When Microsoft quietly signed a lease last week for a 700-megawatt data center in Abilene, Texas, the announcement from Oracle and OpenAI was characteristically sunny. We are very proud of our relationship and our progress in bringing capacity online, Oracle said. That statement appeared designed to reassure anyone watching that the Stargate project was fine, the partnership was strong, and nothing to see here.
But Microsoft had already signed a lease with Crusoe Energy for a facility that Oracle and OpenAI spent most of 2025 designing and then abandoned. The Abilene site sits directly next door to the flagship Stargate campus that Oracle and OpenAI announced with enormous ceremony in January 2025. Reuters first reported the Microsoft-Crusoe lease on Tuesday. Oracle denied any delay. The relationship, the company implied, was proceeding in lockstep.
The documents tell a different story.
According to Reuters and Dallas News, Oracle, OpenAI, and Crusoe had been negotiating since mid-2025 to expand the Abilene complex from 1.2 gigawatts to roughly 2 gigawatts — doubling the footprint of one of the world's most ambitious AI data center builds. The expansion never happened. And the reasons for that collapse reveal fault lines that the press release language was designed to paper over.
Nvidia, for its part, was not a passive infrastructure supplier watching from the sidelines. The chip designer paid Crusoe a $150 million deposit and personally helped broker Meta as a replacement tenant for the expansion buildings, according to reporting by Reuters and Dallas News. The money was not charitable: Nvidia was paying to keep Advanced Micro Devices, a competitor, out of a facility it had helped design. One of the world's largest semiconductor companies wrote a nine-figure check to determine which chips would power a building it did not own. That is infrastructure politics, not just chip sales.
Relations between Oracle and Crusoe, the operator, had also frayed considerably. Earlier this year, multiple data center buildings at the Abilene site went offline for days because winter weather disrupted liquid cooling machinery, Dallas News reported. Crusoe is not a mature operator at this scale — the failures suggest the engineering challenges of running a 1.2 gigawatt AI factory in West Texas are not purely theoretical. Oracle and OpenAI were negotiating expansion terms while the existing facility was having multi-day outages. That context makes the abandonment of the expansion look less like a financing disagreement and more like a confidence problem.
Microsoft, meanwhile, appears to have seen an opportunity in someone else's retreat. The company has massive Azure AI compute commitments and a stated goal of building out hyperscale capacity faster than anyone else. Renting a 700-megawatt facility from Crusoe — with Stargate's Oracle campus sitting next door, presumably already connected to the local grid and permitting infrastructure — is a cleaner path to capacity than starting from scratch. Microsoft gets a nearly-built facility at a site that already has power and connectivity. The fact that it negotiated this deal at all suggests the Stargate expansion was more fragile than the warm statements implied.
The Stargate initiative — a joint project of SoftBank Group, OpenAI, and Oracle, announced by President Donald Trump in January 2025, with up to $500 billion and 10 gigawatts of planned capacity — has now seen its first visible contraction. Oracle and OpenAI confirmed in early March that they had ended plans to expand the Abilene site, according to Reuters and Data Center Dynamics. The Microsoft lease, reported by Data Center Dynamics, suggests the contraction was not a pause but a retreat: someone else is now occupying the infrastructure Oracle and OpenAI decided not to fill.
Meta, which Nvidia reportedly recruited as a backup tenant, is spending lavishly on AI infrastructure regardless. The company has projected capital expenditures of up to $135 billion in 2026 alone, per Dallas News. If Meta does end up as a significant tenant at Abilene — or at one of the adjacent facilities — it would mark a notable consolidation of AI training compute among a small number of very large players, all clustered in the same Texas grid corridor.
The broader pattern is becoming legible: the public posture of AI infrastructure commitment — announcements, partnerships, presidential appearances — and the private allocation decisions are not always the same thing. Oracle and OpenAI announced a 2-gigawatt expansion, negotiated for most of a year, and then walked. Nvidia paid $150 million to influence which chips would be inside. Microsoft rented capacity that its partners abandoned. Crusoe's buildings went dark for days in cold weather, and the expansion talks ended anyway.
Each company has a rational explanation for its own moves. Taken together, they describe an industry that is spending hundreds of billions of dollars on infrastructure it is still figuring out how to operate — while insisting, publicly, that everything is proceeding exactly as planned.
Watch for whether Crusoe's operational reliability becomes a recurring issue as more hyperscale capacity comes online in Abilene. Also watch whether Microsoft's lease with Crusoe signals a broader shift toward developers renting capacity from specialists rather than building their own — a structural change in how AI compute gets delivered, if it holds.
