Public AI models keep climbing the compute curve.
Epoch AI's public database tracks thousands of notable AI systems, including publication dates, organizations, model domains, parameter counts, training-compute estimates, and other metadata. For this chart, the focus is the subset of models with known estimated training compute.
The long-run picture is still striking. Public estimates for frontier models have moved into the 10^25 to 10^26+ FLOP range, with labeled systems such as GPT-4, Gemini 1.0 Ultra, Llama 3.1 405B, GPT-4.5, Grok 4, and Composer 2.5 near the top of the public compute curve.
The important caveat is 2026. The dashed 2026 segment should not be read as proof that frontier training compute declined. It reflects what is visible in Epoch's public estimates so far. Many current frontier labs do not disclose training compute, and public estimates often lag the newest commercial releases.
For investors, the chart still matters because it shows the scale of the infrastructure problem. Frontier AI progress remains tied to enormous training runs, and those runs connect model competition to chips, memory, networking, power, cooling, and data-center capacity.