Microsoft Swaps Out OpenAI: Why MAI Models Are Quietly Taking Over Excel and Outlook

Microsoft is replacing OpenAI. Not everywhere, not all at once, but in the two apps you probably open every single workday: Excel and Outlook. And most people haven’t noticed a thing.

That’s the strange part. Microsoft’s own MAI family of models has started handling production workloads that used to run on OpenAI and Anthropic frontier models. Tens of thousands of weekly prompts inside Excel and Outlook are now processed by MAI systems instead. If you use Copilot features in those apps, some of your requests may already be going to a Microsoft model. You just weren’t told.

Why would Microsoft ditch the models it helped fund?

Let me be direct: this is about control and cost, not just capability. Microsoft has poured billions into OpenAI. But relying on someone else’s model for features embedded in your flagship products is a risky place to sit. Every prompt routed to an external lab is a dependency, a cost, and a bit of leverage you’ve handed away.

The Microsoft MAI models change that math. When Microsoft runs its own models on its own Azure infrastructure, it captures the margin instead of paying it out. At the scale of Excel and Outlook, which have hundreds of millions of users between them, even small per-prompt savings compound into serious money. Multiply tens of thousands of weekly prompts across every enterprise seat and you see why the finance team loves this.

There’s also a strategic signal here. For a couple of years the story was “Microsoft = OpenAI’s distribution arm.” Swapping in MAI rewrites that narrative. It says Microsoft intends to own the full stack, from the chips to the chat box. Not everyone agrees this is wise, and honestly, they have a point, because OpenAI’s frontier models still lead on the hardest reasoning tasks. But for routine work like summarizing an email or cleaning up a spreadsheet, you may not need frontier-level power.

What This Means For You

If you’re a regular Microsoft 365 user, the honest answer is: probably not much, day to day. A well-tuned MAI model summarizing your inbox or drafting a formula feels about the same as the model it replaced. Most everyday tasks don’t stress the ceiling of what these systems can do.

If you’re an IT admin or a developer building on Microsoft’s stack, pay closer attention. Model swaps can quietly shift behavior. A prompt that produced one style of output last month might produce a slightly different one now. Test your critical workflows. Don’t assume the output is identical just because the interface is.

And if you run an AI startup, take note of the pattern. The big platform companies are learning that they don’t have to rent intelligence forever. They can build it in-house and route their own traffic to it. That squeezes the middle. Sound familiar? It’s the same playbook cloud providers used with databases and storage.

There’s a data angle too, and it’s easy to miss. When Microsoft runs its own models, it controls where your prompts go and how they’re processed. For enterprises with strict data-residency or compliance requirements, an in-house model on Microsoft’s own infrastructure can be an easier sell than routing sensitive content through a third-party lab. Some IT departments will quietly welcome that. Others will want proof that the new setup is at least as private as the old one. Ask the questions before you assume.

Scale is the number that makes all of this real. Microsoft doesn’t disclose exact figures, but Excel and Outlook together reach hundreds of millions of users, and Google’s data-center electricity use just jumped a record 37% as these AI features spread everywhere. Running frontier-grade models for every one of those interactions is enormously expensive. A cheaper in-house model that’s good enough for routine tasks, while reserving the pricey frontier models for the genuinely hard requests, is simply smart economics.

What Happens Next

Expect the MAI rollout to widen. If Excel and Outlook go smoothly, Word, Teams, and the broader Copilot layer are the obvious next targets. Microsoft won’t announce it loudly, because quiet migration is the point. Fewer headlines, fewer worried customers.

Watch the broader arms race too. This move lands in a week where TSMC’s revenue jumped 34% to $40.2 billion on AI chip demand, Anthropic opened talks with Samsung about a custom accelerator, and OpenAI is reportedly floating an equity stake to the US government. Everyone is racing to own more of the stack: the chips, the models, the deployment muscle. Microsoft building MAI is one more piece of that same land grab.

The open question is quality. If MAI keeps pace with the frontier for everyday tasks, users won’t care who’s behind the curtain. If it slips, expect grumbling and maybe a quiet route back to OpenAI for the heavy stuff. Either way, the era of Microsoft leaning entirely on outside labs is ending.

Key Takeaways

  • Microsoft MAI models are now handling production prompts in Excel and Outlook, replacing some OpenAI and Anthropic usage.
  • The driver is cost and control: running its own models on Azure lets Microsoft capture the margin and cut external dependencies.
  • Everyday users likely won’t notice, but IT teams and developers should test critical workflows for subtle output changes.
  • The shift signals Microsoft’s intent to own the full AI stack, part of a broader 2026 race spanning chips, models, and infrastructure.
  • Word, Teams, and wider Copilot features are the likely next candidates for MAI.

So here’s the real question: if a Microsoft model quietly replaced OpenAI in your daily tools and worked just as well, would you even want to know?