How Much Does It Cost to Train an AI Model? What to Know
Frontier AI training costs have grown about 2.4 times a year. Here is what the money pays for, why you never see that bill, and what you pay for instead.
How much does it cost to train an AI model? For a frontier model, the largest systems from the biggest labs, published research puts the amortized cost of the final training run growing at about 2.4 times per year since 2016, with the largest runs projected to exceed a billion dollars by 2027. Fine-tuning a small model that already exists is a different activity at a different scale, and nobody publishes a single number for it. The figure you read depends entirely on which of those two things someone means.
The best public source is a paper from Epoch AI and collaborators, The rising costs of training frontier AI models, by Ben Cottier, Robi Rahman, Loredana Fattorini, Nestor Maslej, Tamay Besiroglu and David Owen. It builds a cost model from hardware, energy, cloud rental and staff, because the labs themselves rarely publish what a run cost. Every number below is an estimate from that kind of model, not an invoice.
How much does it cost to train an AI model? What the money pays for
Epoch's summary of the work, How much does it cost to train frontier AI models?, gives the shares that make up the total development cost of a frontier model:
Cost category | Share of total | What it covers |
|---|---|---|
Hardware | 47% to 67% | AI accelerator chips, other server components, interconnect between machines |
R&D staff | 29% to 49% | Researchers and engineers, including equity compensation |
Energy | 2% to 6% | Electricity for the training run |
Two things stand out. First, energy, the part people argue about most, is the smallest line. Second, staff are not a footnote. A frontier model is built by a team, and the people are almost as large a cost as the chips.
The growth rate is the headline: about 2.4 times a year for the amortized hardware and energy cost of the final training run. Compounding at that pace turns a cost measured in millions into one measured in billions within a handful of years, which is the logic behind the billion-dollar projection.
Why "final run" is a misleading number
When a lab says a model cost a certain amount to train, it usually means the last big run. That leaves out:
Failed and experimental runs. Labs train many smaller versions and abandon many attempts before the one that ships. Epoch's model covers the whole development cost, not only the last run, which is one reason it counts staff.
Data work and later stages. Collecting and cleaning training data is real labour, and the stage after pretraining takes its own compute and people.
Amortization. Chips bought for a model are used for other work too. Cost models spread the purchase price over the hardware's life, which is why you see the word "amortized".
If you want the mechanics of the stage that consumes most of the compute, what pretraining in AI is explains it.
Training is not the bill you pay
This is the part that matters if you are building on AI rather than building AI. Four different costs get mixed up in conversation:
Cost | Who pays it | Scale |
|---|---|---|
Training a frontier model from scratch | The lab | Enormous, the figures above |
Fine-tuning an existing model | You, if you choose to | Far smaller, depends on method and data size |
Running the model for your users (inference) | You, per use, via API pricing | Small per request, large in total if usage grows |
Hosting an open model yourself | You | Hardware or rental plus someone to run it |
For almost every builder the relevant line is the third one. You pay per token, as your users send requests, and that is why the per-token price in an API pricing page matters more to your margins than any training figure. Why bigger AI models cost more per token explains how the two connect, and why AI API prices keep falling explains why that bill has tended to drop even as training gets more expensive.
If you are weighing the self-hosting route, the arithmetic in is self-hosting an open-weight model cheaper than an API is more useful than any training number.
What this means for builders
Do not budget from headlines. A story about a model that "cost" a large sum to train says nothing about what it will cost you to call it.
Expect the frontier to stay concentrated. If the cost of the largest runs keeps climbing at that rate, only the best-funded organizations can afford them. The Epoch authors draw that conclusion themselves. For you it means a few providers set the terms, which is a reason to avoid tying your product to one of them.
Remember the cheap path exists. Small and open models, and fine-tuning of existing ones, mean many useful systems cost nowhere near frontier money. What a small language model is shows what is possible at that scale.
Treat every figure as an estimate. Different sources use different scope, so compare numbers only when they count the same things. The pillar guide on how AI models work gives the background for reading them.
Frequently asked questions
How much does it cost to train a large language model?
For frontier models, the best published research finds the amortized cost of the largest runs growing about 2.4 times a year since 2016. The exact amount for a given model is rarely disclosed by its lab, so every figure you see is an estimate.
What is the biggest cost in training an AI model?
Hardware is the largest category in Epoch's cost model at roughly 47% to 67% of total development cost. Research and engineering staff come next at about 29% to 49%, and energy is the smallest at about 2% to 6%.
Is training more expensive than running the model?
For the lab, training is a large upfront cost and running the model for millions of users is an ongoing one. For you as a builder, you never pay the training cost, only the usage cost per request.
Can a small business train its own AI model?
Training one from scratch to frontier quality is out of reach. Fine-tuning or running an existing open model is realistic. Whether it beats paying for an API depends on your volume and skills.
Will training costs keep rising?
The study projects continued growth if the trend holds, with the largest runs exceeding a billion dollars by 2027. Trends can change, so treat that as a projection, not a promise.
How did this land?
About the author

Staff Engineer, Platform
Carlo works on the platform that turns prompts into running apps. He writes the engineering deep dives and the changelog notes worth reading.


