Trends in AI Supercomputers

TL;DR

The paper analyzes trends in AI supercomputers by compiling a dataset of over 500 AI systems from 2019 to 2025 and defining AI supercomputers as systems capable of training large AI models with at least 1% of the leading system's performance. It shows that leading AI compute grows exponentially, with performance roughly doubling every nine months, while power use and hardware cost rise about yearly, and energy efficiency improves due to newer chips. Private industry now dominates AI compute, owning about 80% of total performance by 2025, and the US accounts for roughly 75% of global AI supercomputer performance. Projections suggest that by 2030 the leading system could reach very large scales in chips, cost, and power, prompting potential decentralization of training; these trends carry significant policy and national-competitiveness implications, though data coverage remains incomplete and biased toward certain sectors and regions.

Abstract

Frontier AI development relies on powerful AI supercomputers, yet analysis of these systems is limited. We create a dataset of 500 AI supercomputers from 2019 to 2025 and analyze key trends in performance, power needs, hardware cost, ownership, and global distribution. We find that the computational performance of AI supercomputers has doubled every nine months, while hardware acquisition cost and power needs both doubled every year. The leading system in March 2025, xAI's Colossus, used 200,000 AI chips, had a hardware cost of \$7B, and required 300 MW of power, as much as 250,000 households. As AI supercomputers evolved from tools for science to industrial machines, companies rapidly expanded their share of total AI supercomputer performance, while the share of governments and academia diminished. Globally, the United States accounts for about 75% of total performance in our dataset, with China in second place at 15%. If the observed trends continue, the leading AI supercomputer in 2030 will achieve $2\times10^{22}$ 16-bit FLOP/s, use two million AI chips, have a hardware cost of \$200 billion, and require 9 GW of power. Our analysis provides visibility into the AI supercomputer landscape, allowing policymakers to assess key AI trends like resource needs, ownership, and national competitiveness.

Figures (20)

• Figure 1: The performance of leading AI supercomputers (in FLOP/s, for 16-bit precision) has doubled every 9 months (a rate of 2.5$\times$ per year).
• Figure 2: The cost of leading AI supercomputers (in 2025 USD) has doubled roughly every year.
• Figure 3: Share of aggregate AI supercomputer performance over time in the public vs private sector.
• Figure 4: Share of AI supercomputer computational performance by country over time. We are visualizing all countries that held more than a 3% share at some point in time.
• Figure 5: Performance of top-10 leading AI supercomputers, with performance measured in FLOP/s at 16-bit precision, has increased by 2.5$\times$ per year (90% Confidence Interval (CI): 2.4--2.7$\times$). We start the regression in 2019, but consider pre-2019 AI supercomputers to identify which systems were in the top-10 at the start of 2019. Our pre-2019 data is too limited to include in the regression. We highlight some notable systems.