Skip to content

Cost reference

What it costs to train a 1B model from scratch

A compute-optimal run for a 1B-parameter model, 20B tokens at 20 per parameter, takes about 84 H100-hours, around $370 at $4.39 an hour. Training on 1,000B tokens (1T tokens, closer to how small production models are trained) raises that to about $18,495. Compute-optimal costs a few hundred dollars; a competitive 1B model trained on trillions of tokens costs hundreds of thousands.

By Will Schulz · Checked 15 November 2026 · Prices as of November 2026

Compute-optimal: 20B tokens

GPUGPU-hoursCostTime on 8Time on 64Time on 512GPUs to hold state
RTX 5090209.5 TFLOPS · $1.39/hr398$5532 days6.2 hours47 minutes1
A100 80GB312 TFLOPS · $2.19/hr267$58533 hours4.2 hours31 minutes1
H100 SXM989 TFLOPS · $4.39/hr84$37011 hours1.3 hours10 minutes1
H200989 TFLOPS · $5.29/hr84$44611 hours1.3 hours10 minutes1
B2002,250 TFLOPS · $7.99/hr37$2964.6 hours35 minutes4 minutes1

Longer training: 1,000B tokens

GPUGPU-hoursCostTime on 8Time on 64Time on 512GPUs to hold state
RTX 5090209.5 TFLOPS · $1.39/hr19,889$27,6453.4 months13 days39 hours1
A100 80GB312 TFLOPS · $2.19/hr13,355$29,2472.3 months9 days26 hours1
H100 SXM989 TFLOPS · $4.39/hr4,213$18,49522 days3 days8.2 hours1
H200989 TFLOPS · $5.29/hr4,213$22,28722 days3 days8.2 hours1
B2002,250 TFLOPS · $7.99/hr1,852$14,79610 days29 hours3.6 hours1

Compute = 6 × 1e9 parameters × tokens, at 40% utilisation of dense BF16 peak, with perfect scaling across GPUs. Memory for weights, gradients and Adam state is 16 GB before activations.

What real runs of this size reported

RunSize and dataReported
llm.c GPT-21.6B parameters, 33.6B tokens, 8 H100s for 24 hours$672
Llama 3.2 1Bup to 9T tokens370K H100-hours

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 1B parameters at 20B tokens on H100s.

Estimated cost

$370

H100 SXM-hours

84

Time on 8 GPUs

11 hours

Compute

1.2e20 FLOPs

Full training needs about 16 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 1 GB for 4-bit weights, which fits on a single 32 GB card.

Frequently asked questions

How much does it cost to train a 1B model?
About $370 on H100s at $4.39/hour for a compute-optimal run of 20B tokens (6 × parameters × tokens at 40% utilisation). At 1,000B tokens (1T tokens, closer to how small production models are trained) it is about $18,495. 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 1B model?
About 11 hours on one 8-GPU H100 node for 20B tokens, assuming perfect scaling, or 1.3 hours on 64 GPUs.
How much GPU memory does training a 1B model need?
About 16 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 1B model?
At on-demand November 2026 rates, the B200 ($7.99/hour) gives the lowest compute cost, about $296, 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:

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

  1. [1]Kaplan et al., 2020. Scaling Laws for Neural Language Models (arXiv:2001.08361)
  2. [2]Hoffmann et al., 2022. Training Compute-Optimal Large Language Models (arXiv:2203.15556)
  3. [3]Meta, 2024. The Llama 3 Herd of Models (arXiv:2407.21783)
  4. [4]Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
  5. [5]Runpod. GPU pricing (updated 12 November 2026)
  6. [6]Karpathy. Reproducing GPT-2 (1.6B) in llm.c
  7. [7]Meta. Llama 3.2 model card

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