Microsoft Bets on In-House AI to Cut OpenAI and Anthropic Costs

Microsoft is shifting AI infrastructure strategy, routing tens of thousands of prompts weekly through proprietary MAI models to reduce dependence on OpenAI and Anthropic. The company launched seven MAI models in June 2026, including MAI-Thinking-1, which the company says performs on par with Anthropic's Claude Opus 4.6 on coding benchmarks.

The pivot targets Microsoft's cost structure. By converting rented AI capacity into owned infrastructure, the company aims to improve ROI on its AI spending—a priority as cloud margins face scrutiny. The move also strengthens Microsoft's negotiating position with third-party providers, reducing pricing leverage for OpenAI and Anthropic in future contract renewals.

MAI-Thinking-1's performance against Claude Opus 4.6 suggests Microsoft can now handle more complex workloads internally, potentially slowing the growth rate of its payments to external AI vendors. The scale—tens of thousands of prompts weekly—indicates the initiative has moved beyond pilot phase.