About $1B of training compute is shared for free on the Hugging Face Hub
A rough estimate of what it would cost to rent the GPUs that trained the open-weight models on the Hub. It covers the final training runs of the main model families, not the long tail.
Cumulative cost over time
Each family's cost is placed at its first or latest release month. Shaded: 10th to 90th percentile. "Split by evidence" shows how much rests on published numbers.
By organization
Bar: central estimate. Line: 10th to 90th percentile.
Evidence behind the estimate
Share of the central estimate by type of evidence.
All model families
Costs in $M. Click a column to sort.
| Organization | Family | Released | Evidence | Compute basis | Cost ($M) | Sources |
|---|
Method
- Counted: the final pre-training and post-training runs of 75 official model families.
- Not counted: community fine-tunes, LoRAs and merges (about 900k Hub repos declare a parent model), smaller families outside the list, experiments and failed runs, salaries, datasets. So this is a floor on what was spent, not the full cost.
- Prices: $2.5 per H100-hour, $2 per A100 or H800-hour, $5 per B200-hour. When only FLOPs are known, $1.8M per 10²⁴ FLOP (H100 at $2.5/hour, 40% utilization).
- Compute: GPU-hours from the model card when published. Otherwise 6 × active parameters × training tokens. When tokens aren't published they are guessed, and the family is marked "partly assumed".
- Overlap: a base model is counted once. Official variants only add their extra training. Quantizations and mirrors count as zero.
- Range (): 20,000 simulated runs varying the GPU price ($2 to $3 per H100-hour), utilization (35% to 50%) and each family's training quantities between its low and high estimates. Families with guessed token counts move together. Setting every assumption to its extreme at once gives .
- Dates: some release months come from Epoch AI publication dates.
Data
Families (CSV) Organizations (CSV) Monthly cumulative (CSV) Monthly realistic range (CSV) Release dates and sources (CSV) Realistic ranges per family (JSON) Realistic range summary (JSON) Simulation script Summary (JSON) Previous detailed audit (French)