$1.4 trillion. That was OpenAI's original compute spending target over eight years. The new number? About $600 billion by 2030. The company is pulling back from building its own data centers and shifting to leasing cloud capacity - a strategic retreat driven by infrastructure failures, stalled negotiations, and the reality of going public.
At BlackRock's U.S. Infrastructure Summit last week, Sam Altman was unusually candid about the difficulties. "Anything at this scale, it's just like so much stuff goes wrong," he said.
The Abilene Collapse
The most concrete casualty is the planned expansion of OpenAI's flagship Stargate facility in Abilene, Texas. The existing site - built by Crusoe on Lancium's Clean Campus - has two operational data center buildings with six more due this year, totaling about 1.2 gigawatts of capacity. The plan was to push that to 2.0 GW with an additional 800 MW expansion.
That expansion is dead. Three factors killed it:
- Negotiations between Oracle, Crusoe, and OpenAI dragged on since mid-2025 over financing terms
- OpenAI kept changing its demand forecasts, making it hard for partners to commit
- A multi-day outage earlier this year, caused by winter weather damaging liquid cooling equipment, reportedly soured the relationship between OpenAI and Crusoe
Nvidia stepped in with a $150 million deposit to Crusoe to secure the site for its own chips, and is now brokering talks for Meta to lease that abandoned 800 MW capacity. Meta, which plans to spend up to $135 billion in capital expenditure this year, is evaluating the deal.
From Builder to Buyer
The bigger story is OpenAI's strategic pivot from building infrastructure to buying it. The centerpiece: a $100 billion, eight-year cloud services deal with AWS. As part of that arrangement, Amazon is investing $50 billion in OpenAI during a record $110 billion fundraising round that values the company at $840 billion.
The AWS deal comes with some unusual terms. OpenAI committed to 2 GW of capacity using Amazon's custom Trainium3 and upcoming Trainium4 chips. AWS will exclusively host OpenAI's new "Stateful Runtime Environment" on Amazon Bedrock - an architecture that lets AI agents maintain persistent memory across multi-day tasks. Microsoft keeps exclusive rights for OpenAI's standard API calls and ChatGPT.
That split has created friction. Microsoft is reportedly weighing legal action, believing the $50 billion AWS deal violates its exclusive Azure agreement with OpenAI.
The IPO Math
All of this is driven by one thing: OpenAI plans to go public in late 2026 or early 2027 and wants to raise at least $60 billion. Public market investors have a different appetite for capital expenditure than the venture investors who funded OpenAI's $20 billion annual revenue run rate.
Building your own data centers means massive upfront costs that look terrible on a pre-IPO balance sheet. Leasing shifts that spending from capital expenditure to operating costs, spread over time. It is a financial reframing as much as a strategic one.
Altman also revealed that OpenAI is developing its own inference-only chips (inference is the process of running a trained model to generate responses, as opposed to training the model). These are designed to be the "cheapest inference chip" and "most efficient per watt," with first deployments expected by the end of this year. The long-term vision, as Altman put it: "We see a future where intelligence is a utility, like electricity or water, and people buy it from us on a meter."
The existing 1.2 GW Abilene site and Oracle's broader 4.5 GW development agreement remain intact. But the era of OpenAI trying to build everything itself appears to be over before it really started.