What it costs to train a 125M model from scratch
A compute-optimal run for a 125M-parameter model, 2.5B tokens at 20 per parameter, takes about 1.3 H100-hours, around $5.78 at $4.39 an hour. Training on 10B tokens (80 tokens per parameter, typical of GPT-2-small reproductions) raises that to about $23.12. A weekend project: one rented GPU trains it in hours for a few dollars. It is a teaching model, not a useful assistant.
Compute-optimal: 2.5B tokens
| GPU | GPU-hours | Cost | Time on 8 | Time on 64 | Time on 512 | GPUs to hold state |
|---|---|---|---|---|---|---|
| RTX 5090209.5 TFLOPS · $1.39/hr | 6.2 | $8.64 | 47 minutes | 6 minutes | 1 minutes | 1 |
| A100 80GB312 TFLOPS · $2.19/hr | 4.2 | $9.14 | 31 minutes | 4 minutes | 1 minutes | 1 |
| H100 SXM989 TFLOPS · $4.39/hr | 1.3 | $5.78 | 10 minutes | 1 minutes | 1 minutes | 1 |
| H200989 TFLOPS · $5.29/hr | 1.3 | $6.96 | 10 minutes | 1 minutes | 1 minutes | 1 |
| B2002,250 TFLOPS · $7.99/hr | 0.58 | $4.62 | 4 minutes | 1 minutes | 1 minutes | 1 |
Longer training: 10B tokens
| GPU | GPU-hours | Cost | Time on 8 | Time on 64 | Time on 512 | GPUs to hold state |
|---|---|---|---|---|---|---|
| RTX 5090209.5 TFLOPS · $1.39/hr | 25 | $34.56 | 3.1 hours | 23 minutes | 3 minutes | 1 |
| A100 80GB312 TFLOPS · $2.19/hr | 17 | $36.56 | 2.1 hours | 16 minutes | 2 minutes | 1 |
| H100 SXM989 TFLOPS · $4.39/hr | 5.3 | $23.12 | 39 minutes | 5 minutes | 1 minutes | 1 |
| H200989 TFLOPS · $5.29/hr | 5.3 | $27.86 | 39 minutes | 5 minutes | 1 minutes | 1 |
| B2002,250 TFLOPS · $7.99/hr | 2.3 | $18.50 | 17 minutes | 2 minutes | 1 minutes | 1 |
Compute = 6 × 125e6 parameters × tokens, at 40% utilisation of dense BF16 peak, with perfect scaling across GPUs. Memory for weights, gradients and Adam state is 2.0 GB before activations.
What real runs of this size reported
| Run | Size and data | Reported |
|---|---|---|
| nanochat | GPT-2 capability, 8 H100s, about 2 hours | about $48 |
Published GPU-hours for Llama 3.1 came in 1.2–2.9× above the 6ND estimate; budget for that gap, plus experiments, failed runs and data work, before you treat these numbers as a quote.
Change the assumptions
Pre-filled for 125M parameters at 2.5B tokens on H100s.
$6
1
10 minutes
1.9e18 FLOPs
Full training needs about 2 GB for weights, gradients and optimiser state before activations, so at least 1 H100 SXM just to hold the model. Fine-tuning with QLoRA instead needs about 0 GB for 4-bit weights, which fits on a single 32 GB card.
Frequently asked questions
- How much does it cost to train a 125M model?
- About $5.78 on H100s at $4.39/hour for a compute-optimal run of 2.5B tokens (6 × parameters × tokens at 40% utilisation). At 10B tokens (80 tokens per parameter, typical of GPT-2-small reproductions) it is about $23.12. Treat both as floors: published runs used 1.2–2.9× more GPU-hours than the formula.
- How long does it take to train a 125M model?
- About 10 minutes on one 8-GPU H100 node for 2.5B tokens, assuming perfect scaling, or 1 minutes on 64 GPUs.
- How much GPU memory does training a 125M model need?
- About 2.0 GB for weights, gradients and Adam state (16 bytes per parameter) before activations, so at least 1 × 80 GB GPU with the state sharded.
- Which GPU is cheapest for training a 125M model?
- At on-demand November 2026 rates, the B200 ($7.99/hour) gives the lowest compute cost, about $4.62, by our estimate. Peak FLOPS aren't delivered performance, so benchmark your own code before a large run.
Fine-tuning instead
Open models of this size already carry far more pretraining than most budgets allow. Adapting one costs a tiny fraction:
- Fine-tune Llama 3.2 3B: cost$0.27 per 10M tokens
- Fine-tune Mistral 7B: cost$0.61 per 10M tokens
- Fine-tune Qwen2.5 7B: cost$0.64 per 10M tokens
- Train a 1B modelFrom scratch
- Train a 3B modelFrom scratch
- Train a 7B modelFrom scratch
- Train a 13B modelFrom scratch
- Train a 70B modelFrom scratch
The method in depth is in the free Academy course The Economics of Model Training and the article What it really costs to train an LLM; terms are defined in the glossary under 6ND, Chinchilla scaling and MFU.
Sources
- Kaplan et al., 2020. Scaling Laws for Neural Language Models (arXiv:2001.08361)
- Hoffmann et al., 2022. Training Compute-Optimal Large Language Models (arXiv:2203.15556)
- Meta, 2024. The Llama 3 Herd of Models (arXiv:2407.21783)
- Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
- Runpod. GPU pricing (updated 12 November 2026)
- Karpathy. nanochat
You probably don't need to pretrain.
Kojen's prompt tuning and model search run on hosted models against your own test cases, and text LoRA trains on real GPUs (for accounts with GPU training enabled) when you need weights to change.
Start training free