Oracle and OpenAI Walked. Microsoft Pounced.

Microsoft locked up 900 megawatts at Crusoe's Abilene campus — the same site Oracle and OpenAI walked away from in early March. The deal is not a surprise. It is a land-grab in slow motion, and the real story is the power infrastructure underneath it.

Crusoe, the AI factory company that got its start burning oil-field flare gas to mine bitcoin, announced March 27 that it is developing a new dedicated campus for Microsoft at its Abilene, Texas site: 900 megawatts of new capacity, two buildings each rated for 336 megawatts of critical IT load, and an onsite natural gas plant paired with a medium-voltage battery storage system for behind-the-meter power according to Crusoe's announcement. The first building is expected online mid-2027. Land clearing is already underway. When complete, the full Abilene site reaches approximately 2.1 gigawatts of projected capacity across all tenants.

That 2.1 gigawatt figure is doing a lot of work in the press release. The operational reality is narrower. Only about 200 megawatts of the existing 1.2 gigawatt Oracle and OpenAI project at that site has been powered on, The Register reported. The remaining capacity is expected to come online through 2026 — not 2025, not on the original schedule. The six remaining Oracle and OpenAI buildings, running Nvidia GB200 Blackwell racks, are due mid-2026, Tom's Hardware noted. Getting a gigawatt of data center power online requires substation upgrades, transmission interconnections, on-site generation, and cooling infrastructure at scales that commercial grids were not designed to absorb all at once. The headline number is a commitment, not a switch being thrown.

The site Oracle and OpenAI abandoned was not sitting empty long. Meta had moved to evaluate taking the expansion capacity, Reuters reported, citing sources familiar with the matter. Then Microsoft moved twice in days: a 700 megawatt lease signed March 24, Bloomberg and Reuters separately reported, followed by the 900 megawatt expansion commitment announced March 27. The speed of the land-grab is the story.

Behind the scenes, Nvidia paid approximately $150 million to ensure its chips, rather than AMD's, would be used in the expanded data center, Dallas News reported, citing Bloomberg News reporting that first documented the Oracle and OpenAI exit. That is not a compute subsidy — it is a competitive gatekeeping fee. Nvidia wanted AMD excluded from a high-profile AI training cluster at the deal structure level, and it paid accordingly. The chipmaker is now competing for infrastructure positioning as aggressively as the hyperscalers themselves.

Microsoft's urgency has more than one driver. The company holds roughly 27 percent of OpenAI, Fortune reported, making OpenAI simultaneously its biggest cloud customer and a strategic rival as it builds independent compute capacity. OpenAI's head of compute infrastructure, Sachin Katti, said plainly that the organization considered expanding at Abilene but chose other locations instead, Fortune separately reported. OpenAI has more than half a dozen sites under development across the United States, Katti said, including one it is building with Oracle in Wisconsin. The Abilene site is not the only game — it is one of several.

The gas infrastructure story is less covered. The existing Oracle and OpenAI project at Abilene runs a 350-megawatt gas-fired power plant as backup, Fortune reported. Microsoft is building a larger gas-fired power plant on the same campus — a detail that got less coverage than the headline number. The language in filings and press materials was vague enough that the scale of what was being accommodated did not get scrutiny. Sam Altman, visiting Abilene last year, was direct about the near-term fossil fuel dependency: "We're burning gas to run this data center," Fortune reported, adding that the long-term hope for Stargate is to rely on other power sources. The velocity of AI infrastructure is currently measured in gas turbines, not renewable energy curves.

The velocity, so far, is also uneven. The 200 megawatts running against a 1.2 gigawatt committed build is not a rounding error — it is a supply chain and permitting reality that the headline capacity numbers paper over. Crusoe's first Abilene project, two 100-megawatt buildings, was constructed and energized in under one year, the company noted, setting a "new standard for large-scale AI infrastructure delivery." That is genuinely fast. But scaling from 200 megawatts to 1.2 gigawatts at a single site, in the timeframe remaining, requires solving interconnection, water, and cooling constraints that no one has demonstrated at this specific density in West Texas.

What this week tells us is that the AI infrastructure boom is a competition for electrons first, silicon second. Companies are signing leases for power they cannot yet draw, paying deposits to keep competitors off sites not yet built, and using capacity announcements to signal market position as much as engineering reality. Microsoft locked up 900 megawatts at a site its closest AI partner had already rejected. That is not an accident. It is the market communicating something the press release language obscures: the bottleneck in AI infrastructure is no longer chip supply. It is the grid.

[auto-normalized] 5110 → story_5110