The world's largest technology companies are quietly becoming industrial enterprises.
That is the actual story embedded in the satellite data. A site-by-site analysis published Monday by Epoch AI, verified against satellite imagery, shows that OpenAI's Stargate project has 0.6 gigawatts operational across seven US locations sixteen months after launch. The combined planned capacity is over 9 gigawatts, comparable to New York City's peak power demand. Four companies — Amazon, Google, Meta, and Microsoft — now control 42 percent of US data center capacity and spent more than $330 billion on capital expenditures last year. Google is not buying power from a utility. It is co-funding nuclear reactor construction and taking the output directly. Microsoft has committed over $80 billion in fiscal 2026 capex; Alphabet, $175 to $185 billion.
This is not a cyclical surge. It is a structural reorientation of where the technology sector's capital is going — and the satellite-verified specifics show exactly where the gaps are forming.
At Abilene, Texas, the only operational Stargate site, four of eight planned buildings are operational, housing Nvidia Blackwell chips, drawing power from a mix of on-site natural gas and grid power that includes local wind, per Epoch AI. OpenAI had announced plans to expand that site to 2.1 gigawatts. In early March, it dropped those plans. The reason, per OpenAI's head of compute infrastructure Sachin Katti, writing on X: we considered expanding it further, but ultimately chose to put that additional capacity in other locations.
Microsoft moved in the same week.
Crusoe, the AI infrastructure company that built the existing Abilene campus for OpenAI and Oracle, announced a new 900-megawatt AI factory campus for Microsoft on adjacent land. The total Abilene site — two different companies on the same tract — is projected to reach 2.1 gigawatts. Crusoe's CEO Chase Lochmiller called it continuing to build the industrial foundation for American AI, at a velocity the industry has never seen. The first two Crusoe buildings were built and energized in under one year.
Microsoft holds roughly 27 percent of OpenAI, according to Fortune. The two companies are increasingly pursuing AI development separately, even as they occupy the same dirt. OpenAI redirected its Abilene capacity elsewhere. Microsoft built on the adjacent lot with the same builder. The visible behavior is of two entities with aligned incentives but divergent strategies.
The Stargate model separates capital provider from operator. SoftBank and Oracle own the hardware at various sites. OpenAI runs the workloads. This capital-provider-versus-operator structure is common in real estate and energy infrastructure. It is relatively new in AI, and it is one reason Stargate has attracted financing at a scale OpenAI alone could not have raised.
Under the original announcement, the project was supposed to deliver 20 million H100-equivalents of compute — roughly the total amount of AI training capacity that existed in the world by the end of 2025, according to Epoch AI. At full buildout, Stargate would double the world's AI training capacity in a single program.
To sidestep lengthy queues for grid connection, at least three of the seven sites will draw on dedicated on-site natural gas plants. The approach saves time but raises costs. At least six sites use closed-loop liquid cooling, which avoids the water evaporation that draws local opposition at traditional data centers. This is not a project waiting for the grid to be ready. It is a project building around it.
Google's approach differs. In August 2025, Google signed a deal with Kairos Power and the Tennessee Valley Authority to deliver up to 50 megawatts of advanced nuclear power to its data centers by 2030, with a broader agreement to bring 500 megawatts online by 2035. Nuclear construction timelines are longer than gas, but the long-term cost structure and emissions profile are different in kind, not just degree.
The lock-out consequence is structural. A startup or mid-sized AI lab that wants to train a frontier model needs compute. Compute at the frontier now means gigawatts. Gigawatts require capital relationships with infrastructure developers, power companies, utilities, municipalities, and landholders that most AI companies do not have and cannot easily acquire. The hyperscalers can self-finance what smaller players cannot.
McKinsey estimates 156 gigawatts of AI-related data center capacity demand by 2030, with 125 incremental gigawatts added between 2025 and 2030. If that projection holds, Stargate's 9 gigawatts will be absorbed. If it does not, the investors, along with the municipalities and utilities that enabled the buildout, absorb the gap.
OpenAI CEO Sam Altman, visiting Abilene last year, was direct: we are burning gas to run this data center. The long-term hope, he said, is to rely on other power sources. The immediate reality is natural gas.
In Shackelford County, Texas, directly across the county line from Abilene, Vantage is building a 1,200-acre campus with ten buildings on an on-site natural gas microgrid. In Doña Ana County, New Mexico, STACK Infrastructure is developing four large buildings, also on dedicated gas. In Milam County, Texas, SB Energy — a SoftBank subsidiary — is building what it calls a fast-build site, with the first building expected by October 2026. In Port Washington, Wisconsin, Vantage is building a campus it describes as sustainable-by-design, with 70 percent of power drawn from solar, wind, and battery storage. In Saline Township, Michigan, DTE Energy will provide 100 percent of the power. In Lordstown, Ohio, a joint venture between SoftBank and Foxconn is building a manufacturing facility for AI servers alongside a small data center drawing from an existing substation.
The Lordstown site is also the only location where local opposition has produced an outright ban on future data centers. Regulatory risk is real and concentrated.
None of the six non-Abilene sites are operational. The closest completion dates are late 2026. The rest are 2028 or later.
The infrastructure is real. The timeline is long. The money is enormous. And the gap between the ambition and what currently exists is not a temporary delay — it is the shape of a structural bet. The next milestone to watch: whether the next Abilene buildings hit their energization targets or slip.