewable-power relationship with EDF since Pilot Hill in 2014. EDF and Masdar subsequently became long-term partners across multiple renewable markets. Together they owned Las Majadas. Together they operated Emerge, which was already delivering renewable-power infrastructure to hyperscale data centers by 2022.

Then EDF ran Project Tumbleweed.

Then Soluna emerged from the partner-selection process.

Then EDF and Masdar signed the 166 MW Kati PPA.

Then Microsoft Research began citing Soluna in its wind-powered AI work.

Then Microsoft disclosed that it was working closely with an unnamed large renewable-energy company on Greeninferencing.

Then Microsoft and Masdar expanded their own relationship into energy infrastructure for Microsoft’s global AI growth.

These are separate events, but they continue to populate the same surprisingly small institutional network.

Microsoft. EDF. Masdar. Soluna. Wind generation. Data centers. AI.

The names keep returning.

Power availability explains why this matters

Microsoft’s own infrastructure expansion makes the commercial logic increasingly easy to understand.

The company is trying to expand AI capacity on a scale measured in gigawatts. Fairwater alone is designed around a multi-gigawatt trajectory. Microsoft has contracted tens of gigawatts of renewable generation, is funding grid infrastructure, is participating directly in transmission and interconnection debates, and is partnering with major energy companies to develop the next generation of supply.

At that scale, power becomes a portfolio problem.

No single solution is sufficient.

New utility generation matters.

Transmission matters.

Nuclear matters.

Renewable PPAs matter.

Storage matters.

Grid optimization matters.

And locations where large quantities of existing generation can be paired directly with new computing demand become strategically interesting as well.

That is the category Project Kati occupies.

Las Majadas already exists.

The wind generation already exists.

The substation already exists and is now being expanded.

The first computing phase has already demonstrated the behind-the-meter relationship.

And the second phase is being rebuilt for AI at a scale far beyond its original 83 MW design.

This is where Microsoft’s research, its real-world infrastructure expansion and its energy partnerships begin reinforcing one another.

The research says AI compute can move closer to renewable generation.

The infrastructure organization needs enormous new blocks of power.

EDF has already shown Microsoft it can supply renewable energy to data centers.

Masdar has formally agreed to develop renewable-energy and infrastructure opportunities for Microsoft’s data-center expansion.

EDF and Masdar together own the wind asset underneath Kati.

And both Microsoft Research and Masdar’s CEO have independently placed Soluna inside the emerging AI-and-energy story.

By itself, each connection has its own history.

Together, they form an institutional network around Kati that is becoming increasingly difficult to ignore.

And one final event would soon make that network far more personal.

By 2026, Microsoft had built Fairwater into the centerpiece of its new AI infrastructure generation. Inside Microsoft, teams were coordinating site development, construction, energy, engineering, procurement, commissioning and operations on precisely this class of hyperscale AI campus.

Then, in July 2026, one of the people who had been working inside that Microsoft machine — including on Fairwater itself — left Microsoft and joined Soluna.

His new responsibility was to build Soluna’s AI infrastructure.

His name was Ryan Carver.

7. Ryan Carver — The Microsoft-to-Soluna Bridge

Up to this point, the Microsoft connection had largely been institutional. Microsoft Research had spent years developing architectures for computing directly at renewable-energy farms and had begun citing Soluna inside that work. Microsoft had a long-standing renewable-power relationship with EDF, while Masdar had become a direct Microsoft energy partner even as it remained a co-owner of Las Majadas. Project Kati was developing inside that same EDF–Masdar renewable ecosystem.

In July 2026, that connection became personal.

On July 16, 2026, Soluna appointed Ryan Carver as Chief Development Officer. He joined directly from Microsoft after more than a decade inside its data-center organization, most recently serving as Senior Director of AI Construction & Site Development. At Soluna, his mandate covered essentially the entire lifecycle of the company’s AI/HPC platform: site origination, power procurement, design, construction, commissioning and operations.

This was much broader than a conventional construction appointment. Soluna placed development, construction, technology operations and power within a cross-functional organization led by Carver and reporting directly to CEO John Belizaire. The objective was to create a unified delivery engine capable of taking Soluna’s behind-the-meter AI projects from energy opportunity to operating data center.

Carver arrived from almost exactly the other side of the infrastructure equation Soluna was trying to solve. He knew how Microsoft develops hyperscale AI campuses, how those projects move through land, energy, design, procurement and construction, and what has to happen before hundreds of megawatts can become usable computing capacity.

Now his job was to build Soluna’s.

More than a decade inside Microsoft’s data-center machine

Carver’s Microsoft experience began well before the current generative-AI buildout.

A 2019 Ohio State University profile identified him as Microsoft’s Global Director of Data Center Engineering & Construction Strategy. His work included strategic projects, prototype data-center designs, new form factors and engineering and construction innovations for Microsoft’s cloud infrastructure.

That history matters because it shows that Carver’s experience was not limited to executing an established data-center template. He had spent years working on how Microsoft itself could change the way physical cloud infrastructure was designed and delivered.

His public project history describes that progression at increasingly large scale. It includes Microsoft’s data-center electrical retrofit programs, new delivery models and later AI Data Center Development Mega Projects, including developments ranging from smaller facilities to multiple 340 MW projects running in parallel. His responsibilities extended across complex mechanical, electrical, telecom, civil and structural packages associated with those programs.

By the time AI began transforming Microsoft’s infrastructure requirements, Carver had already spent years inside the engineering and construction organization responsible for translating cloud growth into physical capacity. His later title — Senior Director of AI Construction & Site Development — placed that experience directly inside Microsoft’s new hyperscale AI expansion.

Soluna said that Carver led a construction P&L measured in the tens of billions of dollars across Microsoft’s AI data-center campus program and participated in decisions spanning site selection, permitting, design and delivery across multiple geographies.

The most important project in that period was Fairwater.

Fairwater puts Carver at the center of Microsoft’s AI buildout

Microsoft’s Fairwater campus in Mount Pleasant, Wisconsin provides scale to Carver’s title.

When Microsoft unveiled the facility in September 2025, it described Fairwater as the most sophisticated AI data center it had built. The campus covers approximately 315 acres, with three large buildings totaling about 1.2 million square feet. Its electrical infrastructure includes roughly 120 miles of medium-voltage underground cable, while the computing architecture is designed to connect hundreds of thousands of NVIDIA GPUs into a single massive AI system.

Fairwater was also designed as part of something larger. Microsoft began linking the Wisconsin campus with another Fairwater facility in Atlanta through its AI WAN, creating a geographically distributed AI “superfactory” capable of coordinating enormous workloads across sites. Satya Nadella later told investors that Fairwater Wisconsin alone was expected to scale to 2 GW.

Carver was directly involved in building that infrastructure. Soluna identifies Fairwater as the most notable program in his Microsoft portfolio, while Carver’s own public account says he led construction of the first of several buildings being developed concurrently at the Wisconsin campus. He described that first facility as a first-of-kind Microsoft design carrying the largest individual construction budget in the company’s history, and said it was delivered ahead of its commitments.

The internal organizations he worked alongside reveal just how broad that experience was. After leaving Microsoft, Carver specifically referenced teams spanning Data Center Engineering, Land Development, Contracts & Procurement, Cost, Planning, Networking, Energy, Site Selection, Community Engagement, Commissioning, Operations, Azure and Microsoft 365.

That list is effectively the hyperscale-development stack. A campus like Fairwater begins long before construction, with land, energy, permitting and site selection. It then moves through design, procurement and construction before commissioning, operations and Azure infrastructure turn it into usable AI capacity.

Carver had worked across that environment, and the experience maps remarkably well onto what Soluna was beginning to build at Kati 2.

From Microsoft to Soluna — because power comes first

When Carver explained why he joined Soluna, he centered the decision on power.

Soluna’s appointment announcement quoted him describing the co-location of digital infrastructure directly with renewable generation as “one of the most compelling approaches” he had seen for addressing the industry’s power constraint. During Soluna’s August 2026 earnings call, he expanded on that point, drawing directly from his experience inside hyperscale AI development.

His argument was simple: chips can be procured and buildings can be constructed, but the electricity required to operate them often moves on a much longer timeline. Utility planning, transmission, interconnection and generation development can take years. Soluna’s model offered another development path — locate large computing infrastructure where significant generation already exists and build around that power position.

Coming from someone who had just worked inside Microsoft’s multi-gigawatt Fairwater program, that perspective carries unusual relevance. Carver had spent years inside one of the largest AI infrastructure organizations in the world and had seen firsthand how central energy availability had become to development.

He then chose to move to a company whose operating thesis begins with exactly that issue:

secure the power first, then build the computing infrastructure around it.

His Microsoft mandate and his Soluna mandate nearly mirror each other

The scope of Carver’s new role makes the transition even more striking.

On his first Soluna earnings call, he described his Microsoft work as taking large AI campuses from a piece of land through power procurement, permitting, design, construction, commissioning and ultimately handover into operations.

Soluna’s description of his new responsibilities follows almost the same sequence:

site origination → power procurement → design → construction → commissioning → operations.

Carver therefore did not leave Microsoft for a broad corporate-management position. He moved from Microsoft’s AI construction and site-development organization into Soluna to lead essentially the same end-to-end infrastructure lifecycle under a different development model.

At Microsoft, those capabilities had been applied to one of the world’s largest hyperscale portfolios. At Soluna, they would be applied to renewable generation, behind-the-meter power and a growing pipeline of AI campuses.

At the center of that portfolio was Project Kati 2.

Four weeks after arriving, Carver is already deep inside Kati 2

The timing of Carver’s arrival is especially important.

He joined Soluna on July 16, 2026. Less than a month later, on August 13, he appeared on the company’s Q2 earnings call and personally walked investors through the development status of Kati and the rest of Soluna’s AI/HPC pipeline.

By then, Kati 2 had evolved far beyond the original second 83 MW phase contemplated under the 2024 power agreement. Carver described a campus expected to exceed 350 MW, located directly across from the operating Kati 1 facility. Phase I was designed for more than 100 MW of critical IT capacity, while Phase II was expected to add another 250 MW.

That language itself shows how far the project had progressed. The original Kati structure was generally discussed in terms of an 83 MW power allocation. The new Kati 2 was being described in hyperscale-development terms: critical IT capacity, electrical and mechanical topology, phased campus design, construction documents and long-lead procurement.

Those were precisely the disciplines Carver had spent years working with at Microsoft.

Design was moving toward construction

Earlier in 2026, Soluna had described Phase I at roughly 30% schematic design. By August, Carver said Kati 2 was approaching completion of design development, the stage where the electrical and mechanical systems become substantially defined before the project advances into construction documents.

That progression matters because it places Kati 2 much deeper into real engineering than an early conceptual campus plan.

Soluna had also brought the general contractor into the process while design was still advancing. That allows constructability, procurement and schedule considerations to be incorporated before the drawings are complete rather than waiting until the project is handed over for construction.

Long-lead procurement had begun as well. Carver said commitments were already being made with key electrical-equipment suppliers for components capable of determining the actual delivery schedule of the campus. Drawing on his hyperscale experience, he emphasized that major electrical equipment has to be addressed early because procurement timelines can shape when a data center can ultimately reach operation.

This is where the value of bringing a Fairwater-scale executive into Soluna becomes very tangible. The expertise is not simply knowing what an AI data center looks like when finished. It is knowing which decisions have to be made long before construction reaches the site in order for that facility to turn on when the customer needs it.

Las Majadas itself is expanding with Kati

Carver’s first public Kati presentation also revealed one of the most important pieces of physical infrastructure in the current project.

Engineering was underway to expand the Las Majadas substation by an additional 100 MW in support of future Kati phases, with the upgrade expected in early 2027.

That is a major evolution in the history of the site.

Las Majadas had started as the 273 MW wind asset around which EDF and Masdar eventually brought Soluna as a behind-the-meter computing partner. The first 83 MW became Kati 1 and demonstrated that the relationship could operate physically.

Now the electrical infrastructure of Las Majadas itself was being expanded to accommodate the next stage.

This is what makes the current Kati 2 scale more concrete. A campus does not become larger simply because its announced capacity increases. The underlying high-voltage infrastructure has to grow alongside it.

At Kati, that growth now has a named physical object:

the Las Majadas substation.

And the expansion is being engineered while a former Microsoft AI construction executive leads Soluna’s development organization.

Kati 2 develops into a hybrid hyperscale power campus

The power architecture around Kati 2 was also becoming considerably more sophisticated.

Las Majadas wind remained the renewable foundation of the site, but Soluna was developing additional layers around it to create the power quality and availability expected by large AI customers. The campus could interact with ERCOT while remaining tied directly to behind-the-meter wind, and by August 2026 Soluna had also executed an access agreement with a natural-gas pipeline operator and begun engineering work around a lateral connection to the site.

Carver described onsite generation as part of the firming architecture capable of bringing the campus toward the availability requirements of an AI tenant.

The emerging Kati design therefore combined behind-the-meter wind, ERCOT/grid access, expanded substation infrastructure and planned onsite firm generation rather than relying on a single electrical source.

Soluna was also discussing a broader clustering strategy in which nearby generation assets could eventually support a common computing campus, expanding the power envelope beyond what one renewable project alone could provide.

This direction fits naturally with the infrastructure problem Carver had just left Microsoft to solve. Microsoft Research was exploring power-aware AI systems across renewable sites and hybrid resources. Microsoft’s physical data-center organization was securing generation and electrical infrastructure at multi-gigawatt scale. Carver had now moved directly from that environment into a campus designed around renewable generation, grid interaction, power firming and hyperscale AI load.

The land footprint grows with the power footprint

The campus was expanding physically at the same time.

Carver said another 150-acre parcel for future Kati 2 phases was under agreement and approaching execution. That additional land formed part of a broader development footprint already moving well beyond the original second 83 MW phase.

The relationship between land and power is straightforward at this scale. More critical IT capacity requires additional buildings. Those buildings require additional cooling and electrical infrastructure. New electrical infrastructure requires substations, generation equipment, utility corridors and room for future expansion.

Kati 2 was beginning to assemble all of those elements simultaneously.

The project was no longer being planned as a single additional block beside Kati 1. It was becoming a multi-phase AI campus with land and power being secured around a much larger long-term development plan.

A tenant was already moving through the process

That physical expansion was happening alongside an active commercial process.

By the August 2026 earnings call, Soluna confirmed that Kati 2 had a signed tenant letter of intent, while commercial terms and lease negotiations continued alongside the engineering work.

The timing is significant because development was not being held back until a final lease appeared. Design development was advancing, the general contractor was already engaged, long-lead electrical commitments were being made, substation engineering was underway, additional land was being secured and the firm-power architecture was progressing while tenant negotiations continued.

Those are the activities required to make a hyperscale site ready to move quickly once the commercial structure is finalized.

And Carver arrived precisely during that stage.

He did not join Soluna while Kati 2 existed mainly as a concept. He joined while a prospective tenant was already engaged and the project was moving through the design, procurement, power and land work necessary to turn a lease into operating infrastructure.

Carver understands the hyperscaler side of the table

Carver’s significance also extends beyond construction execution.

His Microsoft background crosses site selection, land development, energy, power procurement, engineering, permitting, contracts, procurement, construction, commissioning and operations. His Fairwater work also placed him alongside Azure, networking and the internal organizations that ultimately consume and operate Microsoft’s infrastructure.

That gives him experience with both the delivery side and the customer side of hyperscale development.

A large AI customer is not evaluating a project solely on a headline electricity price. It is evaluating electrical topology, resilience, design standards, construction schedule, equipment procurement, commissioning, operating capability and whether the entire development organization can deliver the campus to hyperscale expectations.

Carver spent more than a decade inside the type of organization making those assessments.

Now he sits on the developer side of that conversation.

That makes Ryan Carver a direct organizational bridge from Microsoft’s hyperscale infrastructure machine into Soluna’s AI platform.

The chronology keeps tightening

The timing of his move lands inside an already dense Microsoft–Kati timeline.

In 2024, Soluna signed its 166 MW power agreement with EDF and Masdar. During the same period, Microsoft and Masdar began formal collaboration around renewable energy for Microsoft data centers.

In 2025, Microsoft Research explicitly cited Soluna while developing AI inference architectures for wind farms. Microsoft and Masdar then deepened their relationship around energy infrastructure supporting Microsoft’s global AI and data-center expansion, while Microsoft’s Fairwater program became the flagship of its next-generation AI infrastructure.

During the first half of 2026, Kati 2 expanded from its original 83 MW concept toward a 350+ MW AI campus. Soluna and Metrobloks signed their definitive joint venture, design advanced, the development organization expanded, a general contractor entered the process and a tenant LOI was secured.

Then, on July 16, 2026, the executive who had been working inside Microsoft’s Fairwater buildout became responsible for Soluna’s AI/HPC development platform.

Four weeks later, he was publicly discussing Kati 2’s design development, equipment procurement, 100 MW Las Majadas substation expansion, gas access, additional land and tenant negotiations.

The Microsoft overlap had now moved through several distinct layers — research, renewable-energy relationships, the EDF–Masdar ecosystem and finally personnel.

With Ryan Carver, that connection entered Soluna itself.

From Fairwater to behind-the-meter AI

Fairwater represents one of the largest expressions of the conventional hyperscale development model: enormous purpose-built AI campuses, hundreds of thousands of GPUs, dedicated networking, major utility infrastructure and a multi-year chain connecting land and power to operating compute.

Carver helped build that model.

When he left Microsoft, he chose a company approaching the same AI infrastructure race from a different starting point. Soluna begins with access to generation and builds the data center around it. Its model is designed to shorten the distance — both physical and commercial — between large sources of electricity and large computing loads.

Carver’s own explanation for joining the company makes that attraction clear. After more than a decade inside hyperscale data-center development, he viewed access to power as one of the defining constraints on AI growth and saw renewable co-location as a compelling infrastructure solution.

That brings several strands of this research together in one person.

Microsoft Research had spent years moving compute closer to renewable generation and had begun citing Soluna as a real-world example. Microsoft had built direct energy relationships with both EDF historically and Masdar more recently. EDF and Masdar jointly owned Las Majadas, the wind asset underneath Kati. Soluna had already demonstrated the behind-the-meter model there through Kati 1.

And now the executive who helped lead construction inside Microsoft’s flagship Fairwater AI program was leading Soluna’s AI/HPC development organization.

At the center of his new portfolio sat Project Kati 2.

By the time Carver arrived, it was no longer simply the second 83 MW phase of the original Kati agreement. It had become a 350+ MW hyperscale AI campus, with land, power, engineering and commercial activity beginning to match that scale.

8. Kati 2 Becomes a Hyperscale AI Campus

When Soluna first signed the Kati power agreement in May 2024, the structure was straightforward: 166 MW divided into two 83 MW phases. Kati 1 would eventually become the Bitcoin-hosting side of the site, while Kati 2 was reserved for AI and high-performance computing. Even at that stage, the second phase represented a meaningful AI opportunity, but it was still fundamentally an 83 MW extension of the original Las Majadas power agreement.

By the end of 2025, that definition was already beginning to change. Soluna had completed a test fit for an initial Kati 2 data-center design capable of accommodating the latest NVIDIA GPUs, identified architecture and electrical-engineering firms, and reported interest from both neoclouds and hyperscalers. The project still appeared publicly as “83 MW+,” but its development work was beginning to reflect a different class of customer and a different scale of infrastructure.

What happened over the following eight months was one of the most important transformations in the entire Kati story. The second 83 MW phase did not simply get a little larger. It was re-planned as a purpose-built hyperscale AI campus exceeding 350 MW of critical IT capacity.

That change brought new land, a new development partner, a new legal structure, a seven-building campus plan, high-density GPU design, long-lead electrical procurement, onsite generation planning, additional substation capacity and a tenant process reaching formal commercial negotiations. Kati 2 was becoming a very different project from the one originally contemplated in 2024.

January 2026: Metrobloks changes the development model

The first major signal came on January 15, 2026, when Soluna announced a partnership with Metrobloks to co-develop Kati 2.

The initial announcement described 100+ MW of critical IT capacity for AI and HPC, with an expansion roadmap supporting more than 300 MW of total critical IT. This was an important change in terminology as well as scale. Soluna was no longer talking simply about how much renewable power could be delivered to the site. It was now describing the project in Critical IT — CIT — capacity, the measure associated with the actual computing load inside a data center.

That distinction matters because a 100 MW critical-IT facility requires substantially more total site infrastructure than 100 MW of gross electrical service. Cooling, pumps, electrical losses, auxiliary systems, networking and other facility loads sit around the IT load itself. Soluna has since discussed increasing the first-phase approval toward a 150 MW gross power envelope capable of supporting approximately 50–100 MW of CIT, illustrating the difference between power entering a campus and power ultimately reaching computing hardware.

Metrobloks brought a complementary capability to the project. Soluna contributed the power position, site control, electrical infrastructure and experience working around ERCOT and renewable generation. Metrobloks brought an AI-focused data-center development, design, leasing and operating platform. Soluna’s announcement said Metrobloks’ leadership team had collectively enabled more than 12 GW of global data-center capacity, and the initial structure assigned it a major role in design, development, customer engagement and pre-leasing.

The logic of the partnership was closely tied to the power problem already running through this research. Metrobloks CEO Ernest Popescu said customers were looking for power and capacity that could be delivered now rather than years into the future, and described Kati as unusual because the power position was both available and scalable. Soluna had spent years establishing the renewable relationship, interconnection framework and site position; Metrobloks could now build a hyperscale AI development platform on top of that foundation.

The January announcement also revealed that the parties were already working with EDF Power Solutions to source another 50 MW or more from Las Majadas for expansion. That is important because the move beyond 83 MW was not simply a marketing ambition. Soluna and Metrobloks were already returning to the underlying wind asset and looking for additional power around which the larger campus could be developed.

The project moves from 83 MW to seven buildings

Only weeks later, the scale became much clearer.

In February 2026, Soluna said it was engineering Kati 2 as a 350 MW Critical IT Tier 3 AI campus spread across seven buildings.

That was a fundamental change from the original Kati structure. An 83 MW AI phase could have been developed as one large power block beside Kati 1. A 350 MW critical-IT campus requires master planning across multiple buildings, multiple electrical systems, cooling architecture, redundant utility paths, major internal distribution and enough land to allow individual phases to be constructed without constraining the rest of the campus.

Soluna’s 2026 development plan contemplated up to seven approximately 50 MW data-center buildings, designed for high-density racks and current-generation NVIDIA hardware. The design included a hybrid liquid- and air-cooling architecture, reflecting the thermal demands associated with dense GPU systems rather than the relatively simple cooling architecture used for Bitcoin facilities.

The significance of the seven-building plan is not simply that seven is a larger number than two. It shows how Kati 2 had changed conceptually. Instead of expanding the existing Kati data center by another equivalent block, Soluna and Metrobloks were laying out a multi-building hyperscale campus capable of being delivered in stages and expanded around customer demand.

The first approximately 100 MW of critical IT became the anchor phase. The remaining capacity would be added through subsequent buildings and infrastructure expansions rather than waiting for the entire campus to be constructed at once. That is a recognizable hyperscale development model: establish a large master plan, secure land and power for the full campus, then deliver the compute capacity through phased buildings as the tenant grows.

More than 500 acres changes what Kati can become

A campus of that size needed a larger physical footprint.

During the first quarter of 2026, Soluna reported securing more than 500 additional acres for Kati 2. By the Q1 earnings discussion, management was explicitly connecting that land position to the expansion from the original Kati footprint into a campus exceeding 350 MW.

The land expansion is one of the more tangible pieces of evidence behind the change in scale. Data-center announcements can increase in megawatts long before the physical development catches up. Kati 2 was doing both. As the announced computing capacity increased, Soluna was simultaneously expanding the land base needed to support the buildings, electrical yards, cooling systems, generation equipment, roadways and future phases.

The later definitive JV filing makes the land structure even more concrete.

By June 3, 2026, the Kati 2 development had separate Phase I and Phase II LandCos. The Phase I entity already owned the Phase I property, while Soluna had entered into a purchase and sale agreement dated March 10, 2026 for the Phase II property and assigned those rights into the Kati 2 structure when the definitive JV was signed. Soluna committed approximately $19.08 million toward completing the Phase II land acquisition.

That sequence shows the campus being assembled in layers. Phase I was no longer merely a conceptual first building, while Phase II was not merely a future expansion placeholder. Separate land entities and an actual property-acquisition agreement had been created around both phases.

The public legal structure was beginning to mirror the physical master plan.

From concept design to a buildable AI facility

Engineering moved forward at the same time.

By the first-quarter earnings update, Soluna said the initial approximately 100 MW phase had reached 30% schematic design, with architecture and engineering firms selected.

That is the point where the project begins moving beyond a test fit or conceptual rendering. Schematic design establishes how the major systems fit together: building configuration, electrical topology, cooling approach, equipment locations and the broad relationship between the compute buildings and the power infrastructure.

The following months show a steady progression through that design process. In June, after the initial design package had been completed, Soluna launched an RFP covering Phase I’s design-development and construction work. By July, the company had selected a general contractor and was onboarding that contractor to lead the design-build process. Soluna also reported that detailed-design documentation had reached approximately 50% completion.

By August, Ryan Carver said the project was nearly through design development, the stage in which the electrical and mechanical topology becomes substantially defined before the project moves toward final construction documents. The general contractor was already participating while that work was being completed.

The progression is significant:

test fit → schematic design → initial design package → design-development RFP → general contractor selected → approximately 50% detailed design → design development nearing completion.

Over the course of months, the project was moving through recognizable pre-construction stages rather than remaining fixed at the concept level.

Long-lead procurement begins before the lease is complete

Soluna was also beginning to secure the equipment that can determine the construction schedule of an AI campus.

By Q1, procurement workstreams for long-lead data-center equipment were already underway, and Soluna had signed an LOI for long-lead gas-generation equipment. By July, the company said it had signed LOIs with key electrical-equipment providers while detailed design and tenant negotiations were continuing.

Ryan Carver highlighted this directly during the August call. From his experience building AI campuses, he described major electrical components as the items that can become the longest schedule drivers and therefore need to be reserved early rather than after every other element of the project is complete.

That approach says something about how Soluna was preparing Kati 2 commercially. The development strategy was not simply to wait until a tenant signed a final lease and then begin figuring out how to build the site. Soluna and Metrobloks were advancing design, bringing in the contractor and reserving long-lead equipment while the tenant process continued.

For a customer trying to secure hundreds of megawatts of AI capacity, that preparation can materially change the delivery schedule. It is exactly why the original Metrobloks announcement emphasized speed to power.

June 3 turns the partnership into a definitive legal structure

The next major milestone came on June 3, 2026.

The January Metrobloks arrangement had begun as an MOU. In June, Soluna and Metrobloks replaced that preliminary structure with a definitive joint venture agreement, creating Soluna MB KK II JVCo, LLC to develop the campus.

The SEC filing is particularly useful because it removes the ambiguity that existed in some of the earlier “300 MW+” descriptions.

The definitive agreement specifies:

Phase I — 100 MW of critical IT

Phase II — an additional 250 MW of critical IT

That produces 350 MW of critical IT capacity under the legal development structure itself.

This is stronger than an aspirational roadmap in an investor presentation. The 100 MW/250 MW split appears directly in the filed agreement governing the Kati 2 development entity.

The ownership structure also tells us how the parties divided the project. Soluna holds 100% of the Class A interests and serves as manager of the JV. Metrobloks holds the Class B interests. Soluna contributed the Phase I property, the contractual rights to acquire the Phase II property and approximately $3.5 million of previously funded operating expenses. It also committed the capital required to complete the Phase II property acquisition and up to an additional $2 million for certain project expenses.

The economic waterfall is also unusually specific. After capital contributions are returned, contributing members receive a 14% IRR; Soluna then receives $100,000 for each Gross PPA MW of the project, after which additional distributions are split 50/50 between Soluna and Metrobloks.

Those terms show a real development business being constructed around Kati 2. Land, development capital, management rights and future project economics were being allocated before the campus had even reached construction.

The 350 MW plan now existed not only in engineering materials, but in the corporate structure created to develop it.

The original Kati power position becomes a platform

This is where the importance of Kati 1 becomes clearer.

The original 2024 agreement gave Soluna 166 MW across two 83 MW phases. By 2026, the first 83 MW had become an operating computing site while the second phase was expanding far beyond its original allocation.

That means the original Las Majadas relationship had become a foundation rather than a fixed ceiling.

The site already had a renewable-energy counterparty, a behind-the-meter framework, ERCOT planning history, operating compute across the road at Kati 1, local construction activity and established site operations. Additional land and electrical infrastructure were now being built around the same location to support a much larger AI campus.

Metrobloks described this directly in January when it said Soluna had already done the difficult work of securing a renewable-powered site with room to grow. The partnership was designed to build on that position rather than recreate it elsewhere.

That is one of the major strategic advantages behind Kati 2.

The 350 MW campus did not begin as a search for an empty parcel followed by a search for power. It grew outward from a power position Soluna had spent years establishing.

The tenant universe broadens from neocloud to hyperscale

The commercial process was evolving alongside the physical project.

At the beginning of 2026, Soluna said the site was under a non-binding LOI from a potential neocloud tenant. During the spring, however, the company began describing a wider field. April disclosures said hyperscaler and neocloud interest, due diligence and discussions remained active, while the company continued evaluating additional generation and engineering the larger campus.

By June, the language had advanced again. Soluna said tenant due diligence was continuing with hyperscalers and neoclouds, and formal commercial negotiations had begun with at least one potential tenant. Then in July, the company disclosed that negotiations had progressed to a signed tenant LOI, with the parties working on final design, commercial terms and the lease agreement.

The progression is important because the development process and tenant process were moving together. The site was not merely being marketed against a conceptual campus. Prospective customers were evaluating a project whose land, design, contractor, equipment procurement and power expansion were advancing while negotiations took place.

Soluna has separately said that indications of interest for Kati 2 have involved parties conducting due diligence on power, construction schedule and fit with their infrastructure requirements, with the targeted initial customer contract centered around approximately 100 MW of critical IT.

That 100 MW number is worth keeping in mind. It is almost exactly the size of Phase I under the definitive JV.

The campus is therefore being structured so that the first major customer can potentially take an entire hyperscale-class phase, while another 250 MW remains available for expansion.

Kati 2 is being designed around the latest GPU generation

The technical design also shows what class of workload Soluna expects to serve.

Even before the Metrobloks partnership was announced, Soluna said the Kati 2 test fit could accommodate the latest NVIDIA GPUs. By early 2026, the master plan was being developed around seven high-density buildings with a mixture of liquid and air cooling.

Those details matter because modern AI campuses are increasingly constrained not simply by total megawatts but by power density.

A conventional data center can distribute its electrical load across a large number of relatively low-density racks. Modern AI systems concentrate enormous electrical and thermal loads into far smaller footprints. The building, cooling, electrical distribution and backup architecture therefore have to be designed around GPU clusters from the beginning.

Kati 2’s progression toward critical-IT measurements, high-density racks, current NVIDIA hardware and liquid-assisted cooling shows that the campus was being engineered specifically for that environment.

This was not Kati 1 with different servers installed inside it. Kati 2 was becoming purpose-built AI infrastructure.

The power system grows with the campus

As Kati 2 expanded, the power architecture expanded with it.

The first sign was the January effort to source another 50 MW+ from Las Majadas. By August, that work had developed into something more substantial: engineering was underway to expand the Las Majadas substation by another 100 MW specifically to support future Kati phases.

At the same time, Soluna was developing onsite firm-generation capability. Earlier plans contemplated approximately 50–180 MW of onsite natural-gas generation, supplemented by diesel backup and a utility-scale battery-energy-storage system. The broader campus strategy also contemplated renewable clustering — drawing from Las Majadas and other wind assets in the surrounding region as the campus grew.

By August, the gas component had advanced again. Soluna had executed a gas-pipeline access agreement, while engineering around the site connection was moving forward.

The result is increasingly recognizable as the power architecture of a hyperscale AI campus rather than a simple renewable-powered computing facility. Las Majadas provides the renewable anchor, ERCOT provides grid interaction, the substation is being expanded, onsite generation provides firming capacity, battery storage can provide stabilization and supplemental resilience, and additional regional renewable resources can enlarge the long-term power envelope.

The next chapter will go much deeper into that architecture, because it is one of the strongest physical clues in the entire Kati 2 story. But for the evolution of the campus itself, the important point is that every time Kati 2 grew in computing capacity, the underlying power plan grew with it.

The project is now operating on a hyperscale development timeline

By August 2026, the different development streams had begun to line up.

The definitive Metrobloks JV was in place, with 100 MW of Phase I CIT and another 250 MW in Phase II. More than 500 acres had been secured for expansion, separate Phase I and Phase II land structures existed, design development was nearing completion, a general contractor had joined the project, long-lead electrical equipment was being reserved, the Las Majadas substation was being engineered for another 100 MW, gas-pipeline access had been secured, and a potential tenant had progressed into a signed LOI and commercial lease negotiations.

This is what makes the 2026 transformation of Kati 2 so important.

The project did not grow from 83 MW to 350 MW only in an investor presentation. Its land, JV structure, engineering, procurement, construction organization, power infrastructure and commercial process all began expanding around the larger number.

That is the difference between announcing a bigger campus and actually reorganizing a development around one.

The scale now fits a different class of customer

The transformation also changes how Kati 2 should be viewed in the Microsoft thesis.

The original second 83 MW phase was already large enough to interest AI infrastructure companies, but the new configuration was substantially more ambitious. A 350 MW critical-IT campus with an initial 100 MW tenant block gives a hyperscaler the ability to enter through one major first phase and then expand by hundreds of additional megawatts without moving to another site.

The multi-building structure supports phased growth. The additional land allows the campus to expand physically. The power architecture is being enlarged around that growth. And the project is being designed for the same high-density GPU hardware that drives Microsoft’s own infrastructure requirements.

This is also happening at a location where the energy relationships were established years earlier. EDF had already selected Soluna through its behind-the-meter strategy. EDF and Masdar already owned Las Majadas. Masdar had already entered direct energy and infrastructure collaboration with Microsoft. Microsoft Research had already cited Soluna as an existing compute-at-wind operator. And Ryan Carver had now moved directly from Microsoft’s Fairwater program into leadership of Soluna’s AI-development organization.

Kati 2’s physical transformation therefore matters because it supplies the scale underneath those connections.

The institutional network was already there. By 2026, the project itself was growing large enough to fit it.

From an 83 MW second phase to a 350 MW platform

The full evolution can now be reconstructed clearly.

In May 2024, Kati was a 166 MW agreement split into two 83 MW phases. By September 2025, Kati 2 was still publicly described as an 83 MW AI/HPC expansion, although hyperscaler and neocloud interest was already emerging.

By January 2026, Metrobloks and Soluna had moved the first AI phase beyond 100 MW of critical IT, with a roadmap exceeding 300 MW. By February, the master plan had grown to a 350 MW Tier 3 AI campus across seven buildings.

During the spring, Soluna secured more than 500 acres, selected architecture and engineering firms, advanced Phase I to schematic design and began long-lead equipment procurement. On June 3, the definitive JV formally established 100 MW of Phase I critical IT plus 250 MW of Phase II critical IT.

By July and August, a general contractor was engaged, detailed design was well advanced, major electrical procurement was underway, the Las Majadas substation was being expanded, gas access had been secured and the tenant process had reached a signed LOI.

In less than a year, the second half of the original Kati project had been transformed into something far larger.

Kati 2 was no longer simply Phase Two of Kati. It had become a hyperscale AI platform built around the power position that Kati had created.

And that brings the investigation to the most physical part of the story.

A 350 MW AI campus is ultimately only as valuable as the electricity it can deliver reliably to the GPUs inside it. At Kati 2, that system is now being assembled from several layers at once: Las Majadas wind, ERCOT, an expanding substation, battery storage, onsite firm generation and gas infrastructure.

Understanding how those pieces fit together is essential.

It is also where the parallels with Microsoft’s own work on AI power stabilization — and Soluna’s work with Siemens — become much harder to overlook.

9. The Power Architecture — Turning Wind Into Hyperscale AI Power

The transformation of Kati 2 from an 83 MW second phase into a 350+ MW AI campus created a second challenge beyond land, buildings and GPUs: the power system had to grow with it.

That is where Kati 2 becomes more interesting than the description “a data center next to a wind farm” suggests. Soluna is not designing the campus around the assumption that Las Majadas alone must supply a perfectly flat 24/7 load. The architecture being assembled combines the wind farm with ERCOT grid access, an expanded substation, utility-scale battery storage and onsite firm generation, with each layer serving a different function.

By August 2026, Ryan Carver described the strategy as an integrated power system. The renewable plant provides the low-cost generation at the center of the site, the grid provides another path for electricity when conditions require it, and onsite generation provides the firming needed to deliver the availability contracted by an AI customer.

That distinction is central to understanding Kati 2.

Wind is the anchor. It is not the entire architecture.

Las Majadas remains the foundation

Everything still begins with Las Majadas.

The approximately 273 MW wind farm was the reason Soluna arrived at this particular location in the first place. EDF’s search for a behind-the-meter computing partner eventually produced the Kati relationship, and the first 83 MW became Kati 1. By 2026, that first phase had provided something especially valuable for the larger AI campus: a working demonstration that computing could actually be integrated with the Las Majadas energy infrastructure.

Kati 2 therefore begins with an unusual advantage. It is not a greenfield data-center developer trying to obtain hundreds of megawatts after choosing a site. The renewable asset, the existing interconnection and the commercial relationship around that power came first.

Soluna described the expanded campus in early 2026 as using a dual-fed power and clustering model, drawing from the wind farm and other sources as well as the grid. Management estimated the resulting power cost at approximately $43/MWh under the configuration it was then developing.

The economic logic is important, but so is the sequencing.

The company did not first decide where to build a 350 MW AI campus and then begin asking utilities where the electricity might come from. The campus grew around an existing power position that had already been commercially and physically developed.

At Kati, power created the site. The AI campus followed.

The grid is built into the model

Behind-the-meter can sometimes be misunderstood as meaning disconnected from the grid.

That is not how Soluna is designing Kati 2.

Carver explained the architecture directly in August 2026: the campus can draw electricity from the renewable plant or from the grid depending on what the system requires at a particular moment. In Soluna’s view, that dual access is part of what allows a renewable-linked project to serve a long-duration AI contract while retaining the economic advantages of being located beside generation.

This turns the grid into a complementary resource rather than the only source of electricity.

When Las Majadas is producing strongly, the data center has generation immediately beside it. When additional power is needed, the grid can become part of the supply architecture. When the broader system values the wind generation elsewhere, the site’s electrical and workload controls can respond accordingly.

That flexibility is closely related to the original Soluna thesis from years earlier: computing infrastructure becomes more valuable when it can operate around the characteristics of the power system rather than forcing the power system to behave as though the load never changes.

The difference with Kati 2 is scale.

Instead of applying the idea primarily to modular Bitcoin computing, Soluna is now designing it around high-density AI workloads and Tier 3-class availability.

The Las Majadas substation is being expanded by another 100 MW

One of the strongest pieces of physical evidence behind the expansion appeared during Soluna’s August 2026 earnings presentation.

Carver disclosed that engineering was underway to expand the Las Majadas substation by an additional 100 MW in support of future Kati phases, with the upgrades expected in 2027.

This is important because it places the expansion inside a named piece of electrical infrastructure.

The Las Majadas substation is not an abstract future power source. It is part of the existing wind-farm system around which Kati was originally developed. Kati 1 already demonstrates the relationship between generation and compute at the site, and the same substation is now being engineered for another substantial block of capacity.

That is how the move from an 83 MW second phase toward a 350 MW campus begins to show up physically.

Additional buildings require additional IT capacity. Additional IT capacity requires additional gross power. Additional gross power requires electrical equipment capable of moving it.

And Soluna is now expanding the electrical infrastructure at the wind farm itself.

The timing also aligns with the rest of the Kati 2 development. While tenant negotiations were underway, the general contractor was already involved, long-lead electrical commitments were being made and substation engineering was progressing in parallel. Carver presented those activities together as the pre-construction work required for the eventual lease and build.

The power system is therefore being developed alongside the tenant process, not after it.

The 350 MW campus requires more than the original wind allocation

The scale of the new campus makes the logic clear.

The original Kati PPA allocated 166 MW across Kati 1 and Kati 2. But the definitive Metrobloks JV now contemplates 100 MW of critical IT in Phase I and another 250 MW of critical IT in Phase II.

Critical IT is only part of the electrical demand of a data-center campus. Cooling, pumps, electrical losses and auxiliary systems add additional load around the computers themselves.

Soluna therefore began designing a power system that could expand beyond the original Las Majadas allocation.

In its 2025 year-end presentation, the company described four major components behind the larger Kati 2 plan: the phased AI campus itself, 50–180 MW of onsite natural-gas generation, a renewable-clustering strategy using additional regional generation, and utility-scale battery storage for grid stabilization and supplemental backup.

That is a materially different architecture from simply signing a larger wind PPA.

The strategy is to assemble different power resources around the compute campus and make them operate as one system.

Natural gas becomes the firming layer

The gas component began moving from planning toward execution during 2026.

By the Q1 update, Soluna had already signed an LOI for long-lead gas-generation equipment for Kati 2. Additional onsite generation options, including both gas and solar, continued to be evaluated during April, while negotiations around the initial gas-engine purchase progressed in May.

The earlier campus design contemplated 50 MW to 180 MW of gross onsite natural-gas generation for primary and backup purposes, with diesel generation providing another resiliency layer.

Then came a more important development.

By August 13, 2026, Carver disclosed that Soluna had executed an access agreement with a natural-gas pipeline operator and that engineering on the lateral connection to the Kati 2 site was underway. He directly connected that onsite generation to the level of availability an AI tenant would contract for.

That moves the gas strategy beyond a conceptual backup-generator plan.

A pipeline-access agreement means Soluna is working on the fuel-supply side of the onsite-generation architecture itself. Engineering a lateral to the campus creates the possibility of bringing continuous pipeline gas directly to generation equipment at the site rather than treating gas only as an emergency fuel.

The role of that generation is complementary to Las Majadas. Wind provides the renewable foundation, the grid provides another electrical path, and gas provides a dispatchable source that can run when additional firm power is required.

Together, those resources give the campus a much broader operating envelope.

Willacy County already sits inside a major gas infrastructure corridor

The geography adds another interesting dimension.

Willacy County is already crossed by significant natural-gas infrastructure, and that network is expanding.

One particularly important current project is the Rio Bravo Pipeline, which runs through Willacy County on its way from the Agua Dulce supply area toward the Rio Grande LNG complex near Brownsville. Rio Bravo says the system is designed to transport up to 4.5 billion cubic feet of natural gas per day, while the project has targeted service during the second half of 2026.

Another major South Texas system, Valley Crossing, has been operating since 2018 and carries up to 2.6 Bcf/d from the Agua Dulce area toward the Gulf Coast and Mexico.

The significance for Kati is the surrounding infrastructure environment. South Texas is not only a major renewable-power region; it also contains large natural-gas corridors capable of supporting industrial-scale energy demand.

Soluna has not publicly named the operator behind its Kati 2 pipeline-access agreement, but the regional infrastructure helps explain why onsite gas generation is a realistic component of the campus. Rio Bravo is particularly notable because its route includes Willacy County and its construction timeline overlaps directly with Kati 2’s 2026 development period.

The infrastructure around Kati is therefore becoming richer at the same time the data center itself is growing.

Battery storage provides another layer of control

The next component is battery energy storage.

Soluna’s Kati 2 campus plan includes a utility-scale BESS — Battery Energy Storage System — designed to support grid stability and provide supplemental backup power.

The role of the battery is broader than traditional emergency backup.

A large battery-inverter system can respond extremely quickly to changes in electrical load. That makes it particularly useful at AI facilities, where synchronized GPU workloads can create rapid changes in power consumption on timescales much faster than conventional generation can comfortably follow.

Soluna has described the broader AI-site architecture as using battery-inverter systems to shield the power system from GPU fluctuations while also providing ride-through capability during outages. Depending on the site design, those batteries can work alongside diesel backup, onsite gas generation, the renewable plant and the grid to create multiple levels of resilience.

That matters because AI creates two power problems simultaneously.

The first is energy quantity: finding enough megawatts to run the GPUs.

The second is power quality: managing how those GPUs consume electricity from moment to moment.

Kati 2’s architecture is increasingly being designed around both.

Microsoft had identified the same GPU power problem

This is where the Siemens work becomes particularly relevant to the broader Microsoft thesis.

In August 2025, Microsoft Research published “Power Stabilization for AI Training Datacenters.” The researchers described large training workloads involving tens of thousands of synchronized GPUs and showed how the movement between compute-heavy and communication-heavy phases can create large changes in electrical demand.

Microsoft treated the problem as a cross-stack engineering challenge. Software could modify workload behavior, GPU controls could alter power consumption, and data-center infrastructure — including energy storage — could help smooth the electrical profile presented to the utility.

That research fits naturally with the broader Microsoft progression already described in this paper. Virtual Battery made workloads respond to renewable generation. Heron and CWind routed inference around power availability at wind farms. Power Stabilization attacked the problem from the other side, making enormous GPU workloads behave more intelligently toward the electrical infrastructure supporting them.

Soluna was moving into the same technical territory commercially.

Project Grace brings Siemens into Soluna’s AI power stack

On January 8, 2026, Soluna and Siemens announced Project Grace, a 2 MW Texas pilot specifically designed to test rapid GPU-driven power changes in a behind-the-meter renewable environment.

The objective was highly specific: deploy and validate a power-and-controls architecture capable of capturing, monitoring and managing fast changes in AI and HPC electrical demand while those workloads operate directly on renewable energy.

Both Soluna and Siemens described the goal as creating a repeatable blueprint for future behind-the-meter AI deployments at renewable-generation sites.

Siemens was not participating only as an equipment vendor.

Project Grace was designed around Siemens electrical infrastructure, controls and monitoring. The planned system included transformers, switchgear, power converters and ancillary equipment, while Siemens’ current AI data-center materials identify its SICAM SCADA platform as part of the Soluna use case for monitoring and control.

The purpose is closely connected to the engineering challenge inside a campus like Kati 2.

A renewable-linked AI facility has variation on both sides of the electrical interface. The generation source can change with renewable conditions, while the GPU load itself can move rapidly as AI workloads transition between computational phases.

Controls sit between those two moving systems.

That is what makes Grace strategically important.

Soluna is not only assembling more sources of electricity. It is developing the control layer needed to make those resources and AI loads behave like one coordinated power system.

Siemens PTI takes the work into power-system modeling

The Grace program continued to become more technical during 2026.

In June, Soluna reported that PSS®E/PSCAD modeling was underway with the Siemens Power Technologies International — PTI — team.

Those tools are particularly relevant to the problem Grace is trying to solve. Siemens describes PSS®E as a platform for transmission planning, power-flow analysis, dynamic and transient-stability simulation, while its PSS®E–PSCAD capabilities allow detailed electromagnetic-transient models to interact with broader grid-stability simulations.

In other words, the Grace work was progressing beyond equipment selection toward modeling how the electrical system behaves dynamically.

That matters enormously for AI.

A conventional steady-state analysis can tell an engineer whether a system has enough megawatts. Dynamic and transient studies address a more sophisticated question:

what happens electrically when the load changes very quickly?

That is precisely the challenge created by synchronized GPU clusters.

Siemens itself increasingly presents this as an important data-center problem. Its AI-infrastructure platform now highlights advanced controls that coordinate onsite generation, UPS systems, storage and other power infrastructure to manage load profiles and improve readiness for AI workloads. On the same Siemens page, Project Grace appears as the case study for “Taming GPU power swings.”

That places Soluna’s work inside a much larger industry effort to redesign electrical infrastructure for AI.

Grace is intended as a blueprint, not an isolated experiment

The physical Grace pilot was initially associated with Soluna’s Texas Dorothy infrastructure. By August 2026, Carver said the 2 MW technical-validation work with the Siemens PTI team had been allocated to Dorothy 3 capacity as Soluna built out that larger AI campus.

From the beginning, Siemens and Soluna described the project as a repeatable blueprint for future behind-the-meter AI deployments, rather than a technology experiment intended to remain confined to one 2 MW location.

The technical lessons are therefore relevant across Soluna’s AI platform.

Kati 2 and Dorothy 3 differ physically, but both require Soluna to solve the same fundamental engineering problem: combine large AI loads with renewable generation, grid access, firming resources, storage and sophisticated electrical controls.

Grace provides a place to validate that control architecture before similar concepts are deployed at much larger scale.

That is particularly important when the potential end state is measured not in two megawatts but in hundreds.

The Microsoft–Siemens parallel is unusually close

There is an important parallel here.

Microsoft Research describes synchronized GPU power swings and studies how software, GPU controls and storage can stabilize them. Siemens and Soluna describe rapid GPU-driven power swings and are building a behind-the-meter pilot using electrical equipment, monitoring, SCADA controls and power-system modeling to manage them.

The research programs are separate, but they are solving the same emerging AI infrastructure problem from complementary positions.

Microsoft begins with the GPU fleet and asks how its electrical behavior can be stabilized.

Soluna and Siemens begin with the energy infrastructure and ask how it can safely accommodate those GPU workloads.

In both cases, the old model of treating electricity as a passive input to computing is disappearing.

Power behavior itself is becoming part of AI systems engineering.

That connection matters because Kati 2 is being developed for the exact class of large GPU deployment where these issues become significant. Its seven-building campus plan, high-density NVIDIA infrastructure and 100 MW first phase place it far beyond the scale at which power fluctuations can simply be ignored as an internal server-room issue.

The electrical system has to be designed for the workload.

Kati 2 increasingly resembles an energy campus with a data center inside it

Seen as a whole, the current Kati 2 architecture is much more sophisticated than the original behind-the-meter concept might suggest.

At the center sits Las Majadas, providing large-scale renewable generation immediately beside the development.

Around that sits ERCOT connectivity, allowing the campus to interact with the broader electrical system.

The Las Majadas substation is being engineered for another 100 MW to support future phases.

The campus plan includes 50–180 MW of onsite gas generation and diesel backup.

A gas-pipeline access agreement has been executed, with engineering underway for a lateral to the site.

A utility-scale BESS is part of the stabilization and backup strategy.

Renewable clustering provides a path toward incorporating additional regional generation as the campus expands.

And Soluna’s Siemens/Project Grace work is developing the controls and power-system knowledge needed to manage AI workloads inside that type of multi-resource environment.

The result is closer to an integrated energy campus than a conventional data center connected to a single utility feed.

That architecture is also what makes the 350 MW ambition more credible. The campus does not depend on finding one new 350 MW power source. Soluna is assembling a portfolio of resources around an existing renewable anchor and building the infrastructure that allows those resources to work together.

This is essentially the physical version of the principle that has appeared repeatedly throughout this research:

do not wait for one perfect source of power to reach the compute. Build the compute where power already exists, then combine generation, grid access, storage, firming and controls around it.

The overlap with Microsoft’s infrastructure thinking keeps growing

By this point, several parts of Microsoft’s own AI strategy have begun to map onto the Kati 2 architecture.

Microsoft Research has studied computing directly at renewable farms. CWind incorporates renewable generation, batteries, grid interaction and workload controls into distributed AI infrastructure. Microsoft’s Power Stabilization work treats GPU fluctuations as a major electrical-engineering problem and examines storage and cross-stack controls as part of the solution.

Microsoft’s real-world CO+I organization is securing enormous new quantities of generation and paying for the substations and transmission infrastructure needed to support AI campuses. Ryan Carver spent more than a decade inside that system before joining Soluna to lead development of Kati 2.

Now the project under his responsibility is being built around renewable generation, dual grid access, a 100 MW substation expansion, utility-scale BESS, onsite firm generation, pipeline gas and an AI power-controls program with Siemens.

These are not decorative additions to the campus. They are the systems that determine whether hundreds of megawatts of GPUs can operate at the reliability expected by a hyperscale customer.

And that takes the story directly into the next question.

Kati 2 is being engineered for a large AI tenant. Soluna has moved from broad interest, to due diligence, to formal negotiations, and finally to a signed LOI around the first major phase.

The project now has the scale, the land, the power architecture and a hyperscale development team.

The next question is who the campus is being built for.

10. The Tenant Process — From Interest to a 100 MW Anchor Customer

By the summer of 2026, Kati 2 had reached the point where the infrastructure story and the customer story were beginning to merge.

The project had grown into a 350+ MW AI campus, the first 100 MW of critical IT had been formalized through the Metrobloks joint venture, design was moving through development, a general contractor had joined the project, long-lead electrical equipment was being reserved, and the Las Majadas power infrastructure was expanding around it. At the same time, Soluna’s tenant process had moved from general inbound interest to due diligence, formal commercial negotiations and ultimately an exclusive Letter of Intent for the first major phase.

That progression matters because Kati 2 is not being developed around a hypothetical future customer profile anymore. A prospective tenant is already participating in the engineering process, reviewing the power architecture, aligning the building with its GPU roadmap and negotiating the commercial structure of a long-term lease.

The tenant’s name remains confidential. But the public record tells us considerably more about the customer than the absence of a name might suggest.

The customer universe was already visible in late 2025

The first public signs appeared before Metrobloks entered the project.

In late 2025, Soluna said Kati 2 had attracted significant interest from both neoclouds and hyperscalers. The initial test fit had already been completed for a data-center design capable of accommodating the latest NVIDIA GPUs, and architecture and electrical-engineering firms had been identified for the next stage of development.

That customer mix is important because Soluna has been unusually explicit about whom it wants to serve.

In a 2026 investor Q&A, CFO Michael Picchi described the target customer universe for large-scale AI infrastructure. On the hyperscaler side, Soluna specifically named Microsoft, Google, Meta and Amazon. On the neocloud side, it named companies including CoreWeave, Nebius, Crusoe, Lambda and TensorWave.

This was not simply a generic statement that “AI companies” might be interested in Kati. Soluna was designing the business around two clearly defined groups of large infrastructure buyers: hyperscalers capable of taking major long-duration power blocks and AI cloud operators looking for high-density GPU capacity.

Kati 2 was being built to accommodate either.

January 2026: the first LOI is described as neocloud

On January 13, 2026, Soluna disclosed that Kati 2 was already under a non-binding Letter of Intent from a potential neocloud tenant. Two days later, the Metrobloks announcement repeated that disclosure as the companies unveiled their plan for an initial 100+ MW critical-IT phase and a roadmap beyond 300 MW.

That early LOI is an important part of the chronology, but the tenant process did not stop there.

As the campus expanded during the spring, Soluna continued speaking with a wider customer field. In April, the company said hyperscaler and neocloud interest, due diligence and discussions remained active, while the company continued evaluating additional generation and engineering the larger campus.

By June, Soluna’s language changed again. Tenant due diligence was continuing with hyperscalers and neoclouds, but formal commercial negotiations had now started with at least one potential customer.

That is the point where the tenant process begins moving decisively toward a transaction.

The customer was performing real infrastructure diligence

The due diligence itself gives us an important window into the sophistication of the potential buyers.

Before the current LOI was signed, Soluna said interested parties were examining the site’s power, build-out schedule and fit with their infrastructure requirements. The targeted initial contract was approximately 100 MW of critical IT load.

Those are not superficial sales discussions.

A customer evaluating 100 MW of critical IT is effectively evaluating a billion-dollar infrastructure project. It needs to understand where the power comes from, how firm it is, how the site is interconnected, how quickly equipment can arrive, how the data halls will be configured, whether the cooling architecture supports its hardware and whether the campus can scale with its future requirements.

Soluna was preparing Kati 2 around precisely those questions. As the customers were conducting diligence, the company was simultaneously expanding Las Majadas power capacity, securing additional land, developing the gas-firming strategy, selecting the architecture and engineering team, advancing the design and starting procurement of long-lead equipment.

The tenant process and the engineering process were beginning to shape each other.

June: formal commercial negotiations begin

By June 9, 2026, Soluna said formal commercial negotiations had begun with at least one of the parties conducting diligence.

The timing is important. This was the same period in which Soluna and Metrobloks signed the definitive Kati 2 joint venture establishing 100 MW of Phase I critical IT and another 250 MW in Phase II. The project therefore entered formal tenant negotiations just as its long-term physical and legal structure was being locked into place.

The initial tenant block and the Phase I design line up almost exactly.

Soluna had repeatedly said it was targeting approximately 100 MW for the first AI contract. The definitive JV then established a first phase of 100 MW critical IT. The commercial process was therefore centered around an anchor customer capable of taking what is effectively the entire first hyperscale phase.

The rest of the campus would remain available for expansion.

That structure becomes particularly important once we examine what prospective customers were asking Soluna about during negotiations.

July: the process moves into exclusivity

On July 14, 2026, Soluna disclosed that tenant due diligence had progressed into formal commercial negotiations with one potential customer and that a Letter of Intent had been signed. The company said the work had shifted toward finalizing the design, commercial terms and lease agreement.

Three days later, John Belizaire provided considerably more detail in an interview with Power Analysis.

He said several companies and potential tenants had been evaluating Kati 2 during the first half of the year. After multiple rounds of conversations and diligence, Soluna signed an LOI containing exclusivity for formal commercial negotiations with one tenant. The objective was to emerge from that process with a 100 MW Phase I lease and a required-service date aligned with the customer’s schedule.

That is an important step beyond general interest.

Exclusivity means the relationship had progressed far enough for both sides to spend concentrated time working toward a definitive commercial structure. The next phase was no longer simply deciding whether Kati 2 was an attractive site. It was determining exactly how the campus would be delivered for that tenant.

And the customer was becoming deeply involved.

The tenant is now helping shape the facility

Belizaire said the prospective tenant had moved into deeper diligence involving its power and operations teams and was actively participating in the design process.

That design involvement is particularly revealing.

The tenant was reviewing whether Kati 2 would support the latest GPU generation and whether the design could also accommodate the next generation of hardware that would follow it. The customer was examining the battery system and its relationship to the data center for power smoothing and ride-through, reviewing regulatory matters and working through the emerging legal structure as the lease negotiations advanced.

By then, the project had reached approximately 50% design, meaning the customer was not commenting on a generic rendering. Its technical teams were engaging while the electrical and mechanical architecture was still at a stage where their requirements could influence the finished campus.

That is exactly what would be expected from an anchor tenant taking a large dedicated AI facility.

The project was increasingly being designed with the tenant rather than merely marketed to the tenant.

The hardware roadmap matters

The discussion around GPUs deserves particular attention.

Modern AI infrastructure moves quickly enough that designing only for the current generation of accelerators can make a facility obsolete before the lease has run even a fraction of its term. A customer taking a long-duration 100 MW position therefore has to think beyond the GPUs it plans to install on day one.

Belizaire said that was already happening at Kati 2.

The prospective tenant was examining the facility against the latest GPUs while also making sure that future generations could be supported by the design.

That has implications across the entire facility: rack density, liquid cooling, electrical distribution, transformers, backup architecture and the power-quality systems needed to support rapidly changing AI loads.

It also fits the broader Kati 2 development strategy. Soluna had already completed a test fit for current NVIDIA hardware, moved toward liquid-assisted high-density cooling and begun working with Siemens on the electrical behavior created by GPU workloads.

The tenant was now bringing its own hardware roadmap into that process.

The first Kati 2 lease was becoming an infrastructure-design exercise as much as a real-estate transaction.

The tenant is looking beyond the first 100 MW

Another detail from Belizaire’s July interview is particularly relevant to the scale of the potential customer.

He said interested parties were not only asking whether they could take the initial 100 MW. They were also asking what the next phase looked like and how quickly additional capacity could be delivered. Customer conversations involved aligning Soluna’s site ramp with the buyer’s own growth requirements.

That is exactly why the expansion from an 83 MW project into a 350 MW campus matters commercially.

A customer with a rapidly growing AI fleet does not necessarily want to complete a 100 MW deployment and then restart the entire site-selection process somewhere else. A campus that can deliver the first block and then add another 250 MW of critical IT gives the anchor tenant a much longer runway.

Soluna has gone even further in its marketing materials. It has said the Kati roadmap can ultimately be shown to customers as expandable toward 1,000+ MW, using the broader clustering strategy around regional power resources.

The current 350 MW plan therefore functions as the first major expansion envelope, not necessarily the theoretical end of the site.

For a hyperscaler planning AI capacity across multi-year hardware cycles, that scalability is particularly valuable.

Soluna is choosing for bankability as well as price

The tenant-selection process is also about more than who is willing to pay the highest lease rate.

Belizaire described the first customer as strategically important to Soluna’s broader AI platform. The company has been evaluating potential tenants based on their commercial terms, infrastructure fit and, crucially, bankability.

That becomes understandable once Kati 2’s financing model is considered.

Soluna estimates a 100 MW AI buildout could require approximately $1.2–$1.3 billion of capital. Its current plan is to finance roughly 70–80% of that amount using project-level debt, with the debt sized against the contracted cash flows under the tenant lease. Soluna has illustrated the model using a 100 MW, 15-year, long-duration triple-net lease.

The credit quality of the customer therefore directly affects the financing of the campus.

A strong tenant does more than generate rental income. Its contract becomes part of the financial asset against which the data center itself can be funded.

Belizaire made this distinction explicit when discussing the different customer categories. For hyperscalers, the credit support can come directly from the customer itself. Other counterparties may require different forms of financial backing or additional structures around the contract.

That makes the first customer decision especially consequential.

Soluna is not merely filling 100 MW. It is selecting the credit anchor for its first billion-dollar-scale AI campus.

The current LOI is no longer publicly classified as neocloud

This is an important point in reconstructing the timeline.

The January 2026 LOI was explicitly described as being with a potential neocloud tenant. But during the months that followed, Soluna continued conversations with both hyperscalers and neoclouds, added new parties to the process, conducted several rounds of diligence and ultimately entered an exclusive LOI with one prospective tenant in July.

On the August 13, 2026 earnings call, Compass Point analyst Michael Donovan asked management directly whether the tenant currently under LOI was a hyperscaler or a neocloud.

CFO Michael Picchi responded that Soluna had not specified which of those two categories the tenant belongs to. John Belizaire then added that the broader customer interest the company was seeing spans both hyperscale and neocloud buyers.

That answer is significant for this investigation.

The current exclusive tenant relationship should therefore be evaluated from the full hyperscaler-or-neocloud universe, rather than automatically carrying forward the neocloud description attached to the earlier January disclosure.

By August, Soluna was deliberately preserving the category of the tenant along with its identity.

Demand around Kati continued even after exclusivity

The exclusive tenant process did not eliminate broader demand for the site.

In the July interview, Belizaire described additional inbound interest from companies asking whether Kati 2 remained available and whether future power at the campus could still be accessed. He said Soluna was already having to explain that it was in discussions with another party on the current phase, while potential customers were asking about the remaining power and future rights to capacity.

That is a valuable signal because it shows the commercial logic behind building Kati 2 larger than the first 100 MW contract.

The first tenant does not consume the entire development opportunity. Phase II creates another 250 MW, while additional clustering could potentially expand the campus further.

Soluna has said that only one tenant is likely to occupy the initial Kati 2 opportunity, allowing other interested customers to become candidates for Dorothy 3 and subsequent AI campuses.

The company is therefore beginning to build something broader than a one-customer project.

Kati 2 is becoming the first customer-conversion point for an AI development pipeline that had grown to more than 1.6 GW by August 2026.

Ryan Carver arrives after the LOI — and directly into delivery

The timing of Ryan Carver’s appointment adds another layer to the process.

Soluna publicly disclosed the signed tenant LOI on July 14.

Two days later, on July 16, it announced that Carver — Microsoft’s former Senior Director of AI Construction & Site Development and a leader in the Fairwater program — had joined Soluna as Chief Development Officer.

The tenant process was already well advanced when Carver formally joined Soluna. His importance lies in what happened next: Soluna had reached exclusivity with a customer, and then immediately placed an executive with direct Microsoft hyperscale-AI development experience over the organization responsible for delivering the campus.

Less than a month later, Carver was the executive publicly explaining the Kati 2 build.

He described the design-development work, the contractor, long-lead procurement, the additional Las Majadas substation capacity, pipeline-gas access, additional land and the active lease negotiations. He said the work being performed during the negotiation was precisely the work required to move once a signed lease is in place.

The combination is powerful.

The tenant was already engaged. The campus was already being customized. Then Soluna added Microsoft hyperscale execution experience directly into the delivery organization.

By August, the negotiations had become the center of Kati 2

On the August 13 earnings call, Carver described the current tenant process in straightforward terms: an LOI had been signed, commercial terms and lease negotiations were underway, and Soluna would update the market when it had a definitive agreement to announce.

The important detail is what was happening while those negotiations continued.

Design development had advanced.

The general contractor was onboard.

Long-lead orders were being placed.

Substation engineering was progressing.

The tenant’s technical teams were engaging with the facility design.

Power and operating requirements were being reviewed.

The lawyers were working toward the lease.

That is the profile of a project moving toward execution.

The commercial agreement is becoming the final organizing layer around infrastructure that is already being prepared for the customer.

And this is where Microsoft fits increasingly well

The Microsoft thesis becomes more interesting when the characteristics of the Kati 2 tenant process are compared with the Microsoft infrastructure story developed in the previous chapters.

Soluna itself has publicly named Microsoft as one of the hyperscalers it wants to secure for large-scale AI infrastructure. Kati 2 offers the type of initial capacity block Soluna says it is marketing to those customers: approximately 100 MW of critical IT, expandable to hundreds of megawatts at the same campus.

The prospective tenant is bringing power and operations teams into due diligence. It is actively shaping the design around current and future GPU generations. It is studying battery behavior and power smoothing. It is examining how rapidly subsequent phases can be delivered. And its credit profile is important enough to become the foundation for billion-dollar-scale project financing.

Those characteristics align naturally with a sophisticated hyperscale infrastructure buyer.

Microsoft, meanwhile, is building AI capacity at a scale measured in gigawatts. Fairwater alone is expected to scale toward 2 GW. Microsoft is contracting enormous amounts of generation, funding electrical infrastructure, studying power-aware AI at renewable farms and working directly with Masdar on energy projects for its global AI expansion.

Masdar co-owns Las Majadas.

EDF has supplied Microsoft renewable power before.

Microsoft Research repeatedly cites Soluna.

Ryan Carver has moved from Microsoft’s Fairwater program into leadership of Soluna’s AI-development platform.

And Kati 2’s current anchor tenant is operating inside the same broad category Soluna has deliberately left open:

hyperscaler or neocloud.

The tenant process therefore gives the Microsoft thesis a much more concrete commercial frame.

Microsoft Research has already shown sustained interest in the architecture. Microsoft and Masdar are already collaborating on energy for AI infrastructure. Microsoft has historical ties to EDF. A former Microsoft hyperscale-development executive now leads Soluna’s AI buildout. And an anonymous, sophisticated buyer is in exclusive negotiations for the first 100 MW of Kati 2 while helping shape the campus around its power requirements, GPU roadmap and expansion needs.

The tenant has moved from interest to diligence, from diligence to commercial negotiations and from commercial negotiations to exclusivity.

The next layer of the investigation is therefore the physical and legal footprint underneath the transaction itself — the land companies, title insurance, hidden schedules, purchase agreements and external development parties that show exactly how much of Kati 2 has already been locked into place.

11. The Land Beneath Kati 2 — The 350 MW Campus Becomes Physical

By the summer of 2026, Kati 2 was no longer defined only by megawatts, design drawings and tenant discussions. The physical footprint underneath the project was being assembled around the same 100 MW Phase I + 250 MW Phase II structure that now defined the 350 MW campus.

Soluna had already reported securing more than 500 additional acres for Kati 2 as the project expanded beyond its original 83 MW concept. That land provides the space required for a multi-building AI campus, including data halls, electrical infrastructure, cooling systems, roads, battery storage, onsite generation and future phases.

The definitive Metrobloks joint venture signed on June 3, 2026 makes the property structure even more concrete. The Kati 2 JV owns separate entities for the two major stages of the development: Soluna MB KK II LandCo Phase I, LLC and Soluna MB KK II LandCo Phase II, LLC. The legal structure therefore mirrors the physical development plan: approximately 100 MW of critical IT in Phase I and another 250 MW in Phase II.

Phase I land is already inside the project

The JV agreement states that Phase I LandCo was already the owner of the Phase I property when the definitive agreement became effective.

Soluna subsequently contributed that property into the Kati 2 JV together with development costs and additional capital. Its funded contribution at formation totaled approximately $6.7 million, including roughly $1.1 million of Phase I land, approximately $3.6 million of pre-formation development costs and another $2 million in cash.

That matters because the first 100 MW phase is not being planned around land that still has to be identified. The property underlying Phase I is already part of the development structure being used to negotiate with the anchor tenant.

The Phase I property also carries an Owner’s Policy of Title Insurance issued by Sierra Title Insurance Guaranty Company, adding another concrete property-level layer beneath the first phase.

Phase II was already under contract

The expansion land is similarly advanced.

The JV filing reveals that Soluna entered into a Purchase and Sale Agreement dated March 10, 2026 for the Phase II property. When the definitive Metrobloks JV was created, those acquisition rights were transferred into Phase II LandCo.

Soluna committed approximately $19.1 million toward completing that acquisition.

This is particularly important because Phase II represents the additional 250 MW of critical IT behind the long-term Kati 2 campus plan. The expansion was therefore already supported by a contractual property-acquisition process while Soluna was negotiating the first 100 MW tenant lease.

The land strategy was being built ahead of the customer ramp rather than waiting for Phase I to fill before planning the next stage.

Crawford Lands provides an independent development link

The public filings are reinforced by another direct source from the development itself.

Luke Crawford, founder of Crawford Lands, has publicly said that his firm supported Soluna’s land-acquisition efforts for Project Kati 2 in Willacy County. Crawford Lands also lists Soluna among its clients, and Crawford has posted publicly about working alongside the company on the South Texas development.

John Belizaire responded to one of those posts by describing Crawford as “a true partner and our secret weapon in data center development behind-the-meter.”

That gives the land program an independent public development node outside Soluna’s SEC filings and investor materials.

By May 2026, Soluna had also completed geotechnical work on the expansion parcel, showing that the additional acreage was already moving into the engineering work required for future construction.

The physical footprint matches the commercial plan

The importance of the land evidence is its consistency with everything else happening at Kati 2.

The engineering plan calls for a 350 MW critical-IT campus.

The definitive JV establishes 100 MW in Phase I and 250 MW in Phase II.

Separate LandCos exist for those same two phases.

The first-phase property is already owned.

The second-phase property is tied to a signed purchase agreement and approximately $19.1 million of committed acquisition capital.

More than 500 acres are being assembled around the larger campus.

And the tenant process is centered on approximately the same 100 MW first phase.

These are different parts of the project arriving at the same structure.

Kati 2’s 350 MW expansion is no longer simply a roadmap. The land underneath both phases is already being organized around it.

That makes the next stage of the story much more significant. Soluna now has the power position, the expanding electrical infrastructure, the land, the development partner and an anchor tenant in exclusive negotiations.

The remaining question is no longer whether Kati 2 can become a hyperscale campus.

It is which customer is preparing to occupy the first 100 MW — and how the full network of Microsoft, EDF, Masdar, Soluna and Kati 2 fits together around that decision.

12. The Convergence — Why Microsoft Fits Kati 2

The Microsoft thesis around Project Kati 2 becomes most compelling when the evidence is no longer viewed as a collection of isolated coincidences, but as several independent paths arriving at the same physical project. One path begins inside Microsoft Research, where engineers have spent years studying how AI compute can move toward renewable generation rather than waiting for conventional grid infrastructure to reach the data center. Another runs through EDF and Masdar, the two renewable-energy companies behind Las Majadas, both of which have direct institutional relationships with Microsoft. A third runs through Microsoft’s physical AI infrastructure organization and Ryan Carver, who moved directly from building Microsoft’s Fairwater AI campus to leading Soluna’s AI/HPC development platform. The final path is commercial and physical: Kati 2 has simultaneously grown into a 350+ MW critical-IT campus while an anonymous anchor customer has advanced through technical diligence and into a signed LOI for the first approximately 100 MW phase.

The argument therefore does not depend on one document revealing a hidden customer name, because the stronger signal comes from the accumulation of relationships, technology choices, personnel movements and project-development decisions that all point toward the same type of buyer. Microsoft increasingly fits Kati 2 not because of one clue, but because the company appears repeatedly on both sides of the infrastructure Soluna is assembling.

Microsoft was already trying to solve the problem Kati 2 was built around

Long before Kati 2 became a 350 MW AI campus, Microsoft researchers had begun challenging one of the traditional assumptions of data-center development: that computing infrastructure should remain fixed while electricity is transmitted over increasingly constrained grids to reach it.

That idea evolved through Microsoft’s renewable-computing research and became much more concrete in 2025. On May 15, 2025, Microsoft researchers published AI Greenferencing: Routing AI Inferencing to Green Modular Data Centers with Heron, proposing modular AI clusters physically co-located at wind farms and using software to route inference workloads between sites as renewable generation changed. The study used real wind data and Azure production traces, estimated that more than 640 GW of large wind capacity globally and more than 150 GW in the United States sat within 50 milliseconds of Azure data centers, and found that its Heron router could improve aggregate AI-compute goodput by as much as approximately 80% in the evaluated conditions.

The most relevant part of Heron for this investigation is that Microsoft did not discuss wind-farm computing only in theoretical terms. The paper specifically identified windCORES, Soluna and WestfalenWIND as companies already deploying compute at wind farms, and described high-value AI inference as a major opportunity for this emerging class of co-located-compute operator. Microsoft explicitly argued that such companies could benefit by moving beyond existing workloads into AI, while AI providers could gain additional compute capacity and wind farms could monetize electricity locally.

That places Soluna inside Microsoft’s own technical framing of the future AI-power architecture.

Microsoft researchers were effectively describing the business model Soluna had been developing for years: locate compute directly beside renewable generation, consume electricity behind the meter, use software and infrastructure controls to manage variability, and reserve conventional hyperscale infrastructure as a complementary part of the system.

Soluna appears again when Microsoft’s research becomes more ambitious

The connection did not end with Heron.

Microsoft’s follow-on work appeared in May 2026 and was revised in July under the title CWind: A Cross-site Router for Large Language Model Inference Serving at Renewable Energy Farms. CWind significantly enlarged the opportunity identified by Heron, estimating that more than 890 GW of operating or under-construction wind capacity of at least 100 MW sits within 50 milliseconds of Azure data centers, with 73% within 20 milliseconds. Under the paper’s right-sizing assumptions, Microsoft estimated that more than 10 million H100-equivalent GPUs could potentially be deployed at wind farms in this architecture.

The experimental system was no longer purely analytical. Microsoft ran CWind on a 64-GPU A100 testbed emulating three wind-powered sites, used Azure production traces for the AI workload and real U.S. wind-farm data for the power profiles, and reported reductions in P99 end-to-end latency of 22–52% relative to its strongest alternative and up to 98% against simpler baselines. The three wind profiles were selected from the Central United States specifically because their generation patterns complemented one another geographically.

Soluna appears again in the paper’s discussion of renewable-powered modular data centers. CWind lists windCORES, Soluna and WestfalenWIND as real companies deploying compute at wind farms before explaining how Microsoft’s Greeninferencing architecture extends the concept toward high-value AI workloads.

The recurrence matters. Soluna did not merely appear once in an obscure bibliography and disappear from Microsoft’s thinking; it remained one of the small number of real-world companies Microsoft researchers chose to identify while advancing essentially the same architecture from Heron into a hardware-tested Azure-linked system.

CWind introduces one of the most interesting unnamed relationships

CWind also contains a sentence that becomes unusually important when placed beside the rest of the Kati record.

While summarizing the commercial potential of its architecture, the Microsoft researchers state that they are “working closely with a large renewable energy company” that sees significant value in the strategy. The paper does not identify that company, but this means Microsoft’s work has already moved beyond an internal research concept into direct engagement with a major renewable-energy organization interested in bringing AI compute to renewable-generation sites.

That immediately raises the question of which renewable company would naturally fit the relationship.

EDF is one of the strongest candidates because it has already done exactly what CWind describes. Through Project Tumbleweed and its broader Asset Optimization work, EDF deliberately searched for flexible behind-the-meter computing customers willing to co-locate at renewable substations. That process helped bring Soluna to Las Majadas, where EDF and Masdar later announced a 166 MW behind-the-meter agreement explicitly intended to power advanced computing, including AI, directly beside the wind farm’s substation.

Masdar strengthens the possibility further because it is not simply a passive co-investor in Las Majadas. It has become a direct Microsoft energy partner whose public mandate now includes helping develop energy infrastructure for Microsoft’s global AI expansion.

The three-site detail creates an additional fingerprint

There is another detail inside CWind worth placing alongside the EDF–Masdar relationship.

Microsoft’s hardware experiment uses power profiles derived from three wind-farm sites in the Central United States, selected because their generation complements one another geographically. The paper does not publish the identities of those sites, but Microsoft’s decision to model a three-site wind architecture becomes interesting when compared with Masdar’s original EDF U.S. wind investment.

When Masdar entered the EDF North American portfolio in 2020, it acquired interests in exactly three utility-scale wind projects: Coyote, 243 MW in Scurry County, Texas; Las Majadas, 273 MW in Willacy County, Texas; and Milligan 1, 300 MW in Saline County, Nebraska. Together, those three projects represent 815 MW of wind capacity across Texas and Nebraska, and Las Majadas is the asset that subsequently became the physical home of Soluna’s Kati development.

The three-site count and Central U.S. geography are best viewed as a fingerprint rather than an identification, because the CWind paper does not name the individual wind farms behind its traces. What makes the overlap noteworthy is the broader context around it: Microsoft researchers are working closely with an unnamed large renewable company, their experiment uses three complementary Central U.S. wind sites, Soluna is repeatedly cited as a real compute-at-wind operator, and the EDF–Masdar portfolio contains a three-wind-farm U.S. platform whose Las Majadas asset already hosts Soluna.

That combination gives the unnamed-renewable-company line considerably more weight than it would carry in isolation.

EDF and Microsoft already have a direct history

EDF also requires no hypothetical bridge to Microsoft because the two companies have worked together directly for more than a decade.

On July 15, 2014, EDF announced its investment in the 175 MW Pilot Hill Wind Project in Illinois, backed by a 20-year PPA with Microsoft. The project was located on the same electrical grid serving Microsoft’s Chicago-area data center, and Microsoft described the transaction as part of its effort to transform the energy supply chain for cloud computing.

The language Microsoft used at the time is striking in retrospect. Its energy leadership described the company’s objective as transforming the cloud-energy supply chain “from the power plant to the chip.” Pilot Hill therefore established a direct Microsoft–EDF commercial relationship around the same fundamental idea that would later become far more important in the AI era: Microsoft using long-term purchasing power to secure renewable generation for data-center infrastructure.

By 2015, Pilot Hill was operational, with EDF stating that Microsoft’s long-term commitment would support 100% of the energy requirements of its Illinois data center.

The relevance to Kati 2 is not that Pilot Hill and Las Majadas are identical transactions, because they represent different stages in the evolution of data-center energy strategy. The relevance is that EDF has already been inside Microsoft’s data-center power supply chain, and a decade later EDF became the renewable developer that deliberately searched for a behind-the-meter computing partner and selected Soluna’s model at Las Majadas.

Masdar creates an even more current Microsoft bridge

If EDF provides the historical Microsoft connection, Masdar provides the contemporary one.

On November 5, 2024, Masdar, ADNOC and Microsoft announced a strategic collaboration under which the parties would evaluate opportunities to power Microsoft’s data centers with renewable energy through Masdar. Masdar CEO Mohamed Jameel Al Ramahi said the company looked forward to working with Microsoft and other partners to deliver clean energy to the data centers that would power the AI future.

That language is unusually direct for the Kati thesis because Masdar was already a 50% investor in Las Majadas.

One corporate entity was therefore sitting on both sides of the emerging story. In Texas, Masdar co-owned a 273 MW wind project where Soluna was developing behind-the-meter computing. At the global level, Masdar was simultaneously entering an explicit collaboration to evaluate renewable power for Microsoft data centers.

The relationship became substantially deeper one year later. On November 2, 2025, Masdar, XRG, ADNOC and Microsoft announced an expanded agreement under which Masdar and XRG would develop energy projects and infrastructure in support of Microsoft’s global AI and data-center expansion. This was no longer framed simply as evaluating future opportunities; the agreement explicitly linked Masdar’s energy-development capabilities to the infrastructure required for Microsoft’s growing AI platform.

For Kati 2, that creates a direct institutional chain:

Microsoft needs AI power → Microsoft partners with Masdar to develop energy infrastructure → Masdar co-owns Las Majadas → Las Majadas powers Soluna → Soluna develops Kati 2 for a hyperscale AI customer.

Mohamed Jameel Al Ramahi spans the entire Masdar side of the story

The continuity becomes even more interesting at the executive level.

Mohamed Jameel Al Ramahi has served as Masdar’s CEO throughout each relevant phase. He was CEO when Masdar agreed in 2020 to acquire its stake in EDF’s U.S. renewable portfolio containing Las Majadas, Coyote and Milligan. He remained CEO when Masdar entered its 2024 Microsoft collaboration around renewable energy for data centers, and he remained at the helm as that partnership expanded toward energy infrastructure for Microsoft’s global AI growth.

More importantly, Al Ramahi personally highlighted the Las Majadas–Soluna transaction when it became public in 2025. He described rising demand from advanced computing and AI, said renewable energy had become increasingly important for powering data centers, and highlighted the agreement through which Masdar and EDF would supply clean power from Las Majadas to Soluna’s co-located Texas data center.

That means the same Masdar executive was publicly engaged with both sides of the emerging infrastructure model: Microsoft’s need for clean energy to power AI data centers and Soluna’s use of Masdar-owned wind generation for co-located computing at Las Majadas.

The significance is organizational continuity rather than mere name repetition. Masdar did not make one isolated investment in Texas and later develop an unrelated relationship with Microsoft under entirely different leadership; both relationships matured while the same CEO was leading the company and publicly describing renewable power and data-center infrastructure as increasingly connected markets.

The chronology compresses rapidly after 2024

When the dates are placed in sequence, the convergence becomes much easier to see.

In May 2024, Soluna signed the definitive 166 MW Kati power agreement with EDF and Masdar. In November 2024, Microsoft and Masdar announced that they would evaluate powering Microsoft data centers with renewable energy through Masdar. In March 2025, EDF and Masdar publicly identified Las Majadas as the wind asset supplying Soluna’s co-located Kati data center and explicitly positioned advanced computing and AI as part of the opportunity.

Roughly ten weeks after the public Las Majadas announcement, Microsoft researchers published Heron and explicitly cited Soluna as an existing compute-at-wind company capable of benefiting from high-value AI workloads. Later in 2025, Microsoft and Masdar expanded their institutional relationship toward energy projects and infrastructure supporting Microsoft’s global AI growth.

During the first half of 2026, Kati 2 then changed scale dramatically. The project moved from the second 83 MW phase of the original Kati agreement toward a 350+ MW critical-IT campus, with the June joint venture legally defining a 100 MW first phase and another 250 MW second phase. At the same time, Microsoft published CWind, again citing Soluna, again advocating AI compute directly at wind farms, and revealing that its researchers were already working closely with a large renewable-energy company that saw significant value in the model.

Then came July.

Ryan Carver turns the Microsoft connection from institutional into personal

On July 16, 2026, Soluna announced that Ryan Carver would become its Chief Development Officer.

Carver had spent more than ten years at Microsoft and most recently served as Senior Director of AI Construction & Site Development, with responsibility across Microsoft’s AI data-center campus development. Soluna said he had led a construction P&L measured in the tens of billions of dollars and had been involved in Microsoft’s Fairwater campus in Mount Pleasant, Wisconsin. His new Soluna mandate covers the full AI/HPC development lifecycle, from site origination and power procurement through design, construction, commissioning and operations.

Microsoft describes Fairwater as its largest and most sophisticated AI factory, occupying approximately 315 acres, with three major buildings totaling 1.2 million square feet and roughly 120 miles of medium-voltage underground cable. The facility was designed to connect hundreds of thousands of NVIDIA GPUs into a massive AI system, and by June 2026 the first facility had become fully operational while construction on the adjacent second facility continued.

Carver therefore arrived at Soluna immediately after working on one of the clearest examples of how Microsoft builds physical AI infrastructure at hyperscale. During Soluna’s August earnings presentation, he described his Microsoft role as taking AI campuses from land through power procurement, permitting, design, construction, commissioning and handover to operations, before explaining why Soluna’s behind-the-meter model had attracted him: he had repeatedly seen power availability become the limiting factor while compute hardware and buildings could move more quickly.

This is perhaps the most direct bridge in the entire thesis because it connects Microsoft hyperscale development expertise to the exact Soluna organization responsible for Kati 2.

Carver arrives just as Kati 2 reaches the tenant stage

The timing of Carver’s move matters because Kati 2 was already entering its most commercially important stage when he joined.

By August, Kati 2 was expected to exceed 350 MW, with Phase I at 100+ MW of critical IT and Phase II adding another 250 MW. Design development was nearing completion, the general contractor had joined the project, commitments had been made with key long-lead electrical suppliers, engineering was underway for another 100 MW at the Las Majadas substation, additional land was being assembled and Soluna had executed a gas-pipeline access agreement to support onsite firming generation.

Most importantly, Soluna had signed a tenant LOI, with commercial terms and lease negotiations underway while final design work continued. Carver’s own description of the development is revealing because he emphasized that the contractor was already onboard, long-lead orders were already being placed and substation engineering was running in parallel, describing those activities as the work required to move once the lease is signed.

That places a former Microsoft AI campus executive inside Soluna at precisely the point when Kati 2 is transitioning from development into customer-specific execution.

The significance becomes greater when we remember that the customer was already conducting the kind of diligence expected of a sophisticated AI buyer: reviewing power, construction schedule and fit with its infrastructure requirements around an initial 100 MW critical-IT contract.

Soluna itself places Microsoft inside the target customer set

There is also no need to infer whether Microsoft fits Soluna’s intended customer profile, because Soluna has named it directly.

When CFO Michael Picchi was asked what type of AI/HPC customers the company wanted to secure, he identified the hyperscaler universe as Google, Meta, Amazon or Microsoft, alongside large neocloud providers such as CoreWeave, Nebius, Crusoe, Lambda and TensorWave. In the same Q&A, Soluna said it was marketing Kati 2 around an initial 100 MW critical-IT Phase I, with expansion above 300 MW and a customer roadmap showing potential scale beyond 1,000 MW through future clustering.

That makes the Microsoft thesis structurally straightforward.

Microsoft is one of the exact hyperscalers Soluna says it wants for the project. Kati 2 is being sized around the kind of large initial block such a hyperscaler can consume. The campus provides hundreds of megawatts of follow-on capacity, and the surrounding power strategy gives Soluna a path to show even greater long-term expansion.

The fit becomes stronger because Microsoft itself is operating at a scale where additional 100 MW blocks are meaningful but no longer extraordinary. By February 2026, Microsoft said it had contracted 40 GW of new renewable-energy supply across 26 countries through more than 400 contracts, with 19 GW already online.

Kati 2 therefore sits comfortably inside the scale and energy-procurement behavior of the company being hypothesized.

The Kati tenant profile resembles a hyperscale infrastructure buyer

The prospective Kati 2 customer’s behavior is also consistent with the way a sophisticated hyperscaler would evaluate a purpose-built AI campus.

The customer process has focused on power, delivery schedule and infrastructure fit, rather than simply price and floor space. Soluna’s development team has been advancing design and power infrastructure while negotiations continue because the customer is effectively evaluating whether the project can deliver a major AI deployment on its required schedule.

The economics reinforce that profile. Soluna estimates that a 100 MW AI phase could require approximately $1.2–$1.3 billion of construction capital and expects the majority of that capital to be raised as project-level debt underwritten against contracted tenant cash flows. The first tenant therefore becomes more than a user of the facility; its credit and long-term lease become part of the financial infrastructure supporting the build.

For Microsoft, that structure would be familiar territory. Its own renewable-energy strategy has relied heavily on long-duration contractual commitments capable of enabling large infrastructure investments, from Pilot Hill in 2014 to the hundreds of energy contracts that now form its 40 GW renewable portfolio.

The Kati 2 customer is therefore being selected for exactly the combination of characteristics a large institutional technology company can provide: substantial capacity demand, technical sophistication, future expansion needs and a credit profile capable of supporting project financing.

The architecture itself increasingly looks Microsoft-compatible

Kati 2’s technical architecture also aligns remarkably well with the power problems Microsoft has been studying.

The site starts with a large wind asset immediately beside the campus. It adds ERCOT connectivity, an expanding substation, onsite firm generation and planned storage and controls, creating a hybrid system capable of serving dense AI workloads without abandoning the economic advantage of renewable power at the source. Ryan Carver described the model as combining the renewable plant, the grid and firming resources into an integrated system capable of reaching the availability contracted by an AI customer.

Microsoft Research has independently moved in almost exactly the same direction. Heron and CWind place modular AI compute at wind farms, use distributed routing to manage renewable variability and preserve conventional data centers as complementary capacity. Microsoft’s separate AI power-stabilization work has focused on controlling the rapid electrical swings created by large GPU clusters, while Soluna and Siemens are developing the physical power-and-controls side of that same problem through Project Grace.

The overlap therefore reaches beyond the simple statement that both companies care about renewable energy.

Microsoft is researching how AI should behave when compute is moved toward variable renewable generation, while Soluna is engineering the physical infrastructure required to make that model commercially deployable at scale.

Two paths have been approaching the same point

The easiest way to understand the convergence is to trace the story from both directions.

From the energy side, Las Majadas was developed as a major Texas wind asset. Masdar acquired a 50% interest alongside EDF. EDF later searched for behind-the-meter flexible computing demand, selected Soluna through the commercial process that became Kati, and the first 83 MW developed into Kati 1. Kati 2 subsequently expanded into a 350+ MW AI campus while Masdar simultaneously deepened its relationship with Microsoft around renewable power and infrastructure for global AI growth.

From the compute side, Microsoft spent years researching how workloads could respond to renewable generation. That work moved from Virtual Battery and modular renewable data centers into Heron and CWind, where wind-farm AI became the explicit deployment model and Soluna was cited by name. Microsoft researchers then revealed they were working with a large renewable-energy company on the strategy, while a Microsoft executive with direct Fairwater experience left the company and joined Soluna to lead the development organization responsible for Kati 2.

Those two paths now meet at the same physical location in Willacy County.

That is the central observation of this research.

The pattern is stronger because the connections are independent

What makes the Microsoft case particularly interesting is that the individual connections come from different organizations and different types of evidence.

The Microsoft Research connection comes from academic papers. The EDF connection comes from a historical Microsoft PPA and EDF’s later behind-the-meter strategy. The Masdar connection comes from direct strategic agreements with Microsoft and Masdar’s ownership of Las Majadas. The Al Ramahi connection comes from executive continuity and his own public promotion of the Las Majadas–Soluna data-center model.

The Carver connection comes from a direct personnel move from Microsoft’s AI construction organization into Soluna. The Kati 2 connection comes from SEC filings, engineering milestones, land acquisition, substation expansion and tenant negotiations. The commercial fit comes from Soluna itself naming Microsoft among its intended hyperscaler customers and structuring Phase I around an initial 100 MW critical-IT lease.

These signals were not created by a single press release or repeated from one underlying source. They emerged independently across Microsoft, Microsoft Research, EDF, Masdar, Soluna, regulatory filings and individual career histories.

That independence is what gives the accumulation its weight.

The timing may be the strongest part of the pattern

The sequence becomes especially striking from 2024 onward.

Soluna and EDF/Masdar formalize Kati. Microsoft and Masdar formalize collaboration around renewable power for Microsoft data centers. EDF and Masdar publicly reveal that Las Majadas will supply Soluna’s co-located computing facility. Microsoft Research publishes Heron and names Soluna.

Microsoft and Masdar then deepen their relationship toward global AI energy infrastructure. Kati 2 expands into a hyperscale-class 350+ MW campus. Microsoft Research publishes CWind, again names Soluna and discloses work with a large renewable company. Soluna enters advanced negotiations for the first approximately 100 MW AI tenant.

Ryan Carver then moves from Microsoft’s AI data-center organization directly into Soluna and takes responsibility for delivering the project. By August 2026, he is publicly explaining the Las Majadas substation expansion, the gas connection, long-lead procurement, design development and tenant lease process around the same campus.

When viewed chronologically, the connections are no longer spread randomly across a decade. They increasingly cluster around the exact period in which Kati 2 is turning from a renewable-computing concept into a hyperscale AI project.

Why Microsoft now stands out

There are many companies capable of consuming 100 MW of AI capacity, and Soluna has publicly identified both hyperscalers and neoclouds as potential customers. What makes Microsoft unusually compelling within that universe is the number of additional links that exist specifically around the underlying infrastructure.

Microsoft has a direct historical relationship with EDF. Microsoft has an active strategic energy relationship with Masdar. Masdar co-owns Las Majadas, and Las Majadas powers Kati. Microsoft Research repeatedly cites Soluna as a real-world compute-at-wind operator and is actively working with an unnamed large renewable-energy company on the same architecture.

Microsoft’s current research uses real wind power and Azure workloads to test multi-site Greeninferencing systems. A Microsoft hyperscale AI development executive has moved directly into Soluna. And Kati 2 has a 100 MW anchor tenant under LOI at exactly the moment the campus is being engineered toward the requirements of large AI buyers.

No other candidate currently identified in the public Kati customer universe has surfaced across as many independent layers of the project.

That is why Microsoft has become more than simply one name on a list of plausible tenants.

It is the company that most consistently appears wherever the Kati 2 story touches research, renewable-energy partners, hyperscale development expertise, power constraints and AI infrastructure scale.

The convergence

The story began years before Project Kati existed, with Soluna developing the idea that computing could become a flexible industrial load attached directly to renewable generation. EDF later encountered the same opportunity from the power-owner side and went searching for exactly that kind of customer, while Masdar entered Las Majadas as EDF’s co-owner and later became a strategic Microsoft energy partner.

Microsoft Research independently developed the technical argument for moving AI compute directly to renewable farms and then began citing Soluna as one of the companies already operating there. Microsoft’s physical AI organization simultaneously ran into the real-world power constraints that its researchers were trying to solve architecturally, and Ryan Carver experienced that problem while helping develop Fairwater before moving directly to Soluna because he believed behind-the-meter renewable computing offered a compelling answer.

The project now under his development organization is Kati 2: more than 350 MW of planned critical IT beside the EDF–Masdar Las Majadas wind farm, with a 100 MW first phase, expandable land and power infrastructure, advanced design, long-lead procurement and an anonymous anchor tenant already under LOI.

The paths have been separate for years, but by 2026 they are arriving at the same place. Microsoft has the demand, Masdar and EDF sit inside the energy chain, Microsoft Research has articulated the architecture and repeatedly identified Soluna as a real-world example, Ryan Carver brings Microsoft’s hyperscale execution experience directly into Soluna, and Kati 2 has become large enough, sophisticated enough and expandable enough to serve exactly that class of customer.

For the Microsoft thesis, that is the real significance of the evidence accumulated throughout this investigation. The case is no longer built around finding one mysterious clue that happens to point toward Redmond; it is built around the fact that independent technical, institutional, commercial and personnel trails keep converging around the same renewable-powered AI campus in South Texas.

At some point, accumulation itself becomes information, and at Kati 2, Microsoft is now the name around which the largest number of those independent signals align.