Frontier labs used only 21% of global AI compute in 2025, per EpochAI, challenging the narrative of compute concentration.
OpenAI, Anthropic, and xAI used only 21% of global operational AI compute in late 2025, per EpochAI data. The finding contradicts the narrative that frontier labs monopolize AI infrastructure.
Key facts
- 21% of global AI compute used by OpenAI, Anthropic, xAI.
- 16M deployed H100-equivalents worldwide in late 2025.
- 20M H100-equivalents sold, 4M gap indicates deployment lag.
- 79% of compute runs outside the top three frontier labs.
- Data source: EpochAI gradient update.
A new data point from EpochAI reveals that the three most prominent frontier AI labs — OpenAI, Anthropic, and xAI — collectively consumed only about 21% of the world's operational AI compute at the end of 2025. At that point, the global installed base of AI accelerators stood at roughly 16 million deployed H100-equivalents, with an additional 20 million H100-equivalents sold but not yet operational.
The finding, shared by @rohanpaul_ai on X, directly challenges the common assumption that a handful of well-funded labs control the majority of AI compute resources. Instead, the majority of compute — nearly 79% — is spread across other companies, research institutions, and cloud providers. This suggests that the AI compute landscape is far more distributed than the hype around frontier labs implies, and that the total addressable market for AI hardware and services is much larger than the spending of the top three labs alone.
Why This Matters
The data implies that the narrative of an AI compute oligopoly is overstated. While OpenAI, Anthropic, and xAI command significant media attention and investment, the bulk of compute is deployed elsewhere — likely in enterprise applications, smaller AI startups, and non-frontier research. This distribution has implications for hardware demand forecasting, cloud pricing dynamics, and the competitive landscape for AI models. If the top labs do not dominate compute, then the market for inference and training at scale is more fragmented than previously thought.
Context from Prior Reports
Earlier EpochAI reports have tracked the exponential growth in training compute for flagship models like GPT-4 and Gemini, but this new data provides a snapshot of the total installed base. The gap between sold and deployed hardware — 20 million vs. 16 million H100-equivalents — also indicates supply constraints or deployment delays that could affect near-term capacity.
What’s Missing
The source does not specify how the remaining 79% of compute is allocated by sector or geography, nor does it break down usage by model type (training vs. inference). The data is a point-in-time estimate and may not reflect rapid changes in deployment or utilization rates.
What to watch
Watch for EpochAI's next quarterly update on compute distribution, which will reveal whether the top labs' share is growing or shrinking. Also monitor GPU lead times and cloud provider capital expenditure disclosures for signs of deployment acceleration.
[Updated 30 Jun via light_reading]
Meanwhile, South Korea announced plans for an 18.4GW AI data center megaproject to be built by 2035 at an estimated cost of $650 billion, in partnership with SK Group, GS Group, and Naver [per Light Reading]. This massive buildout, if realized, would dramatically increase global compute supply and could further dilute the share held by frontier labs.
Originally published on gentic.news
Top comments (0)