Microsoft is slowly pivoting away from reliance on OpenAI. The company is embedding its own AI models into Excel and Word, using in-house solutions to power a percentage of user requests. It’s a strategic shift driven by OpenAI’s pricing and Satya Nadella’s new caution about frontier AI safety.

This month, Microsoft announced $2.5 billion for a new division, Microsoft Frontier Company, dedicated to enterprise AI deployment. The division has 6,000 engineering and industry experts focused on delivering working AI solutions, not just selling access to models.

The OpenAI Relationship Changes

Microsoft and OpenAI remain partners. But the partnership has friction now. OpenAI announced GPT-5.6 would be the preferred model for Microsoft Copilot 365. That’s contractual obligation. Meanwhile, Microsoft trains cheaper alternatives in-house for specific use cases.

Excel and Word AI features now run on Microsoft’s MAI models for certain operations. Users won’t see the difference. The magic is silent. But internally, Microsoft avoids paying OpenAI per token for every completion.

The Safety Warning

Nadella has become more public about frontier AI risks. In mid-July, he issued a warning to companies deploying AI without guardrails. The message is clear: frontier models can fail in unpredictable ways. Companies need safeguards, testing, monitoring. It’s not just deploy and pray.

This stance justifies Microsoft’s investment in controlled deployment at scale. Microsoft Frontier Company exists to help enterprises avoid catastrophic AI failures.

The Competitive Reality

Google Cloud and AWS are competing on AI APIs and infrastructure. Microsoft is competing on enterprise workflow integration. Copilot in Office, Teams, and now deployed through Frontier Company creates stickiness. Customers adopt one Microsoft AI tool, then expand because it’s already in their workflows.

Microsoft’s AI strategy is less about leading frontier research and more about dominating enterprise deployment. OpenAI builds models. Microsoft will build the systems that use them reliably.