Those poor AI companies are paying for a lot more than just training runs, explaining their sensitivity to threats such as model distillation

InfoWorld is a source of news, analysis, and commentary on technology trends, IT strategies, and business innovation. With a focus on enterprise technology and digital transformation, InfoWorld offers insights and guidance for IT decision-makers, software developers, and technology professionals. From  articles on cloud computing and cybersecurity to product reviews and industry trends, InfoWorld helps readers navigate the complexities of modern IT environments and make informed decisions to drive business success.

InfoWorld

Epoch AI research reveals that final training runs account for only a small fraction of total AI model development costs. Analysis of OpenAI's $5B R&D spend shows roughly 10% went to final training, with the bulk going to scaling, synthetic data generation, and basic research. Similar patterns were found at Chinese AI companies MiniMax and Z.ai. This cost structure explains why AI companies are highly sensitive to model distillation and IP theft — a competitor who learns what works from a frontier model could replicate results at a fraction of the original cost. Google and Anthropic have both raised concerns about such extraction attempts.

Final training of AI models is a fraction of their total cost