AIAcademy · AIAcademy · 2026-05-16
Epoch AI — frontier training compute trends
Ten to the twenty-sixth floating-point operations. That number sits inside California SB 53, the EU AI Act's Article 3 definition of GPAI with systemic risk, and the (now-revoked) Biden Executive Order 14110. It is the load-bearing scalar of frontier-AI regulation. It deserves a clear translation.
What it looks like in chips and time. An Nvidia H100 delivers roughly 1,000 TFLOPS sustained on BF16 — call it 10¹⁵ operations per second. Run 100,000 H100s at full utilisation, you reach 10²⁰ FLOPs per second. To accumulate 10²⁶ FLOPs at that rate takes 10⁶ seconds — about 12 days. A more realistic 30–50% sustained utilisation across a 100K-GPU cluster pushes that closer to a month of training. A 25K-GPU cluster at the same utilisation needs roughly four months. That is the rough envelope GPT-5, Gemini Ultra 2.5, and Claude Opus 4.x sit in.
Why this specific exponent. The threshold was not chosen for theoretical reasons. It was set just above the largest disclosed 2024 training runs and well above the largest disclosed 2023 runs, so that roughly 5–8 firms would be in scope at the law's effective date. Epoch AI's frontier-training tracker plots actual training compute over time; 10²⁶ is currently the upper envelope, not the median. The choice is a regulatory targeting decision dressed as a physical constant.
Why it will not hold. Compute efficiency improves. Epoch's training-compute analysis shows frontier runs growing at ~4.5× per year, while effective compute — what you get per nominal FLOP, after algorithmic and architectural gains — improves another 3× annually. The implication is straightforward: a 10²⁶ FLOP-equivalent model in 2028 will be trainable with roughly 10²⁵ nominal FLOPs, and by 2030 the threshold will be inside the budget of mid-tier industrial labs rather than only the biggest five.