Final training runs account for only a small fraction of AI model development costs, research finds

Training Runs Are Just 10% of AI Development Costs

AI companies spend far more money building models than most people realize, and most of that spending happens before training ever begins.

Research firm Epoch AI analyzed the R&D budgets of major AI developers and found a consistent pattern: final training runs account for only about 10% of total development costs. OpenAI's $5 billion R&D budget breaks down this way, with the remainder going to scaling infrastructure, synthetic data generation, and basic research.

The finding extends beyond OpenAI. Chinese companies MiniMax and Z.ai disclosed similar spending patterns, suggesting this is how frontier AI development works across the industry.

Why This Matters for Competition

The cost structure creates a significant competitive advantage for companies that can learn from competitors' work. Epoch AI noted the risk plainly: "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 concern has driven US AI companies to protect their intellectual property aggressively. Google has raised alarms about IP theft. Anthropic accused MiniMax of attempting to extract capabilities from its Claude model to improve its own offerings.

The Real Cost of AI Development

For managers overseeing AI initiatives, the breakdown matters. Companies investing in model development face substantial costs long before they see returns:

• Infrastructure scaling and compute resources
• Synthetic data generation and curation
• Research and experimentation
• Final model training (the visible part)

Understanding these costs helps explain why AI companies view competitive intelligence and IP protection as business-critical issues.

For those managing AI strategy, AI for Executives & Strategy covers the business economics driving these decisions. Learn more about generative AI and LLM development to understand the technical factors behind these costs.