Inside Stargate: Bloomberg gets exclusive access to OpenAI's $500B Abilene, Texas datacenter build

Summary

OpenAI, Oracle, and SoftBank's $500 billion Stargate project has moved from announcement to ground-truth construction. Bloomberg gained exclusive access to the first site in Abilene, Texas, a 900-acre facility that will cost $12 billion to build, according to Chase Lochmiller, co-founder and CEO of Crusoe, the startup managing the data center development.

Execution, not money, is the constraint. The three partners—none with hyperscaler experience—have pledged capital they say they can deploy incrementally. OpenAI and SoftBank are each committing $19 billion to start. Oracle and Abu Dhabi's MGX are on the hook for $7 billion each. That reaches roughly $52 billion in initial equity, with JPMorgan providing debt financing and structured deals to follow for future sites. SoftBank founder Masayoshi Son framed the approach directly: we don't need $500 billion in one day. We'll go step by step. Primary Digital Infrastructure is orchestrating the capital structure across sites.

The Abilene facility will have 1.2 gigawatts of capacity. Local utilities could not meet that demand, so Crusoe built its own natural gas power plant on site, enabled by Texas's deregulated energy market. The data center is projected to reach full operational status slightly more than two years from first shovels in June 2024, though parts will go online sooner.

Crusoe promised local officials 357 full-time jobs once construction finishes. For a $12 billion build, that staffing level is remarkably thin. The company imported workers from across the country because Abilene's 129,000 residents could not supply the construction labor needed. Lochmiller told Bloomberg he has had to learn infrastructure development on the job, a candid acknowledgment that none of these companies have built anything at this scale before.

The underlying question is whether compute scaling still works. The project traces back to a 2019 OpenAI research paper on scaling laws, co-authored by Dario Amodei, who later founded Anthropic. That paper argued AI capability requires exponentially more data and compute. Amodei has since publicly questioned Stargate's seriousness, calling it chaotic. Elon Musk called it fake. The math creates pressure: if GPT-4 training cost $100 million and Stargate infrastructure costs $12 billion per site, the model outputs have to justify a 100x leap in capex. That payoff is not obvious, especially if synthetic data hits diminishing returns or if the architecture fundamentally changes before these facilities are done.

Sam Altman has dismissed critics. He argues the constraint is real: GPUs are scarce, rate-limiting remains endemic with users throttled by Google and other cloud providers even as paying customers, and compute capacity is a precious resource. OpenAI needs more compute and more capital to scale AI development and deployment. Whether that claim is correct or rationalization for the build will likely determine whether Stargate becomes a cornerstone of AI infrastructure or a cautionary tale about capex discipline.

Stargate was revealed at the White House on day two of Trump's second term, with Altman, Son, and Oracle chairman Larry Ellison standing behind the president. The whole production came together in days, according to event planning sources. Altman chose the name because an early data center design resembled the ring-shaped portal from the 1990s sci-fi film. It is now a live construction site in red-clay Texas, staffed day and night, with cooling tubes dangling from ceilings and GPU racks being installed in real time.