The New York Times has published a visualization based on Epoch AI research, illustrating the computational requirements for training modern AI models. The analysis utilizes Epoch AI's metrics for training compute, measured in floating-point operations (FLOP), and capability scores.
Data indicates that training compute for frontier language models has increased by approximately 5x annually since 2020, effectively doubling every 5.2 months. Since 2010, notable AI models have experienced an annual growth in training compute ranging from 4.4x to 4.7x.
Epoch AI identifies hardware quantity as a primary growth driver, with a median increase of roughly 1.69x per year. As of mid-2026, there are approximately 20 million AI chips in use globally, calculated in NVIDIA H100 equivalents. Projections indicate this number could reach 200 million by the end of 2028, doubling every nine months.
In September 2026, Epoch AI launched an AI chip users explorer. Its database tracks 86 sites that represent about 44% of global AI compute capacity. Additionally, an October 2, 2026, report suggests that high-bandwidth memory (HBM) shipped between 2025 and 2027 could support between 30 million and 170 million concurrent frontier-model AI agents, with potential scaling to 1.9 billion agents for cheaper models.
