The glamorous version of AI strategy is capability.
The real version is the bill.
TechCrunch reports that Microsoft has begun relying more on its own in-house MAI models for some prompts in Word and Excel, reducing dependence on OpenAI and Anthropic for parts of Microsoft 365. Microsoft still uses third-party models, but the direction is clear enough: even the company closest to OpenAI wants more control over cost, latency, routing, and margins.
That should make everyone sit up.
Not panic. Just sit up.
Microsoft is not a random startup trying to save a few dollars on inference. It is the company with Azure, GitHub, Office, Windows, Teams, Copilot, enormous distribution, and deep model partnerships. If Microsoft is still pushing hard toward its own model family, then the lesson is not subtle.
Owning the interface is not enough.
Owning the model is not enough either.
The advantage is in controlling the whole operating surface: which model answers which prompt, which workload deserves premium reasoning, which task can run on a cheaper model, and which internal data can safely tune something better for the company.
Token cost is becoming strategy
Microsoft’s own MAI announcement framed its models around real-world tasks, coding, transcription, image generation, reasoning, and workflow adaptation. The company also claimed major efficiency gains from tuning models for specific enterprise work.
That is the next phase.
The market is moving away from “use the biggest model for everything” because that was never going to survive contact with finance teams. Frontier models prove what is possible, but production systems need routing, budgets, and cheaper specialized models that do most of the daily work without burning the house down.
This is not bad news for AI.
It is AI growing up.
The winners will not be the teams with the most romantic model loyalty. They will be the teams that know when to use the expensive brain, when to use the fast brain, when to use the local brain, and when to not use AI at all.
That last one may become the most underrated optimization in the whole industry.
Sources: TechCrunch, Microsoft AI, GitHub Changelog