AI supercomputers may cost $200B by 2030, demand power of 9 nuclear plants

Power demand is outpacing efficiency gains

While data centers are becoming more efficient, it’s not enough to curb overall energy needs.

The study found that computational performance per watt has improved 1.34x annually between 2019 and 2025.

Yet, over the same period, power requirements have climbed 2x each year.

“AI supercomputers are improving in energy efficiency, but the shift is not quickly enough to offset overall power growth,” notes Epoch AI.

Their estimate shows that 9GW of power could keep the lights on in 7 to 9 million homes — a scale far beyond today’s capabilities.

Currently, the most advanced AI supercomputer is xAI’s Colossus, which was built in 214 days and cost $7 billion.

It runs on 200,000 AI chips and consumes 300 megawatts of electricity, equal to the usage of 250,000 households.

Industry now dominates the AI supercomputer landscape

Back in 2019, private companies owned 40% of AI supercomputing power. By 2025, that figure has surged to 80%. The shift reflects how AI has moved from academic exploration to a commercial goldmine.

“Where once supercomputers were used just as research tools, they’re now being used as industrial machines delivering economic value,” explains the Epoch AI paper.

As a result, the size and scope of private sector systems have grown rapidly, at 2.7x annually compared to 1.9x for public sector machines.

OpenAI’s $500 billion Stargate project and NVIDIA’s recent $500 billion commitment are just two examples of the massive capital flowing into AI infrastructure.

The U.S. leads global AI computing

According to the study, the United States controls about 75% of all AI supercomputing power today.

China follows with 15%, while traditional supercomputing nations like Japan and Germany play smaller roles.