The AI race has fundamentally shifted its focus. It is no longer a contest of elegant algorithms, but a battle for raw electrical power. Sam Altman’s OpenAI recently announced it has secured 10 gigawatts (GW) of computing capacity in the U.S.—a bold claim, especially considering the company originally aimed to hit this milestone only by 2029.
To put this scale into perspective, 10 GW can power roughly 7.5 million American homes. In effect, the company is staking a claim on the energy consumption of a major metropolis, attempting to reserve the physical world's resources for its needs before its competitors can react.
The speed at which OpenAI is "consuming" gigawatts suggests an aggressive sprint: according to their own data, three of those ten gigawatts were contracted in just the last 90 days. The primary driver of this paper victory is a 2 GW deal with Amazon. Rather than breaking ground on its own facilities, OpenAI is opting to leverage existing hyperscaler infrastructure. For the business community, the signal is clear: the bottleneck has shifted from a shortage of NVIDIA chips to the limited capacity of national power grids. If you don't have a direct line to a nuclear plant, your ambitions in Large Language Models remain purely theoretical.
However, an uncomfortable reality of localized failures hides behind these celebratory figures. While OpenAI cheers for completing its "five-year plan" ahead of schedule, physical construction projects are stalling or shutting down. The ambitious Stargate project in Texas—positioned in early 2024 as a $500 billion collaboration with Oracle and SoftBank—is effectively frozen. A similar fate befell sites in Norway and the UK, where OpenAI simply could not stomach the local electricity costs.
This creates a strange dissonance: the company claims to be four years ahead of schedule while simultaneously scaling back expansion in key regions. We are likely witnessing a classic bait-and-switch. OpenAI is trading difficult, ground-up infrastructure projects for quick-fix capacity rentals from cloud giants. This allows for impressive investor reporting today, but leaves a critical question unanswered: where will Altman find real electricity once the cloud reserves of Amazon and Microsoft run dry and his own nuclear dreams remain confined to slide decks?