Final training runs account for a minority of R&D compute spending
Most of the money spent on AI research goes toward experiments rather than the final model training.
Recent analysis shows that final training runs—the ones that produce a named model—are only a small part of the total R&D cost. For example, estimates suggest that only about 10% of OpenAI’s 2024 R&D compute went to the final training runs for released models. The rest was spent on scaling tests, synthetic data, and other research workloads that were never released.
This matters because it shows that the cost of a “finished” AI tool is much higher than just the final step. For public sector leaders, this highlights a key point for budgeting and governance. We cannot just look at the final product; we must account for the extensive exploration and failed experiments required to make a model reliable and ready for public use.
How does this change your perspective on the actual costs of building AI?
#AI #PublicSector #TechPolicy #DataScience #Innovation