Definition
The full bill — chips, electricity, data, and expert salaries — to train one of the most advanced AI systems, now tens of millions to over $100M per run.
At a glance
- A single frontier run costs about $40M to $190M today: GPT-4 near $78M-$100M, Gemini Ultra near $190M[4].
- Chips and their power eat half to two-thirds of the bill; expert salaries are the next slice (about a third)[1].
- The headline figure counts only the final successful run, so true program cost runs several times higher.
- Costs have grown about 2.4x per year since 2016[2].
What you pay for
Mostly scarce machines and scarce people, not electricity. Renting GPUs and powering them is roughly 47-67% of cost; researcher salaries are 29-49%; raw power is just 2-6%[1].
Why the number understates it
The advertised price is one run that worked. Teams also pay for failed runs, experiments, and data prep. DeepSeek’s reported $5.6M covered only final compute, not infrastructure or failures[4].
Where it’s heading
If the trend holds, the biggest runs top $1 billion around 2027[3]. Only a few giants can compete — for most businesses, renting access beats building.
Bottom line
A tens-to-hundreds-of-millions undertaking dominated by chips and talent, doubling yearly — a race only a few giants can run, so nearly everyone else should rent, not build.
References
- How much does it cost to train frontier AI models? — Ben Cottier, Robi Rahman Epoch AI epoch.ai
- The rising costs of training frontier AI models — Ben Cottier, Robi Rahman. arXiv arxiv.org
- Training compute costs are doubling every eight months for the largest AI models. Epoch AI epoch.ai
- AI Training Costs 2026. Local AI Master localaimaster.com
Comments
Questions, corrections, and links welcome. Be specific and civil.