Frontier AI labs own a surprisingly small slice of the world’s AI compute. A May 21 analysis by Epoch AI researcher Josh You estimates that OpenAI, Anthropic, and xAI combined held somewhere between 20 and 30 percent of global operational AI compute at the end of 2025. Add in the portions held by Google DeepMind and Meta Superintelligence Labs, and the five best-resourced frontier developers still held less than half. The installed global base sat at roughly 16 million Nvidia H100-equivalent units. The top two labs accounted for perhaps 3 to 4 million of those.
That gap represents real opportunity, and real risk. The bull case says Anthropic and OpenAI are growing compute far faster than the broader industry. OpenAI tripled data center power capacity in both 2024 and 2025, and You estimates its computing power grew roughly fourfold per year after accounting for hardware efficiency gains. Anthropic’s pace may be faster still, given the funding its revenue growth has unlocked, which lets it secure more compute than almost anyone else. Both labs have signed multi-gigawatt agreements in 2026, including Anthropic’s arrangements with Amazon, Google, CoreWeave, and an agreement to rent SpaceX data center capacity.
The math compounds quickly. You models a scenario where Anthropic and OpenAI start at 20 percent of global compute and grow 33 percent faster than the rest of the industry each year. Under those assumptions they would hold roughly 80 percent of global AI compute within five years. At that point, continued capability scaling would require the overall compute buildout to keep accelerating, not just the labs’ share of an existing base.
That is where the bear case enters. Global AI capital expenditure is approaching $1 trillion per year, per the Epoch AI analysis. Hyperscalers are growing their capex at roughly 70 percent per year and guiding similar rates for 2026, but sustaining that trajectory well beyond this year would require economic conditions that do not currently exist. You puts it plainly: once the frontier labs have absorbed most of the available compute, further growth in model capabilities at the current rate would require a dramatic economic transformation, meaning AI itself driving enough productivity growth to justify a multi-trillion-dollar annual compute budget.
Neither conclusion is inevitable, and Epoch AI is careful to frame these as scenarios, not forecasts. Two skeptical notes deserve weight. First, the concentration scenario assumes the current revenue trajectory persists, and the recent growth rates are extraordinary enough that extrapolation is suspect. Second, the economic-transformation threshold is analytically fuzzy. It depends on how much of the compute buildout non-frontier uses absorb, including open-weight inference, recommendation systems, and biological modeling, which the analysis does not fully resolve.
What is concrete, and what matters for near-term planning, is that even flat capex does not mean flat capability. Chips keep improving in price and energy efficiency. A fixed annual budget buys more H100-equivalent compute every year. A plateau in capital spending would slow the growth rate, but the installed compute base would keep expanding, and training runs and research would continue. The competitive dynamics shift when capex plateaus, because labs with larger existing fleets keep a structural advantage over challengers building from scratch.
For operators making infrastructure bets or model-roadmap decisions for 2027, the Epoch AI framing implies a specific risk. Planning around a continuation of the 2023 to 2025 capability curve assumes capex acceleration continues beyond 2026, and that assumption carries more uncertainty now than it did eighteen months ago. Teams that need a specific capability level to ship a product should pressure-test whether that capability arrives on the current curve or only on an accelerated one.
Analysis by Josh You, published by Epoch AI on 2026-05-21.
