Numbers age fast in AI. Epoch AI’s May 2026 update makes that concrete.
According to Epoch AI’s latest compute data, median frontier training compute grew approximately 44 times per year between 2023 and 2026. The baseline from 2010 to 2022 was approximately 3.8 times per year. That’s not incremental acceleration. It’s a structural break in the growth curve, and it’s happening while the regulatory frameworks designed to contain frontier AI risk are still being implemented.
The number that will get the most immediate attention from compliance teams: Epoch AI’s May 2026 update indicates more than 30 models now exceed the 10^25 FLOP training compute threshold that triggers mandatory systemic risk reporting under the EU AI Act. The Act’s compliance timeline is fixed, with high-risk system registration deadlines approaching. As recently as April 20, Epoch AI’s own data showed 12 models above the threshold.
That’s more than 18 newly threshold-crossing models in roughly 12 days. The increase likely reflects a combination of rapid new model releases and possible refinements to Epoch AI’s counting methodology. Epoch AI has not specified the specific drivers, and this discrepancy should be treated as a data point requiring confirmation rather than a settled count. But even accounting for methodology variation, the directional signal is consistent: the number of models triggering mandatory regulatory obligations is growing faster than compliance infrastructure can absorb.
The efficiency side of the data adds a complication. Epoch AI reports compute efficiency – measured as petaFLOP per dollar, improving at approximately 37% per year. Lower cost per FLOP doesn’t slow the growth in total compute; it accelerates the model count crossing regulatory thresholds, because falling costs let more organizations train at frontier scale. More efficient compute means more models, not fewer.
On infrastructure concentration, Epoch AI’s data indicates the largest known AI data center – the Anthropic-Amazon joint infrastructure, exceeds 1.1 gigawatts of capacity. For context, a standard utility-scale power plant runs between 0.5 and 1 GW. A single AI facility now exceeds that range. Epoch AI references additional large-scale compute clusters in its tracking data; specific facility names and equivalent GPU counts in this package could not be independently verified and are excluded here.
Practically speaking, the 30+ model count matters most for two audiences. EU AI Act compliance teams working from the April 20 figure of 12 models are now operating on outdated data, the scope of mandatory obligations for general-purpose AI model providers has expanded materially in the same period their compliance programs were being built. Infrastructure investors and hyperscaler strategists watching the 44x growth figure should treat it as a signal that the infrastructure economics underpinning current capex commitments were calibrated to a baseline that no longer applies.
One practical note the compute data doesn’t address directly: the 37% annual efficiency gain and the 44x compute growth rate are running in opposite directions relative to what they imply for compliance workload. More efficient compute lowers the technical barrier to crossing the regulatory threshold; faster overall growth increases the number of organizations that find themselves newly subject to mandatory EU AI Act obligations. Compliance teams can’t offset growing scope with efficiency arguments.
What to watch: Epoch AI’s methodology note on the threshold count increase (12 to 30+) is the most important clarification to track. The specific explanation, new model releases, methodology change, or both, determines how compliance teams should respond and whether the April 20 figure from prior briefs needs to be treated as superseded or supplemented.