How Much Does It Cost to Train an AI Model?

Executive summary

Training a frontier AI model has become one of the most expensive things a company can do. This report ranks the known training-compute costs of 15 notable AI models, from GPT-4's roughly $37m in 2023 to the about $388m spent on Grok 4 — a rise of roughly 10× in two years. Epoch AI, whose estimates this uses, finds frontier training-compute cost growing about 2.4× a year, putting the first billion-dollar training runs within reach by 2027. One caveat matters: these are the costs of the compute used to train each model, not the all-in cost including researchers, data and amortised hardware, which is higher. Every figure is sourced to Epoch AI and the full table is downloadable under CC BY 4.0.

Key findings

The most expensive AI models to train

The known training-compute costs of frontier AI models, ranked (US$ millions). Grok 4 tops the list at about $388m; the full table below has every model with a cost estimate.

How fast is the cost rising?

The cost of the most expensive training run each year (US$ millions). It has climbed from about $37m in 2023 to $388m — roughly 2.4× a year, one of the steepest cost curves in technology.

What makes training an AI model so expensive?

Almost all of the cost is compute: the thousands of specialised AI chips (mostly Nvidia GPUs) running for weeks or months, and the electricity to power and cool them. A frontier run uses tens of thousands of chips at once — hardware that costs tens of thousands of dollars each — so the compute bill dominates.

Cost scales with training compute, which itself is growing about 4–5× a year as labs chase capability. Because compute rises faster than chips get cheaper, the dollar cost keeps climbing. The correlation between a model's training compute and its cost is near-perfect — the chart below plots the two.

What these figures leave out matters. Training-compute cost excludes the researchers' salaries, the data pipeline, failed experiments and the amortised cost of buying (rather than renting) the hardware. Counting those, the true bill for a frontier model is materially higher than the compute cost alone.

Full data — 15 models by training cost

Every notable AI model with a published training-compute cost estimate, with its compute and release date. Source: Epoch AI (2023 USD).

Limitations & notes

These are estimates of training-compute cost in 2023 US dollars, from Epoch AI, which models the number of chips, their utilisation, training time and cloud/hardware prices. They are not company-disclosed figures — labs rarely publish training costs — and carry real uncertainty, so treat each as a central estimate, not a precise invoice.

Crucially, training-compute cost is not the all-in cost of a model: it excludes salaries, data, failed runs and amortised hardware. It also excludes inference — the cost of running the model for users — which over a model's life can exceed its training cost. The full dataset, with the underlying compute figures, is downloadable below.

Scoreboard (machine-readable data)

Every headline indicator with its value, period, source and confidence. Free to reuse under CC BY 4.0.

↓ CSV · ↓ JSON

Methodology & verification

Training-compute cost estimates from Epoch AI's Notable AI Models database, in 2023 US dollars. Epoch estimates the cost of the compute used to train each model from the number and type of chips, hardware utilisation, training duration and cloud/hardware prices. These are compute costs only — they exclude staff, data, failed experiments, amortised hardware and inference. Models without a published Epoch cost estimate are excluded. Nothing is estimated by Affärslivet. Confidence: Medium-High (estimates, not disclosures).

Data dictionary

Frequently asked questions

Glossary

Training-compute cost

The estimated cost of the computing power used to train a model — chips, their utilisation and the electricity — in 2023 US dollars. Excludes staff, data and hardware amortisation. ↗

Training compute (FLOP)

The total number of arithmetic operations used to train a model; the main driver of training cost. ↗

Inference

Running a trained model to produce outputs — a separate, ongoing cost from training. ↗

FLOP

A floating-point operation; AI training is measured in the total FLOP performed, often 10^25 or more for frontier models. ↗

Sources

• Epoch AI — Notable AI Models database (CC BY 4.0)
• Epoch AI — training-cost trends
• Stanford HAI — AI Index Report 2026 (context)

Related reports & data

• AI Compute Index
• The Largest AI Companies
• AI Model Benchmarks
• The Biggest AI Data Centers