Trends in Frontier AI Model Count: A Forecast to 2028

TL;DR

This paper analyzes compute-threshold governance for frontier AI and forecasts how many released models will exceed fixed training compute thresholds ($>10^{25}$ FLOP and $>10^{26}$ FLOP) and frontier-connected thresholds through 2028. It combines forecasts of total AI compute stock growth with allocations across model sizes, anchored by LMS behavior and an allocation gradient, using Epoch AI's Notable Models dataset to fit size distributions. The results project a median of $165$ models above $10^{25}$ FLOP and $81$ above $10^{26}$ FLOP by 2028 (with wide intervals), while frontier-connected thresholds imply near-stable annual captures (roughly 14–16 within 1 OOM, 6–8 within 0.5 OOM, and 20–24 within 1.5 OOM). The work emphasizes governance implications, noting the rapid growth in frontier AI model counts and the need for flexible policy design, all tempered by dataset selection effects and parameter uncertainties.

Abstract

Governments are starting to impose requirements on AI models based on how much compute was used to train them. For example, the EU AI Act imposes requirements on providers of general-purpose AI with systemic risk, which includes systems trained using greater than $10^{25}$ floating point operations (FLOP). In the United States' AI Diffusion Framework, a training compute threshold of $10^{26}$ FLOP is used to identify "controlled models" which face a number of requirements. We explore how many models such training compute thresholds will capture over time. We estimate that by the end of 2028, there will be between 103-306 foundation models exceeding the $10^{25}$ FLOP threshold put forward in the EU AI Act (90% CI), and 45-148 models exceeding the $10^{26}$ FLOP threshold that defines controlled models in the AI Diffusion Framework (90% CI). We also find that the number of models exceeding these absolute compute thresholds each year will increase superlinearly -- that is, each successive year will see more new models captured within the threshold than the year before. Thresholds that are defined with respect to the largest training run to date (for example, such that all models within one order of magnitude of the largest training run to date are captured by the threshold) see a more stable trend, with a median forecast of 14-16 models being captured by this definition annually from 2025-2028.

Figures (11)

• Figure 1: Historical data (2017-2023; blue) and a sample of our model’s predictions (2024-2028; red) for the number of AI models exceeding $10^{25}$ and $10^{26}$ FLOP.
• Figure 2: Compute allocation across model sizes for years 2020-2023, x-axis unnormalized.
• Figure 3: Compute allocation among model sizes for years 2020-2023, x-axis normalized by largest model trained that year. Linear fit shown.
• Figure 4: Historical distribution of models over training compute - scatter plot.
• Figure 5: Historical distribution of models over training compute - KDEs