Training a frontier AI model — one of the largest, most capable systems available at a given time — is one of the most expensive undertakings in modern technology, with costs that have climbed dramatically as models have grown larger.
Where the cost comes from
The biggest expense is compute: specialized processors called GPUs or TPUs, run for weeks or months across thousands of interconnected chips, consuming enormous amounts of electricity. On top of raw compute, there’s the cost of collecting, cleaning, and licensing training data, paying the research and engineering teams who design and tune the model, and running the extensive safety testing required before release.
Why costs keep rising
Larger models trained on more data have generally produced better capabilities, which has pushed labs toward bigger training runs. At the same time, competition for advanced chips and the electricity needed to power large data centers has driven up costs across the industry, making frontier model training accessible to only a small number of well-funded labs.
Why this shapes the industry
The enormous cost of frontier training is a major reason the AI industry has consolidated around a handful of well-capitalized labs and their major cloud and investment partners. It’s also why smaller companies and researchers often build on top of existing models — through fine-tuning or API access — rather than training frontier-scale models from scratch, and why open-source models trained by well-funded organizations have become an important resource for the broader research community.