Those poor AI companies are paying for a lot more than just training runs, explaining their sensitivity to threats such as model distillation Credit: Rob Schultz / Shutterstock / Unsplash AI models cost a lot more to develop than you may think. AI research company Epoch AI has set out all the costs of building a new AI model — and explaining why AI companies are so concerned about perceived threats to their intellectual property. It has looked into this before: Last year, it estimated that of OpenAI’s $5 billion expenditure on R&D, only about 10 percent went on the final training runs, with the majority going on scaling, synthetic data generation, and basic research. At the time, Epoch was unsure whether this was a peculiarity of OpenAI but now two Chinese companies, MiniMax and Z.ai, have also disclosed their R&D compute spending, and Epoch has found that, despite the differences in company size, final training runs are only a small part of the Chinese companies’ R&D expenditure too. Epoch set out more detail about the issue. It said that if “most of the spending is exploration rather than execution, then a competitor who learns what works from the frontier could replicate the results for a fraction of the original cost.” This has been a concern of US AI companies for some time. Google has already expressed concerns about intellectual property theft. And Anthropic has fingered MiniMax as a company that has sought to extract Claude’s capabilities to enhance its own offerings. It’s clear that any business looking to develop AI models is going to be committing to spend huge sums of money: The training is just a small part of it. Artificial Intelligence