OpenAI CEO Sam Altman has presented a proposal to the Biden administration advocating for the construction of a vast network of AI datacenters across the United States. Each of these datacenters would consume up to five gigawatts of power, aiming to bolster U.S. national security and maintain its technological edge over China.
Massive Energy Requirements
I initially reported on OpenAI's infrastructure plans earlier this month, including seeking outside investment to fund the project. Altman along with other executives at OpenAI have been actively meeting with potential investors to fund the project. Key interest is expected from investors in countries like Canada, Korea, Japan, and the United Arab Emirates. Microsoft which holds a significant stake in OpenAI, is likely to join this venture.
Altman's propsal to the government, reviewed by Bloomberg, suggests building several such facilities, with initial plans for one datacenter. Each facility would require an immense amount of power, equivalent to the output of five pressurized water nuclear reactors. The power stations needed for these datacenters would be among the largest in the United States, second only to the Grand Coulee hydro plant in Washington state, which has a capacity of 6.8 gigawatts.
The next largest power plants, such as Georgia's Alvin W. Vogtle and Arizona's Palo Verde, are nuclear facilities with capacities of approximately 4.6 and 3.9 gigawatts, respectively. Excess capacity is already a limiting factor for many datacenter developments, as highlighted by a CBRE report indicating that power shortages and equipment constraints are causing delays.
Strategic Partnerships and Challenges
Cloud providers are already taking significant steps to secure power for their AI operations. Microsoft has entered a 20-year agreement with Constellation Energy to restart the 837-megawatt Three Mile Island Unit 1 nuclear power plant. Similarly, Amazon has partnered with Talen Energy, acquiring the Cumulus datacenter facilities and gaining access to up to 960 megawatts of power.
Oracle's founder Larry Ellison has also discussed deploying small modular reactors to support the company's AI expansion. However, sourcing enough accelerators to fill these datacenters remains a challenge. A five-gigawatt facility could support over 35,000 Nvidia Grace-Blackwell NVL72 rack-scale systems, or roughly 2.5 million Blackwell GPUs, assuming a power use effectiveness (PUE) of 1.1.
Last Updated on September 30, 2024 1:40 pm CEST
