LLM training cost calculator
10,000 H100 GPU-hours on the cheapest bookable 8-GPU H100 node (Lambda, $3.99 per GPU-hour) cost $39,900 in compute today. Enter your own hours, estimate them from model size and training tokens, or use a throughput you measured.
Training budget formula
Compute bill = hours × GPU count × rate per GPU-hour. The runtime comes from one of three places:
- Hours you expect to run, entered directly.
- Model size and tokens: training takes about 6 floating-point operations per parameter per token (2 forward, 4 backward), so hours = 6 × parameters × tokens × epochs ÷ (GPU count × the GPU's dense 16-bit TFLOPS × utilisation) ÷ 3,600. Utilisation is the share of that datasheet figure your run actually sustains. The 40% it starts on is an example to replace, not a measured figure: it varies with the model, the batch, the software and the number of GPUs, so use a figure measured on your own setup, or the measured-throughput option, when you have one. The estimate covers full training of a dense model; adapter fine-tuning (LoRA) skips most of the backward pass for the frozen weights and costs less.
- Your measured throughput: hours = tokens × epochs ÷ tokens per second for the whole node ÷ 3,600. The most reliable figure, once you have run a few steps.
What the estimate leaves out
Model size does not tell you whether the run fits in memory: optimizer state, activations, context length and precision all add to it. Compare inference, adapter training and full-training memory.
Rates checked 22 October 2026. This estimate excludes storage, data transfer, checkpoint overhead, evaluation, failed runs and idle time outside your runtime input. Add those hours explicitly and use the job cost calculator for published storage and egress charges.