Final Training Runs Account for a Minority of R&D Compute Spending

Key Takeaways

• Final training runs <25% of total AI R&D compute.
• OpenAI’s R&D spend $5 B, training share $500 M.
• MiniMax and Z.ai spend $141 M/$216 M R&D respectively.
• Majority of spend goes to experiments, data generation, research.
• Catch‑up firms need less experimentation when learning from frontier.

Summary

The analysis shows that final training runs represent only a minority of AI R&D compute spending. Across OpenAI, MiniMax and Z.ai, final runs account for 9.6%, 22.6% and 12.3% of total compute respectively. OpenAI’s 2024 R&D compute bill was about $5 billion, with roughly $500 million devoted to the final training of released models, while MiniMax and Z.ai spent $141 million and $216 million in total. The pattern persists despite differences in scale, geography and business models, indicating that most compute is consumed by experimentation, data generation and other research activities.

Pulse Analysis

Understanding how AI firms allocate compute reveals a hidden layer of expense that most industry reports overlook. While headlines often cite the massive GPU hours required to train a flagship model, the bulk of spend actually fuels a relentless cycle of scaling experiments, synthetic data creation, and prototype training that never reaches production. This exploratory phase consumes the majority of the budget, as shown by OpenAI’s $5 billion R&D outlay in 2024, of which only a tenth funded the final GPT‑4.5 run. The same dynamic appears in smaller Chinese players, where total R&D budgets of $141 million (MiniMax) and $216 million (Z.ai) are largely devoted to non‑final work.

The distinction matters for policymakers and investors alike. Compute‑based thresholds used to gauge AI risk or to set regulatory caps often assume that the headline training cost reflects total development expense. By ignoring the extensive exploratory spend, such metrics can dramatically underestimate the resources required to bring a model from concept to market, leading to lax oversight or mispriced investment risk. Moreover, the heavy front‑loading of experimentation creates a barrier to entry: newcomers must either replicate costly trial‑and‑error cycles or find ways to shortcut them by learning from frontier firms.

A key question is whether lagging firms can reduce their exploratory overhead by borrowing insights from leaders—a hypothesis known as the catch‑up effect. MiniMax’s higher training‑to‑R&D ratio hints that it may be leveraging frontier knowledge to focus more on final runs, whereas Z.ai’s ratio mirrors OpenAI’s, suggesting a less efficient learning curve. However, with only three data points and wide confidence intervals, the evidence remains tentative. Expanding the dataset through more IPO disclosures or voluntary reporting will be essential to confirm whether knowledge diffusion truly compresses R&D compute for emerging competitors.