Sam Altman reportedly pitches major AI datacenter development plans directly to Biden administration officials.

OpenAI CEO Sam Altman is reportedly trying to convince the Biden administration that an enormous network of AI datacenters, each consuming up to five gigawatts of power, is imperative to ensuring US national security and maintaining its technological lead over China. According to a document reviewed by Bloomberg, Altman may be planning as many as 5-7 such datacenters across the United States, though he will start with one. This proposal follows recent meetings between Altman and other tech leaders at the White House to discuss the proliferation of AI technologies and infrastructure.

Building even one of these facilities presents extraordinary power challenges. Five gigawatts requires roughly equivalent output to five pressurized water nuclear reactors. Such power stations would be among the largest in the US, second only to the Grand Coulee hydro plant in Washington state, which has a rated capacity of 6.8 gigawatts. The next largest plants—Georgia's Alvin W. Vogtle and Arizona's Palo Verde—are nuclear facilities capable of generating about 4.6 and 3.9 gigawatts, respectively. A CBRE report from late last month found that shortages of power and equipment necessary to harness it are already leading to delays in datacenter developments.

Cloud providers are already taking extreme measures to secure sufficient power. Last week, Microsoft announced a 20-year power purchase agreement with Constellation Energy to bring the 837-megawatt Three Mile Island Unit 1 nuclear power plant back online. Earlier this year, Amazon acquired Cumulus datacenter facilities from Talen Energy for $650 million, which will eventually provide access to up to 960 megawatts of power from the Susquehanna nuclear plant. Oracle's founder Larry Ellison is even considering deploying small modular reactors to fuel the database giant's AI expansion.

Sourcing enough accelerators to fill these datacenters represents another significant hurdle. Assuming a power use effectiveness of 1.1, a five-gigawatt facility could support in excess of 35,000 Nvidia Grace-Blackwell NVL72 rack-scale systems—or roughly 2.5 million Blackwell GPUs. The Uptime Institute estimated that Nvidia shipped roughly 600,000 H100s in 2023 and predicted the GPU giant would ship somewhere between 1.5 and 2 million chips this year. Beyond that, TSMC's CoWoS packaging capacity, already constrained, may struggle to keep pace with demand.

Altman's history of proposing ambitious projects adds uncertainty to the plan's seriousness. Earlier this year, he reportedly floated a $7 trillion project to establish a network of chip factories. However, speaking at Intel's Foundry event in February, the OpenAI CEO reminded observers that "not everything you read on the internet is true." The datacenter proposal may ultimately be an exercise in encouraging the US government to consider long-term investments necessary to support AI development, rather than a concrete plan.