Category
Compute
Training runs, chips, data centers, and the cost of a token.
10 numbers
Compute · Global
5×
per year growth in training compute for frontier language models since 2020
Epoch AI · Feb 5, 2026
Compute · Global
$0.07
per million tokens was the cost to query a GPT-3.5-level model by October 2024
Stanford HAI · Apr 7, 2025
Compute · Global
280fold
reduction in the cost of querying a GPT-3.5-level model in about 18 months
Stanford HAI · Apr 7, 2025
Compute · Global
$38billion
in up-front capital expenditure for a typical AI data center with 1 gigawatt of IT power
Epoch AI · Feb 5, 2026
Compute · Global
49%
per year growth in AI chip performance per dollar since 2023
Epoch AI · Feb 5, 2026
Compute · Global
30%
of OpenAI's compute spending in 2024 was used on inference
Epoch AI · Feb 5, 2026
Compute · Global
3.5×
per year increase in the cost to train frontier language models since 2020
Epoch AI · Feb 5, 2026
Compute · Global
1.1million
NVIDIA H100-equivalent chips of computing capacity in the largest known AI data center
Epoch AI · Feb 5, 2026
Compute · Global
1.4x
per year growth in how long frontier LLM systems spend training, since 2020
Epoch AI · Jul 25, 2025
Compute · Global
9months
is the training length past which frontier runs may be inefficient
Epoch AI · Jul 25, 2025