xAI to Power Data Center with Overseas Power Plant
Source: tomshardware.com
Elon Musk's xAI is planning new data centers that will house millions of AI chips and require significant power. According to Dylan Patel from SemiAnalysis, xAI intends to buy a power plant overseas and ship it to the U.S. Musk confirmed this statement.
xAI's Colossus AI supercomputer already uses 300 MW to power 200,000 Nvidia Hopper GPUs. Powering the next AI data center, which is expected to house one million AI GPUs, presents a significant challenge, potentially requiring as much power as 1.9 million households.
xAI's Computing Resources
Patel noted that xAI has been assembling computing resources and a team of researchers to develop its Grok AI models. xAI has already encountered difficulties in powering its existing data center.
The Colossus data center, located near Memphis, Tennessee, uses 35 gas turbines producing 420 MW and Tesla Megapack systems to manage power usage. xAI is also expanding its facilities, having purchased a factory in Memphis to convert it into additional data center space that can power about 125,000 eight-way GPU servers with networking, storage, and cooling.
Power Consumption
A million Nvidia Blackwell GPUs could consume between 1,000 MW and 1,400 MW, depending on the models (B200, GB200, B300, GB300) and their configuration.
However, power consumption extends beyond GPUs to include CPUs, DDR5 memory, storage, networking, cooling, air conditioning, and power supply inefficiencies. Overhead can add another 30% to 50% to the AI GPU power draw, expressed as PUE (power usage effectiveness). Consequently, a million-GPU data center might consume between 1,400 MW and 1,960 MW (with a PUE of 1.4), equivalent to powering 1.9 million homes.
A solar power plant isn't considered viable for a 24/7 load of this size because it would need gigawatts of panels and extensive battery storage, which is expensive and requires a lot of land. Building multiple natural gas combined-cycle gas turbine (CCGT) plants, each producing 0.5 MW – 1,500 MW, is a more practical option because it can be deployed faster, scaled in stages, and integrated with existing grids. xAI may be considering importing such a plant to the U.S.
While nuclear reactors could also meet the load, they take longer to design, permit, and build. Acquiring a power plant in the U.S. can be a lengthy process, so xAI is reportedly buying a plant overseas and shipping it in.
xAI's Colossus currently produces power onsite and purchases power from the grid, and its next data center will likely combine a dedicated onsite plant with grid interconnections.
The Future of AI Development
Patel suggests that leading AI companies are adopting similar strategies: concentrating compute clusters, hiring researchers, and training larger AI models. To remain competitive, xAI needs to develop more advanced and power-hungry data centers.