Trends in AI Supercomputers is an empirical report by Konstantin F. Pilz (Georgetown), James Sanders (Epoch AI), Robi Rahman (Epoch AI), and Lennart Heim (Epoch AI / GovAI), published through Epoch AI and arXiv (2052.06106v2) and released in April 2025. It assembles a dataset of 500 AI supercomputers spanning 2019 to 2025 and characterizes their doubling rates, ownership shifts, geographic distribution, and a forward extrapolation to 2030 (Source: https://epoch.ai/blog/trends-in-ai-supercomputers).
Summary of findings
The paper documents three exponential trends across the 2019-2025 dataset: computational performance doubling roughly every 9 months, hardware acquisition cost doubling roughly every 12 months, and power needs doubling roughly every 12 months. It uses the xAI Colossus system (March 2025, Memphis) as the 2025 reference point and extrapolates the observed doubling rates to a projected 2030 leading system.
| Quantity | Doubling rate | 2025 leading system (xAI Colossus) | Extrapolated 2030 leader |
|---|---|---|---|
| Computational performance | 9 months | 200,000 AI chips | 2 × 10²² 16-bit FLOP/s |
| Hardware acquisition cost | 12 months | $7B | $200B |
| Power needs | 12 months | 300 MW (≈ 250,000 households) | 9 GW |
| Chip count | (implied from above) | 200,000 | 2 million |
The Colossus system is presented as the leading 2025 example, with per-system metrics of 200,000 chips, $7 billion in hardware acquisition cost, and 300 MW of power (roughly the consumption of 250,000 households). The 2030 projections — $200 billion in hardware and 9 GW of power for a single system — are cited in subsequent AI-policy and energy-policy commentary.
Geographic distribution
The paper finds AI supercomputer performance heavily concentrated in the United States, which holds about 75% of the global total, against about 15% for China and roughly 10% for the rest of the world combined.
| Country | Share of total AI supercomputer performance |
|---|---|
| United States | ~75% |
| China | ~15% |
| Rest of world | ~10% combined |
The authors note that the US-China gap on this metric is wider than gaps reported on benchmark performance, capability, or model release frequency.
Ownership trajectory
The dataset shows the share of total AI supercomputer performance held by companies expanding over the period, while the share held by governments and academia declined. The authors describe this as a shift from scientific tools to industrial machines:
"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."
This evidence on the private-sector expansion of AI compute is drawn on by the Buchanan & Collins "Grand Bargain" argument that AI infrastructure has outgrown laissez-faire policy and by America's AI Action Plan industrial-policy framings.
Relation to other sources
The dataset's power-doubling-every-year trend is the calibration basis for the "50 GW of new power for AI by 2028" estimate in The AI Grand Bargain — Ben Buchanan and Tantum Collins (Foreign Affairs, October 2025); the paper's 2030 extrapolation of 9 GW per leading system sits at the upper bound of associated energy-demand projections. The per-system power-and-water footprint argument in AI Environmental Impact uses the paper as its primary reference. The 9 GW 2030 leading-system projection has been invoked alongside the Trump DPA Determination 2026-10, on the reading that a single 2030-class system at 9 GW would exceed many state-level baseload capacity additions, making the DPA Section 303 designation of grid hardware as defense-essential conceptually consistent.
The 75-15 US-China supercomputer split is cited as evidence for the industrial-base arguments in America's AI Action Plan, and bears on BIS Framework for AI Diffusion — Interim Final Rule (Jan 13, 2025): a 75% US share means export controls operate from a position of structural advantage, and the distribution shapes the rule's effective scope. The 9-month performance-doubling pace is consistent with continued scaling-law-driven progress, providing empirical grounding for Scaling Laws at the supercomputer level. The transition from government and academic to industrial compute connects to AI and National Security and to Sullivan's argument that the United States should integrate AI into national security through industrial cooperation rather than government-led R&D.
The paper is co-authored by Lennart Heim and anchors Epoch AI's empirical compute-tracking program. Colossus is the xAI compute reference point, and its per-system metrics depend on the fab capacity of Nvidia & TSMC. The report sits alongside the companion Epoch AI papers Epoch AI — Can AI Scaling Continue Through 2030? and Epoch AI — How Much Power Will Frontier AI Training Demand in 2030? as part of the empirical compute-trajectory literature.
Provenance and confidence
Confidence is high for the empirical doubling rates and the 2025 Colossus reference point, which can be cross-checked against xAI's own disclosures. Confidence is lower for the 2030 extrapolations, which assume the observed doubling rates continue — a condition that depends on ASML EUV throughput, fab capacity, energy buildout, and capital availability all remaining on trend. The authors are explicit about these extrapolation assumptions.
Relationships
- supports: AI Environmental Impact (primary reference for per-system footprint), America's AI Action Plan (US-China share evidence), The AI Grand Bargain — Ben Buchanan and Tantum Collins (Foreign Affairs, October 2025) (energy-doubling underpinning), Presidential Determination 2026-10 (Defense Production Act Section 303 — AI Grid Components) (9 GW 2030 system context for grid-hardware DPA designation), Scaling Laws (continued doubling at supercomputer level)
- related: Epoch AI — Can AI Scaling Continue Through 2030? (companion Epoch paper), Epoch AI — How Much Power Will Frontier AI Training Demand in 2030? (companion Epoch paper), Lennart Heim, Epoch AI, xAI (Colossus), Nvidia & TSMC — AI Compute Infrastructure (TSMC fab dependency), BIS Framework for AI Diffusion — Interim Final Rule (Jan 13, 2025) (export-control distributional context)
- instance-of: Compute Governance