Hugging Face founder: $1B of training compute shared free on the Hub, up 25x since 2023
Thom_Wolf · x · 2026-10-18
Thom Wolf, co-founder of Hugging Face, launched "Hub Dollars," an interactive estimate of the training compute behind open-weight models on the Hub.
Key numbers
- $0.95B central estimate (up to $1.28B) across 72 model families from 35+ orgs including Meta, NVIDIA, Alibaba, DeepSeek, Moonshot and Mistral — versus under $40M at the end of 2023.
- Pricing basis: $2.5/H100-hour, $2/A100-H800-hour, $5/B200-hour; or $1.8M per 10²⁴ FLOP when only FLOPs are known.
Method & caveats
- GPU-hours from model cards when published, otherwise 6 × active params × tokens; 20,000 simulations yield the 10th–90th percentile range.
- It's a floor: excludes 900k community fine-tunes, failed runs, datasets and salaries.
- The curve steepens sharply from 2025 with DeepSeek, Kimi and Qwen releases.
"Wikipedia was written in volunteer hours. This AI-age commons is written in GPU-hours and data," Wolf writes. An accompanying HF Space lets you explore costs by organization and model family.
More from Infra
- AMD unveils 'Helion' racks for agentic apps with up to 32,768 EPYC cores or 16 MI355X GPUs — Beth_Kindig · 2026-10-18
- BofA: AI GPUs run near 100% utilization, rental rates up 12%-37% YTD — not dotcom glut — McDonaghMatthew · 2026-10-18
- Reddit take: data centers are becoming the new oil drills — Unique_Ad_330 · 2026-10-18
- Robotics to demand 363 TWh of electricity globally by 2035, per Wood Mackenzie — ingliguori · 2026-10-18
- Leap second during backup silently skips it: Antithesis clock bug detailed — moarbugs · 2026-10-18
- Musk's Terafab vision: AI chip race shifts to full vertical integration for speed — tekbog · 2026-10-18