Final Training Is Only 10% of What Frontier AI Labs Actually Spend

The number that usually gets published when a frontier AI model launches is the cost of its final training run. The number that almost never gets published is everything that came before it. A new analysis from Epoch AI suggests that for the most advanced AI labs, the final training run represents roughly 10 percent of total R&D compute spending. The other 90 percent goes to scaling experiments, synthetic data generation, failed runs, and basic research.

Epoch AI estimated OpenAI 2024 compute spending at approximately $5 billion in total R&D, with only around $500 million attributable to the final training runs that produced released models like GPT-4.5. The remaining roughly $4.5 billion went to the less visible work of figuring out what to build before building it.

The evidence for this split is not limited to OpenAI. MiniMax and Z.ai, two Chinese AI companies, disclosed their R&D compute spending as part of their IPO processes on the Hong Kong Stock Exchange in early January 2026. By matching their reported spending windows against the training compute of models released shortly after, Epoch AI found the same pattern at different scales and geographies. Final training runs accounted for a minority of compute spending in both companies, despite major differences in size and business model.

The implications for policy are the most uncomfortable part. Public debates about compute thresholds for AI regulation typically focus on the cost of the final training run as the relevant input. But if the full cost of developing a frontier model is roughly 10 times the headline number, then the compute threshold that policymakers treat as significant is actually capturing only a fraction of what frontier development actually costs. A competitor who can observe what the frontier produces and then replicate the result with fewer experiments would face a much lower effective cost than the original developer paid.

MiniMax, which is further behind the frontier than OpenAI, showed a higher ratio of training compute to total R&D compute than OpenAI, which is consistent with the catch-up theory: companies that can learn from the frontier should need less experimentation. Z.ai, which is closer to the frontier than MiniMax, did not follow that pattern, and Epoch AI notes that three data points are insufficient to draw strong conclusions about the relationship between position and efficiency.

The methodology relies on assumptions about GPU utilization rates and GPU hour costs, which Epoch AI acknowledges involves uncertainty. The company publishes its full estimation procedure, which is more transparency than the labs themselves offer. But the directional claim is robust: final training runs are the minority of the spending, and anyone trying to understand the actual economics of frontier AI development is missing the bigger picture if they only look at the headline training cost.