The article explains that the widely discussed “training” phase of AI models is actually just a small part of the total cost of building advanced systems. Research from Epoch AI shows that companies spend far more on activities leading up to training—such as experimentation, data preparation, and scaling infrastructure—than on the final training run itself. This challenges the common perception that training large models is the primary expense.
For example, analysis of OpenAI’s reported $5 billion R&D spending found that only about 10% went toward the final training stage, while the majority was spent on earlier processes like generating synthetic data, testing different approaches, and foundational research. Similar spending patterns were observed in Chinese AI firms such as MiniMax and Z.ai, suggesting this is an industry-wide trend rather than a one-off case.
This cost structure has important competitive implications. Since most of the investment goes into exploration and figuring out what works, a competitor that learns from these breakthroughs could replicate similar models at a much lower cost. In other words, the expensive part is discovery—not execution—making intellectual property protection a major concern for AI companies.
Overall, the article highlights a shift in how we understand AI economics. Building cutting-edge models is less about a single expensive training run and more about a long, resource-intensive process of research, iteration, and scaling. This explains why companies invest billions upfront and why they are highly sensitive to risks like model copying or distillation, which could undercut their massive investments.
