INNOVATING TOGETHER
Meet MAI - Microsoft's replacement
Microsoft invested $13B in OpenAI. Now it's replacing OpenAI models in Excel and Outlook with its own.
ALL NEWSAI & STARTUPS
Khanlar Alizada
7/27/2026
The answer is on the cap table
Microsoft has put around $13 billion into OpenAI. After the October 2025 restructuring, it holds roughly 27% of OpenAI Group PBC, valued at about $135 billion at recapitalisation. Its IP rights to OpenAI's models run through 2032. And OpenAI committed to buying an incremental $250 billion of Azure.
Now trace the money.
When Microsoft pays OpenAI, a meaningful share of that value returns — as equity appreciation in an asset it part-owns, and as cloud revenue on infrastructure it operates. The cheque leaves and much of it comes home.
When Microsoft pays Anthropic, the cheque just leaves.
That is the whole quote. It isn't a verdict on model quality — it's a verdict on ownership structure. Suleyman named the vendor Microsoft has no equity in, and left unnamed the vendor whose invoice is partly a transfer between pockets.
If you're building an AI company and a hyperscaler is your largest customer, that distinction is the most important sentence in this story.
What Microsoft has actually built
This isn't a single model. At Build, Microsoft's newly formed Superintelligence Team unveiled seven models built from scratch, spanning reasoning, coding, transcription, voice, image generation and image editing. Three were named: MAI-Thinking-1 for multi-step reasoning, MAI-Code-1-Flash, a 5-billion-parameter coding model wired into VS Code and GitHub Copilot, and MAI-Image-2.5.
Microsoft has claimed one of these matches the coding capability of Anthropic's Opus 4.6 at lower cost. Take vendor benchmarks with the usual salt — but the claim itself is the signal. You don't make that comparison publicly unless you intend to act on it.
Suleyman's framing: "long term self-sufficiency for Microsoft and our partners." And: "these are all the basic building blocks of a superintelligence."
The part your timeline is missing
The standard story goes: rely on external models to ship fast → study the use cases → build your own. That's true, and it's the right general lesson.
But it skips the mechanism. Microsoft didn't simply decide to do this. It was contractually freed to.
The October 2025 restructuring made Microsoft's license to OpenAI's IP non-exclusive and stripped Microsoft of its right of first refusal on OpenAI's compute. Both companies got looser. Suleyman has since said Microsoft was "set free"from OpenAI to pursue superintelligence, and the MAI Superintelligence Team was formed in the aftermath.
So the real timeline is: contract renegotiated in October 2025 → seven in-house models and live traffic in Excel by July 2026. Nine months from permission to production.
That's the number that should worry model vendors. Not that platforms want to internalise — everyone knew that. But that once the legal handcuffs came off, it took Microsoft under a year to ship credible substitutes into its flagship products.
Where I'd temper the panic
One reality check, because most coverage is skipping it.
"Tens of thousands of prompts per week" is a very small number for Microsoft. Microsoft 365 serves hundreds of millions of users. Tens of thousands of weekly prompts is a canary deployment — a controlled test on low-risk, high-volume, low-complexity tasks. Summarising an email is not where model quality gets stress-tested.
Microsoft is still buying from Anthropic. It is still deeply entangled with OpenAI through 2032. What we're watching is direction, not displacement.
And the direction is unambiguous: the commodity tier goes in-house first. Simple, repetitive, well-understood tasks migrate to the cheap internal model. The hard reasoning stays outsourced — for now.
The lesson if you're building
Being a model vendor to a company with tens of billions in annual capex is a rented position, and the rent comes due the moment your use case becomes legible and high-volume.
The defensible layer was never the model. It's distribution, or ownership of the workflow, or a proprietary data loop the platform can't reconstruct. Everything else is a line item on someone's cost-reduction roadmap.
Question for you: if Microsoft can rebuild "good enough for email summaries" in nine months, what's the actual half-life of a model-layer moat? I'd argue it's shorter than most 2026 valuations assume — but I'd like to be argued out of it.
Read Mustafa Suleyman's (CEO of Microsoft AI) Quote Again. He Named Anthropic, Not OpenAI.
"We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost."
Microsoft's AI chief said that in June. By July, Bloomberg reported that tens of thousands of prompts a week inside Excel and Outlook were being served by Microsoft's own MAI models instead of third-party ones. Summarising email. Drafting replies. Formatting spreadsheets. No announcement, no keynote.
Most of the commentary has framed this as Microsoft turns on its AI partners. I think the more precise question is the one the quote raises by itself: why Anthropic, and not OpenAI?
