Nvidia is buying power infrastructure. Not chips. Not networking gear. Not even a model lab. On Friday, The Information reported that the chipmaker is putting up to $3 billion into Lancium, the Blackstone-backed developer behind the OpenAI and Oracle Stargate campus in Texas — $2 billion now for roughly a 20% stake, and another $1 billion conditional on the company securing additional power capacity. It is, by some distance, the largest single corporate commitment Nvidia has made outside its core silicon business, and the reasoning behind it explains a great deal about the shape of the AI buildout for the rest of this decade.
The bottleneck moved. For most of the last three years, the AI conversation has been about compute. H100s, then H200s, then Blackwell, and now Rubin. Each generation shipped more FLOPS per dollar, and each generation tightened the supply curve for the next. The unspoken assumption was that compute was the scarce input and that everything else — power, land, water, cooling, fibre — would accommodate whatever the chip roadmap produced. The Stargate project, which OpenAI and Oracle announced in January 2025 with the backing of SoftBank and MGX, was supposed to be the proof of that thesis at scale. It is not. Stargate is roughly 18 months behind schedule, and the reason is not chips. It is megawatts.
Lancium is a power developer, not a data centre company. That distinction matters. Lancium does not build server halls. It builds behind-the-meter generation — gas turbines, grid-scale batteries, and increasingly direct connections to wind and solar farms in West Texas — and it packages them into contracts that hyperscalers can underwrite. Its pitch to the largest AI labs is that power is now a product, that it can be delivered faster than utility interconnect, and that it can be priced in ways that match the depreciation curves of GPU clusters rather than the decades-long contracts utilities prefer. The OpenAI/Oracle campus in Abilene is the reference deployment. Blackstone, which has been the dominant financier of data centre real estate for a decade, took a majority stake in Lancium in 2023. Nvidia is now joining them.
Why Nvidia, and not OpenAI or Oracle? This is the part that should sharpen every founder's mental model. OpenAI and Oracle are the offtake customers. They have already signed long-dated power purchase agreements with Lancium for the Abilene campus and the next few sites. They do not need to own the power developer — they need the electrons. Nvidia, on the other hand, sells the machines that consume the electrons. If the GPU roadmap keeps accelerating and the demand curve keeps outrunning the build rate of new power, then every quarter of delay at the Stargate campus is a quarter of unshipped systems. By taking an equity stake, Nvidia is converting part of its customer-acquisition cost into infrastructure ownership. The math is brutal: an H200 rack pulls about 1.2 megawatts under load. A 100,000-GPU cluster — which is the unit of planning for any frontier lab in 2026 — is 120 megawatts of continuous draw, or roughly the load of a mid-sized American city. Multiply that across the announced Stargate sites and you can see why Jensen Huang decided that financing the power stack is now part of his job.
What this means for the rest of the industry. Three implications land immediately. First, the major chip companies are going to start writing power-related cheques. AMD has a small nuclear PPA with Constellation; that is going to grow. Intel's foundry strategy looks increasingly threadbare without an equivalent. Second, the venture ecosystem for grid-edge and behind-the-meter power — the companies sitting between utilities and hyperscalers — is now one of the most attractive categories in climate tech. Third, the geographic distribution of AI infrastructure is going to be decided less by where fibre terminates and more by where the gas turbines, batteries, and grid interconnect can be brought online in 24 to 36 months rather than the 7-to-10-year utility cycle. Texas, West Virginia, Wyoming, and the Permian Basin — not Loudoun County, not The Dalles — are the next data centre corridors.
The deeper signal. Nvidia's market cap is now larger than the entire GDP of Texas. The company could afford this deal out of a single quarter of free cash flow. But it chose to do the deal, and to do it as equity rather than as a power purchase agreement, because the binding constraint is no longer priced by the chip market. The chip market will deliver more FLOPS every six months. The power market will not deliver more megawatts on that cadence. So Nvidia is now a power company with a chip business attached. Every founder building in AI infrastructure should update their mental model accordingly: the buyer of your product is increasingly the person trying to keep a 120-megawatt cluster cool, not the person trying to train the next frontier model.
Photo by Mariia Shalabaieva on Unsplash
