AI training cost calculator,
GPU-hours and dollars, not just FLOPs.

Compute guide

What does it actually cost to train a model?

The naive version of this estimate — GPU-hours at the chip’s advertised peak speed, times a sticker price — is almost always wrong, and wrong in a predictable direction: it undersells the real cost. Real training clusters never sustain a GPU’s full advertised throughput, and the gap between “advertised” and “actual” is large enough to matter for anyone budgeting a training run.

The FLOPs formula: 6 × N × T

Total training compute is well-approximated as 6 × N × T, where N is the model’s parameter count and T is the number of training tokens. The 6 comes from 2 FLOPs per parameter per token in the forward pass (one multiply, one add) plus a backward pass that costs roughly twice the forward pass — 2 + 4 = 6. This is the same formula behind the Chinchilla scaling-law paper and essentially every public training-compute estimate since.

Why MFU changes everything

Model FLOPs Utilization (MFU) is the fraction of a GPU’s advertised peak throughput a real training job actually sustains, after accounting for memory bandwidth limits, network overhead between GPUs, and software inefficiency. Even well-tuned production runs on modern hardware typically land in the 35-45% range — never close to 100%. Costing a training run at the advertised peak (skipping MFU entirely) is a common mistake, and it doesn’t just make the estimate a little optimistic — at the 40% default this calculator uses, the naive figure is 2.5× too low. This calculator shows both numbers side by side specifically so that multiplier is visible, not hidden inside a single result.

H100 vs A100, on-demand vs spot

An H100 delivers roughly 989 TFLOPS of dense BF16 throughput against an A100’s 312 TFLOPS — about 3.2× more compute per GPU — but also costs meaningfully more per hour. The right comparison is cost per FLOP, not cost per hour, and it depends on each provider’s current rate, which is why this calculator lets you compare both GPUs directly rather than asserting one is always cheaper. On-demand rates are also 2-4× spot/marketplace rates for the same hardware — spot is substantially cheaper but can be interrupted mid-run, a real tradeoff for anything beyond a rough estimate.

A worked example

Take a 70B-parameter model trained on 1.4T tokens (Chinchilla’s own headline example) on H100s at a representative $12.29/hr on-demand rate, 40% MFU: total FLOPs ≈ 5.88 × 10²³. At H100’s effective throughput (989 TFLOPS × 40%), that’s roughly 413,000 GPU-hours — about $5.1M at that rate. Costing the same run at a naive 100% MFU would show only $2.0M — a genuinely misleading underestimate, not a rounding difference.

Chinchilla-optimal tokens, and why labs often ignore it

Chinchilla’s compute-optimal guideline is roughly 20 tokens per parameter — a 70B model pairs with about 1.4T tokens, a 7B model with about 140B tokens. This ratio minimizes training loss for a fixed compute budget, but it is a guideline for training efficiency, not for the resulting model’s cost to actually run. Production labs frequently overtrain smaller models well past the Chinchilla-optimal ratio on purpose — a smaller, more-trained model is more expensive to train but meaningfully cheaper to serve at inference time, and inference cost is paid on every single request forever, not once.

Disclaimer

This calculator estimates directional training cost from published GPU specifications and commonly-cited cloud pricing — it is not a quote. GPU cloud pricing is one of the fastest-moving cost categories in software and can shift meaningfully within weeks; always confirm current rates directly with your cloud provider before budgeting a real training run. This tool also does not model data-pipeline cost, storage, networking egress, or the compute spent on failed runs and hyperparameter search — all of which add to the real total cost of training a model.