How much does it cost to train an AI model?

Want more stories like this?
Follow Omni on Google.
The 2017 model that gave birth to today's chatbots cost an estimated $900 to train, while one of the biggest models of 2025 cost an estimated $490 million, close to the budget of one of the most expensive movies ever made.
Key takeaways
- The cost of training top AI models has been multiplying by more than two every year.
- The first billion-dollar training run may be only months away.
- The real bill is even bigger than these estimates.
From a laptop to a Star Wars movie
Every AI model starts with a training run, weeks or months in which thousands of computer chips crunch through huge amounts of text and data until the model learns to write, answer and reason. That computing power is expensive, and nobody has tracked its price more closely than the research group Epoch AI.
Here's how the estimated cost of training a single model has climbed, based on Epoch's figures:
- Transformer, 2017: about $900. Google's research model introduced the design behind ChatGPT, Claude and Gemini. Training it cost less than a decent laptop.
- GPT-4, 2023: about $78 million. Six years later, the model that powered ChatGPT's paid plan at the time cost tens of thousands of times more.
- Gemini Ultra, 2023: about $191 million. Google's top model that year more than doubled GPT-4's bill.
- Grok 4, 2025: about $490 million. Epoch called xAI's model the largest AI training run to date when it analyzed it in September 2025.
That last figure is close to the cost of making Star Wars: The Force Awakens, which ran to about $533 million according to company filings reported by Forbes, one of the most expensive films ever made. In other words, training one AI model now costs about as much as a record-breaking blockbuster.
From $900 to $490 million is a jump of more than 500,000 times in about eight years. Epoch's research puts the long-run growth at about 2.4 times a year, meaning the cost of the most expensive training runs has been more than doubling every single year since 2016. If that pace continues, Epoch expects the first billion-dollar training run around the start of 2027.
Why it got so expensive
The simple answer is scale. Each new generation of top models is trained on more data, with more chips, for longer. Computer chips do get cheaper for the same amount of work over time, but nowhere near fast enough to keep up with how much more computing each new model uses.
According to Epoch's research, the AI chips themselves make up the biggest share of the bill, around 44% on average. Other server parts take about 29%, the networking that links thousands of chips together about 17%, and electricity only about 9%.
The real bill is bigger
These estimates cover only the final training run, the last big push that produces the finished model. They leave out a lot.
The researchers and engineers who build the model don't appear in the figure, and neither do the many smaller experiments and failed attempts before the final run, or the cost of collecting and preparing training data. When Epoch looked at the full development cost for a few models, staff alone made up roughly a fifth to a half of the total, depending on whether stock-based pay is counted.
And the chips used to train a model are worth far more than the share of their cost that one run uses up. Epoch estimates that the hardware used for GPT-4, for example, cost about $800 million to buy.
That's the bigger point behind these numbers. With training bills now in the hundreds of millions of dollars and climbing, only a handful of the best-funded companies in the world can afford to build the most capable models.
Methodology
All training costs are estimates of the computing cost of each model's final training run. The figures for the Transformer, GPT-4 and Gemini Ultra come from Epoch AI's estimates as published in Stanford's 2024 AI Index report. The Grok 4 figure comes from Epoch AI's September 2025 analysis. The growth rate, the cost breakdown, the share of staff costs and the GPT-4 hardware value come from Epoch AI's research paper The rising costs of training frontier AI models. The film budget is the production cost of Star Wars: The Force Awakens from company filings, as reported by Forbes in 2023. Dollar figures are shown as originally estimated and are not adjusted for inflation.
loading




