Microsoft is starting to use more of its own AI models inside products like Excel and Outlook, gradually replacing models from OpenAI and Anthropic as the company looks for ways to bring down the Icarus-esque cost of running AI at scale.
Microsoft begins expanding MAI model deployment
According to Bloomberg, tens of thousands of prompts each week in Excel and Outlook are now being handled by Microsoft’s internally developed MAI models.
That’s still only a small slice of Microsoft’s overall AI workload – Copilot processes millions of prompts every week. But it’s a meaningful sign of where the company wants to go next.
Microsoft has been laying the groundwork for this over the past month. At its Build developer conference in June, Microsoft introduced seven new MAI models for things like coding, transcription, image generation, and voice recognition.
Some are already available in GitHub Copilot, and Microsoft has said its transcription model will start showing up in Teams and other products in the coming months.
AI inference costs are reshaping vendor strategy
For the past few years, Microsoft has enjoyed unusually favorable access to OpenAI’s models thanks to its long-running partnership. Even with discounted pricing, though, AI inference is expensive when you’re serving hundreds of millions of users across Office, GitHub, Teams, and Azure.
That seems to be changing Microsoft’s priorities.
“We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost,” Microsoft AI chief Mustafa Suleyman said in an interview with Bloomberg.
The economics aren’t exactly unique to Microsoft. TechCrunch notes that companies, including Amazon, Uber, Meta, and Accenture, have also been looking for ways to rein in AI spending.
Earlier this year, some organizations even explored lower-cost Chinese models despite security concerns, illustrating just how much attention AI operating costs are getting.
Microsoft still relies heavily on OpenAI and other models
This doesn’t mean Microsoft is completely walking away from OpenAI or Anthropic.
Bloomberg says that the company’s own models currently handle only a small percentage of requests, and Microsoft continues to rely heavily on third-party frontier models where they make sense.
Still, the direction is hard to miss. Rather than depending entirely on outside model providers, Microsoft is steadily building technology it controls itself.
Owning more of the AI stack changes the economics
If its in-house models can deliver comparable results for routine workloads at a lower cost, that’s a pretty huge financial advantage – one that compounds quickly across products used by hundreds of millions of people.
It’s also another reminder that the AI race isn’t just about building the smartest model anymore. Companies are increasingly asking a different question: “can we afford to run this at the scale our customers expect?”
For organizations deploying AI broadly across everyday productivity tools, that may prove just as important.
OpenAI recently launched its first formal partner network to expand the adoption of its AI models and services. Microsoft’s decision to use more of its own models inside products like Excel and Outlook shows that even close partners are looking for ways to reduce AI costs. Read more here.