Epoch AI's 11 June data insight estimates that the record for computing capacity inside one AI data centre has grown 3.3-fold per year since August 2024. That log-linear fit implies a doubling about every seven months. It is a sharp measure of construction ambition, but it is an estimated frontier, not a promise about every site or the pace of model progress.
Epoch AI Watch analysis
What happened
Epoch follows whichever individual facility holds the compute record at a given point. It measures chips using their maximum specified throughput at 8-bit precision, then converts the result into H100 equivalents. Most estimates combine a site's peak power capacity with the expected energy efficiency of its likely chips; the power figures draw on satellite imagery, public disclosures and permitting documents. A few facilities have disclosed chip types and counts directly.
The downloadable table retrieved on 31 August contains nine rows labelled as the observed frontier and five as the future frontier. Its final observed point is Meta Prometheus on 28 May 2026 at 763,012 H100 equivalents. There is a small record-keeping splinter: Epoch's method text says the analysis used 10 past observations and five future projections. The mismatch is worth preserving rather than sanding smooth.
Why it matters
Epoch tested the noisy capacity and opening-date estimates with 1,000 simulations. The annual growth fits ran from 2.3-fold at the fifth percentile to 5.1-fold at the 95th, against the 3.3-fold point estimate. The authors call that an illustrative range, not a formal confidence interval. For frontier sites, they expect the true capacity to sit within a factor of 1.3 and the operational date within three months about 80 per cent of the time.
The metric shows how much infrastructure leading firms appear willing and able to concentrate in one facility. It does not identify the largest training cluster. One site may contain several clusters, while a single cluster could eventually span multiple campuses. That distinction matters when turning an impressive line on a chart into claims about models or competitive position.
Our read
Treat the trend as a pressure gauge for the physical AI race. Chips, electricity, cooling, networking, capital and planning all have to arrive together, which is the sort of group project that usually discovers one participant has brought only biscuits. The seven-month number is useful precisely because it raises the harder question: which constraint stops the curve behaving neatly?
What to watch
- Whether projected frontier sites become operational on the recorded dates and at the estimated capacity.
- How revisions to chip mix, power capacity or opening dates alter the fitted growth rate.
- Whether multi-campus training makes single-site compute a less representative frontier measure.
Discussion spark: If single-site compute keeps growing this quickly, which constraint bites first: power, chips, networking, capital, planning permission or public patience?
Sources and evidence
Independent WittyWires tracker for public updates about Epoch AI. Not affiliated with or endorsed by Epoch AI; this is not an official account.