Epoch AI: frontier training compute has doubled every 5.2 months since 2020

Five point two months.

That is how long it takes for the compute used to train a frontier AI model to double, and the rate has held since 2020, per Epoch AI, the research organization whose entire job is measuring this. A doubling every 5.2 months compounds to roughly a fivefold increase every year. Not five percent growth. Not fifty. Fivefold, annually, sustained across six years and counting.

This post makes one argument, and only one. That number, and its quieter siblings below, describe a demand curve that no physical construction process on earth can track, and the gap between the two is the single plainest explanation for why data-center site selection became so urgent and so contested at the same time.

Three clocks, all fast

Epoch tracks several distinct series, and it pays to keep them separate, because they answer different questions.

The 5.2-month figure is training compute for frontier models, the flagship systems at the edge of capability. It measures what the most ambitious single projects consume, and it is the fastest of the three clocks, a roughly 5x annual growth rate.

The second clock is the installed base. In data published February 5, 2026, Epoch put global AI computing capacity on a doubling time of roughly every 7 months, with a 90 percent confidence interval of 6 to 8 months, which works out to a 3.3x annual growth rate sustained since 2022. Epoch measures this in H100-equivalent units, normalizing the whole zoo of accelerator generations into one yardstick so the series tracks actual compute rather than box counts. More than 60 percent of that total capacity is NVIDIA hardware, which means the world's AI compute stock is, to a first approximation, the output cadence of a single company's supply chain.

The third clock is inference, the compute spent running models rather than training them, which Epoch's own analysis puts at roughly a tripling every year. Training builds the model once; inference is the meter that runs every time anyone uses it, which is why its curve tracks adoption rather than ambition, and why it is the series a landlord should watch most closely. And in the essay that matters most for anyone planning past next quarter, "Can AI scaling continue through 2030?", Epoch's answer is that these trends are on track to continue through 2030.

Take the middle, conservative clock and run it forward. At 3.3x annually, two years of compounding hands you roughly an eleven-fold installed base. That is arithmetic, not forecasting, and every unit of that multiple has to land somewhere physical.

Nothing made of concrete moves at this speed

Now put the other calendar next to it, the one this research lane lives on.

An interconnection request in a constrained territory sits in a queue measured in years. A substation upgrade is measured in years. A transmission line is measured in many years and at least one lawsuit. A contested rezoning in a jurisdiction that has already seen its first data-center fight can consume an entire election cycle, and the moratorium that sometimes follows can consume the next one. None of this is a complaint about permitting; some of it is democracy doing its job at the deliberate pace democracy runs at. But the pace is the pace.

Demand doubles in seven months. Supply is a public hearing.

When a demand curve compounds monthly and its supply curve moves in years, the difference between them does not politely wait. It gets absorbed by the only two things that can absorb it, price and conflict. Price shows up as scarcity premiums on the small stock of land that is already entitled, already powered, and already near fiber, the parcels where the multi-year queues have been pre-run. Conflict shows up as everything RealClear's other lane documents weekly, the packed hearing rooms and the moratorium votes. A 7-month doubling time meeting a multi-year entitlement calendar produces both, mechanically, the way pressure produces heat.

There is a practical corollary for anyone reading capacity announcements as demand signals. A press release describes a decision made one or two doublings ago, under an estimate of need the curve has already passed. The interconnection filings and land assemblages that surface in county records run closer to the live edge of the exponent than anything a communications office publishes.

I think this framing also explains something that puzzles people outside the industry, which is why site selection feels frantic when so much capacity is visibly under construction. The answer is in the exponent. Against a 3.3x annual rate, whatever is under construction today was sized against a demand estimate that is one or two doublings stale by the time it energizes. The buildout is not behind because anyone is slow. It is behind because the target moves faster than concrete cures.

The demand side of this market is not a real-estate cycle. It is a technology curve wearing a real-estate cycle's clothes, and the industry keeps underwriting it with tools built for the former.

My bet, in checkable form. Epoch's installed-capacity doubling time stays inside its published 6-to-8-month confidence band through the end of 2027, and when the curve finally does bend, the binding constraint will be interconnection and permission rather than capital or silicon, because those are the inputs no amount of spending compresses. If Epoch's published data shows the doubling time drifting past eight months before 2028, I was wrong, and cheerfully so, because that world gives every county and every developer breathing room that nobody's current model assumes.

This analysis is a source-cited research summary drawn from public reporting and market research, not investment advice. It can contain errors and should be verified independently before any investment decision.