Power bottleneck looms as AI supercomputers double every 9 months

The research reveals that the computational performance of leading AI supercomputers has been doubling every nine months, a pace of growth that dwarfs traditional supercomputing progress. This explosive trend is largely fueled by two symbiotic factors: a 1.6x annual increase in the number of AI chips used per system and a matching 1.6× improvement in performance per chip. Collectively, these shifts are enabling performance increases of 2.5x per year.

AI supercomputing is accelerating at a staggering pace, with its power, cost, and strategic implications drawing increasing global attention. A new study titled "Trends in AI Supercomputers", published in April 2025 on arXiv, delivers a sweeping analysis of the industrial-scale compute powering today’s artificial intelligence breakthroughs.

Compiled by researchers from Georgetown University, Epoch AI, and the Centre for the Governance of AI, the study constructs a dataset of 500 AI supercomputers spanning 2019 to 2025, offering a detailed portrait of trends in compute power, energy use, hardware cost and global distribution. This in-depth survey not only maps the rise in computational performance but also spotlights the rapidly growing dominance of the private sector and outlines a provocative forecast through 2030.

How are AI supercomputers evolving in power and scale?

The research reveals that the computational performance of leading AI supercomputers has been doubling every nine months, a pace of growth that dwarfs traditional supercomputing progress. This explosive trend is largely fueled by two symbiotic factors: a 1.6x annual increase in the number of AI chips used per system and a matching 1.6× improvement in performance per chip. Collectively, these shifts are enabling performance increases of 2.5x per year.

As of March 2025, the world’s most powerful AI system, xAI’s Colossus, operated on 200,000 high-performance NVIDIA chips, demanded 300 megawatts of power (comparable to 250,000 U.S. homes), and cost an estimated $7 billion. This represents an astonishing leap from the 2019 state-of-the-art system, Oak Ridge’s Summit, which utilized 27,648 GPUs and consumed just 13 megawatts.

If these trends continue, by 2030, the leading AI supercomputer could require:

• 2 million AI chips

• $200 billion in hardware costs

• 9 gigawatts of power—equivalent to 9 nuclear power plants

While chip production and capital expenditure seem to be keeping pace with this trajectory, power availability is expected to become a major bottleneck. The researchers suggest that decentralized training, splitting workloads across multiple sites, may be a necessary adaptation.

Who owns these AI supercomputers and where are they located?

Ownership of AI supercomputing power has shifted dramatically toward the private sector. In 2019, companies controlled just 40% of aggregate AI supercomputer performance. By 2025, that number surged to 80%, with governments and academia relegated to less than 20%.

This transition has profound implications. Private AI supercomputers grew at an annual rate of 2.7x, outpacing the public sector’s 1.9x, fueled by skyrocketing investments, commercial incentives, and demand for generative AI products. Private firms are now driving AI progress at the frontier, while public-sector systems like El Capitan lag behind in relative power and application reach.

Geopolitically, the United States leads the global race. In 2025, U.S.-based AI supercomputers account for roughly 75% of total global AI compute capacity, with China a distant second at 15%. The European Union and other traditional computing powers, including Japan and Germany, now play marginal roles.

This dominance stems from U.S. leadership in cloud infrastructure, AI development, and chip design. U.S. firms have trained 18 of the 25 largest AI models, and the government has implemented export controls to preserve strategic advantages over rivals, particularly China.

Can this trajectory of compute and investment be sustained?

While advances in AI chip performance and supply chains appear to be on track to support current growth, power constraints loom large. A single 9 GW system, if built, would be the most power-hungry industrial complex ever created. Building it would require overcoming enormous regulatory, technical, and social hurdles. The study flags this as the single most likely limiting factor for continued exponential growth.

Moreover, the ballooning costs of AI infrastructure raise broader economic questions. While leading firms like Microsoft, AWS, and OpenAI have announced plans to spend upwards of $500 billion on AI infrastructure by the end of the decade, such investments are only justifiable if AI applications consistently generate transformative returns.

The privatization of AI supercomputing also introduces serious visibility and equity issues. Academic researchers face diminished access to cutting-edge compute, and governments struggle to track the rapidly evolving AI capabilities hidden behind corporate firewalls. The study urges policymakers to consider mandating infrastructure reporting and international intelligence-sharing to mitigate these risks.

The researchers stress that compute infrastructure has become a strategic pillar of national AI competitiveness. Tracking its growth and distribution is not just a technical exercise - it is a geopolitical imperative.