The EU AI Act sets 10^25 FLOPs of training compute as the threshold at which a model triggers systemic risk obligations, the Act’s most stringent requirement tier. When that threshold was written, the number of models above it was small. According to Epoch AI’s May 7, 2026, compute tracking update, as reported by The Wire, that count has grown substantially and now reportedly exceeds 30 models. The specific trajectory, how quickly that number moved, cannot be confirmed from the source material available, but the directional claim is consistent with the 44x annual compute growth rate that Epoch AI has documented for the 2023 to 2026 period, a figure corroborated across multiple prior technology-pillar entries in this hub’s coverage.
For context on what that growth rate means: historical compute scaling prior to this period ran at a fraction of the current pace, according to Epoch AI’s sourcing. The specific prior baseline figure has not been independently confirmed and isn’t stated here. What the data does establish is that the current pace is substantially above historical norms, and that the EU AI Act’s threshold, designed to capture a narrow category of frontier models, is capturing a widening one.
Epoch AI also tracks benchmark performance across frontier models. Per The Wire’s sourcing from Epoch AI’s May 7 data, MMLU-Pro scores for top-tier frontier models have reportedly clustered in a narrow range in the high-80s to 90 percent, suggesting that standard academic benchmarks may be reaching saturation as a differentiation tool. Claude Opus 4.7, confirmed in prior coverage, is among the models in this range. Other specific model names cited in the Wire’s sourcing have not been independently confirmed and are not stated here.
Benchmark saturation matters for regulators as much as for developers. If standard evaluation tools can no longer reliably differentiate frontier capability, then conformity assessments that rely on benchmark performance face a methodological gap. The AI Office, gaining full enforcement powers on August 2, will need to address what adequate technical documentation looks like when the benchmarks the documentation references are themselves no longer discriminating.
The regulatory implication, that compliance frameworks designed around a threshold population are being outpaced by a rapidly expanding scope, is a conclusion that AI governance observers are drawing from this data. It’s worth being precise: that interpretation is analysis, not an Epoch AI finding. What Epoch AI documents is compute growth. The policy consequence of that growth is a separate judgment.
The combination of today’s two EU stories is worth reading together. The Omnibus deal extends the high-risk compliance deadline for existing in-scope systems. The Epoch AI data shows the number of systems entering scope is growing faster than the frameworks designed to govern them were built to anticipate. The extension buys time. It doesn’t resolve the widening scope, and compliance teams whose systems are approaching the 10^25 FLOP threshold should be asking, this week, whether they’ve crossed it yet.