Artificial intelligence (AI)-powered supercomputers are likely to face significant power constraints by 2030, according to research institute Epoch AI. In a recent paper, Epoch AI projected that the leading AI supercomputer in June 2030 will require 2 million AI chips, cost $200 billion, and need 9GW of power. This power requirement is equivalent to the output of nine nuclear reactors, a scale that surpasses any existing industrial facilities.
Challenges and Trends
Epoch AI’s research highlights that while the first two requirements – chips and capital – can likely be met based on historical trends and major capital commitments like the $500 billion Project Stargate, the power constraint poses a significant challenge. The institute noted that since 2019, AI supercomputers’ computation performance has grown 2.5 times per year, while their power requirements and hardware costs have doubled each year.
The computational performance of AI supercomputers has been driven by the use of more and better chips. Since 2019, chip quantity has increased 1.6 times per year, and chip performance has also improved by 1.6 times per year. Notably, while systems with more than 10,000 chips were rare in 2019, several companies have deployed AI supercomputers more than 10 times that size in 2024. For instance, xAI’s Colossus features 200,000 AI chips.
Industry Shifts and Solutions
To overcome the impending power constraint challenge, companies may shift to decentralized training approaches. This would allow them to distribute their training across AI supercomputers in several locations. Epoch AI’s research also revealed that the share of AI supercomputers’ computing power owned by companies rather than the public sector has risen from 40% in 2019 to 80% in 2025. Additionally, 75% of AI supercomputers’ computing power is hosted in the United States, with China hosting the second-largest share at 15%.
Notable Developments
xAI launched its Colossus 100k H100 training cluster in September, with owner Elon Musk announcing plans to double its size to 200k (including 50k H200s) within months. Project Stargate, announced by former President Donald Trump in January, aims to build large AI-focused data centers in the U.S., starting with 10 facilities in Texas. The first of these will span 500,000 square feet.
As the AI supercomputer industry continues to evolve, addressing the power constraint will be crucial for future development and deployment.
