Oracle and OpenAI scrap planned Abilene, Texas data-centre expansion

Oracle and OpenAI ended talks over a planned expansion of their flagship Stargate data centre in Abilene, Texas, capping the campus at 1.2 gigawatts rather than the 2.0 gigawatts originally planned across six additional buildings intended to house roughly 400,000 Nvidia Blackwell chips. Reporting attributed the decision to power-grid interconnection delays exceeding a year, with OpenAI’s compute lead, Sachin Katti, saying the additional capacity would instead be placed at other sites. Winter-weather outages earlier in 2026 had also exposed cooling problems at the existing campus, adding to the case against building further there.

The companies’ wider agreement, covering up to 4.5 gigawatts of capacity across multiple sites, remained in force; the pullback affected only the Abilene add-on rather than the broader Stargate build-out. Nvidia reportedly put down $150 million to secure the vacated expansion site and was said to be courting Meta as a replacement tenant, while OpenAI redirected some planned investment toward Nvidia’s newer Vera Rubin chips at other locations.

The episode illustrated a constraint on frontier AI infrastructure that compute and chip supply alone could not solve: physical grid interconnection queues, which by 2026 were running well over a year in parts of Texas and elsewhere, had become as binding on data-centre buildout timelines as capital or hardware availability. It also complicated the framing of Stargate, the joint venture between OpenAI, Oracle and SoftBank announced in January 2025 as a $500 billion commitment to US AI infrastructure: a scaled-back single site did not reduce the venture’s overall multi-site targets, but it showed those targets running up against physical limits that money could not immediately buy past.