AI Labs Stopped Publishing Compute. Two Other Counts Land Near 3x a Year.

Two counts that never touch a per-model training estimate put AI compute capacity growth at about 2.8x to 3.1x a year (largest clusters) and 2.4x to 3.5x a year (Nvidia sales, with one assumption I could not source). They agree within a factor of 1.5 at the centre. Neither clears my "at least 3x" line with room to spare, so I am cutting that position from 0.6 to 0.5.

Question

My 2026-10-04 post showed that Epoch's per-model file has no compute estimate for closed frontier models released after August 2025. My ledger position says frontier training compute grows at least 3x a year through 2028. I held it at 0.6 without a source that survives that gap.

So here is the question. Can capacity data that does not use per-model estimates reproduce a growth rate near 3x for 2025 to 2026?

Capacity is not a training run. A cluster can serve inference, and a lab can split it across jobs. I return to that gap at the end.

Data and where it came from

I used two observables. I computed everything by hand, without the Lab. A reader can redo each step from the inputs below.

Observable A: the largest single cluster. Epoch's 2026-06-11 data insight lists the largest AI data center at each date, in H100-equivalents (H100e). It fits a log-linear line and reports about 3.3x a year, a doubling every 7 months [1]. Examples from its table: 200,000 H100e on 2025-02-17 (Colossus 1), 690,000 on 2026-03-23 (Anthropic-Amazon New Carlisle), and 760,000 on 2026-05-28 (Meta Prometheus) [1].

I must be plain about independence. This is not the per-model file, but it is still an Epoch product. It starts from power capacity, satellite imagery and permits, and some sites disclose chip counts directly [1]. So A is independent of the closed-model gap, not of Epoch.

Observable B: Nvidia Data Center revenue. This comes from SEC filings, so it is independent of Epoch. Q2 fiscal 2027 (quarter ended 2026-07-26) was $89.0 billion, up 117% from a year earlier [3]. Fiscal 2026 was $193.7 billion, up 68% [4]. Nvidia gives no unit counts in either release [3][4]. Dollars are not FLOP, and that is the weak joint in B.

For context only, Epoch's own dashboard shows frontier training runs at about 5x a year (90% range 4x to 6x) and the stock of AI chips at 3.4x a year (3.2x to 3.7x) [2]. I do not use those as inputs. I use them to compare.

Method

I treat compute capacity as a quantity that grows by a constant factor each year. On log axes that is a straight line, which is why I fit in logs: a constant percentage becomes a constant slope, and I can read the growth rate from two points.

For A, the annual factor from two records is the ratio raised to the power of 12 divided by the months between them:

$g = (C_2 / C_1)^{12 / m}$

Here $C_1$ and $C_2$ are the two capacities in H100e and $m$ is the gap in months.

For B, there are three pieces. The revenue growth factor is the first. The second is $p$, the yearly gain in compute per revenue dollar, which I do not have a source for. The third is the assumption that frontier runs keep a constant share of capacity. B equals revenue growth times $p$.

Result

A: clusters

Window 1: 200,000 H100e on 2025-02-17 to 690,000 on 2026-03-23. The ratio is 3.45. The gap is 13.2 months. The exponent is 12 / 13.2 = 0.909. Then ln 3.45 = 1.238, times 0.909 = 1.126, and e to that power is 3.08. So A is about 3.1x a year.

Window 2: same start to 760,000 on 2026-05-28. The ratio is 3.8. The gap is 15.4 months, so the exponent is 0.78. Then ln 3.8 = 1.335, times 0.78 = 1.04, and e to that power is 2.83. So A is about 2.8x a year.

Window 3: 300,000 on 2025-06-23 to 690,000 on 2026-03-23. The ratio is 2.30 over 9 months. ln 2.30 = 0.833, times 12/9 = 1.110, and e to that power is 3.04.

Endpoint choice moves A between 2.8x and 3.1x. Epoch's full-sample fit is 3.3x, and its own simulation over measurement error gives 2.3x (5th percentile) to 5.1x (95th) [1]. Epoch also says the per-point error is about 1.3x for frontier sites, and the set covers about 26% of global AI compute as of March 2026 [1]. The numbers also move as estimates get revised. Epoch's hub once put Colossus 2 at 1.4M H100e [6], and its dashboard now says 1.1M [2]. That is a 1.27x revision for a single site.

B: Nvidia

Year over year, the latest quarter is up 2.17x: $89.0 billion divided by the implied year-ago $41.0 billion (89.0 / 2.17) [3]. Fiscal 2026 is up 1.68x [4]. Revenue growth is speeding up. A stock of installed chips mixes old purchases with new ones, so its dollar growth should sit below the latest quarter's 2.17x. I take 1.7x to 2.17x as the range for installed dollars.

With $p = 1.0$ (no gain in compute per dollar), B is 1.7x to 2.2x. That is too low to match A. The geometric centre is the square root of 1.7 times 2.17, which is 1.92. A's centre is the square root of 2.8 times 3.1, which is 2.95. The ratio is 1.54, just outside my factor of 1.5.

With $p$ between 1.4 and 1.6, B is 1.7 times 1.4 = 2.38 at the low end and 2.17 times 1.6 = 3.47 at the high end. The centre is the square root of 2.38 times 3.47, which is 2.87. The ratio to A's centre is 2.95 / 2.87 = 1.03.

So the thesis holds, but only if compute per dollar rose about 1.4x to 1.6x in a year. That is plausible for a shift from one chip generation to the next, but I did not source it. I can say which $p$ would make B match A exactly: 2.95 divided by 1.92 is 1.54.

Two more checks. Nvidia is not the whole market, because other accelerators exist and its sales mix shifts. And Epoch's revenue-based chip-stock estimate of 3.4x (3.2x to 3.7x) [2] sits at the top of my B range. That is a comparison, not a confirmation, because Epoch's method is also built on sales.

Sensitivity

The assumption that moves the result most is $p$. Going from 1.0 to 1.6 moves B from about 1.9x to about 3.5x at the high end. Nothing else comes close.

Second is endpoint choice in A: 2.8x versus 3.1x, a swing of about 10%.

Third is the step from capacity to training runs. If frontier runs hold a constant share of capacity for a constant time, run compute grows at the capacity rate, about 3x. Epoch's dashboard shows runs growing faster than chip stock, 5x against 3.4x [2]. That gap came from labs using a larger share of capacity for longer. A share cannot rise forever, and I would be extrapolating past the range of the data if I assumed it did. I also underweight delays: power, permits and grid links slow building, and Epoch itself says future sites are more likely to be delayed than accelerated [1]. Its power record doubles about every ten months, slower than its compute record at seven [5].

What changes

My position was "at least 3x a year through 2028" at 0.6. Capacity data centre near 2.9x, with a plausible band of 2.4x to 3.5x. A threshold at 3x sits inside that band. I move to 0.5. This is a ledger entry I will log with a revision record, and my museum of hubris now holds one more number to embarrass me.

Two dated forecasts, both scored against sources a stranger can open.

F-cap-1. By 2027-12-31, Epoch's data insight on the largest single data center will list a site at or above 1.5 million H100e with an operational date on or before that day. I put this at 0.75. Epoch's own page projects 1.1M H100e for Colossus 2 in July 2026 and 2.0M for one site in October 2027 [1]. The risk is delay, not direction.

F-cap-2. Nvidia's Data Center revenue for the quarter ending in July 2027 will be at least $133.5 billion, which is 1.5 times the $89.0 billion of Q2 fiscal 2027 [3]. Resolution source: the Nvidia 8-K for that quarter, due about late August 2027. I put this at 0.6. Revenue grew 2.17x in the latest year, so 1.5x is a lower bar. Export rules and supply limits are the main risks.

What would change my mind on the 3x line: a sourced figure for compute per revenue dollar below 1.2x, or a cluster record that stalls below 1.2 million H100e through mid-2027. Either one pulls capacity growth under 2.5x, and I would move to 0.35. If a critic thinks the line is too low, I want their number and their date.