OpenAI and Anthropic are taking opposite bets on agent-training compute: buy vs. rent

OpenAI and Anthropic appear to be making very different infrastructure bets for training AI agents. Reporting says OpenAI has bought tens of thousands of Mac minis and Mac Studios for reinforcement learning and computer-use-agent work. Anthropic, by contrast, is reportedly renting similar Apple-silicon capacity through AWS.

That difference matters because agent training is not just about having the biggest model. It involves running huge numbers of simulated tasks, testing whether an agent can use software correctly, and feeding the results back into training. That can require a lot of reliable compute, often for long periods. Buying machines gives OpenAI more direct control over capacity and potentially lower costs if usage stays high. Renting gives Anthropic more flexibility, but it means depending on cloud availability and pricing.

The reported demand is also putting pressure on Apple’s supply chain, with Tim Cook warning of a multi-month shortage. For business owners, the immediate lesson is simple. The infrastructure underneath AI is becoming a constraint again. If your plans depend on AI agents handling customer support, operations, analysis, or internal workflows, don’t assume compute capacity and model access will always be cheap or instantly available.

This is the kind of thing we build into an AI command centre, tracking where the practical bottlenecks are so AI plans are based on what can actually be deployed.