Decisions

Microsoft Started Swapping Out the Lab It Funded

On July 23 Microsoft began replacing OpenAI's image models with its own MAI-Image-2 in PowerPoint and Bing. The build-versus-buy lesson for anyone whose product depends on a single AI vendor.

DecisionsBy Servola Tech Desk2026-07-243 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Microsoft Started Swapping Out the Lab It Funded

The decision in front of everyone

The company that put roughly 13 billion dollars into OpenAI is now taking OpenAI's product out of its own flagship apps. On July 23 Microsoft confirmed it has started replacing OpenAI's image-generation models with its own, called MAI-Image-2, in PowerPoint and Bing. Mustafa Suleyman, who runs Microsoft AI, put it flatly: the in-house model is faster, cheaper, higher quality, and drives better retention. Microsoft says MAI-Image-2 generates images about twice as fast as its predecessor on its Foundry and Copilot services.

It does not stop at images. Suleyman has said openly that Microsoft wants to reduce, and eventually eliminate, what it spends with Anthropic by moving more work onto its own MAI models across Copilot, Microsoft 365 and GitHub Copilot. The most important AI partnership of the decade is quietly being routed around by one of its own architects.

Why this is a decision story, not a product story

Every deep vendor relationship is still a build-versus-buy line, and it flips the instant the buyer's volume justifies building. Microsoft bought when it needed speed and had no model of its own. It is building now because it processes so many AI tokens that even a modest per-request saving becomes an enormous absolute number. Nothing about the closeness of the OpenAI relationship changed that arithmetic; it only delayed it.

That is the lesson worth taking. If Microsoft, with a front-row seat and billions invested, treats the model layer as swappable the moment the numbers say so, then any company betting its product on a single model vendor should assume the same move is coming for it, either from its supplier or, one day, by its own hand. Loyalty did not protect the arrangement. Volume decided it.

Read where the swap started

Commoditisation arrives where it is cheap to be wrong. Microsoft did not begin by replacing the reasoning model behind its hardest coding tasks. It began with image generation in PowerPoint and Bing, where a slightly worse picture costs almost nothing and the request volume is vast. That is the classic path: prove the in-house model on forgiving, high-volume work, bank the savings, then move up the difficulty curve as confidence grows.

For an operator that pattern is a map. The parts of your stack most exposed to a supplier swap, by you or by your vendor, are the high-volume, error-tolerant ones. The parts that stay locked to a specialist longest are the low-volume, high-stakes ones where a wrong answer is expensive. Knowing which of your workloads sits where tells you where to invest in portability first.

What to do before your vendor does

Design as if you will switch, because the market now assumes you might. Put every model call behind an internal interface so a provider change is a setting, not a project. Keep your prompts, your evaluation sets and your guardrails in your own repository, independent of any one vendor's tooling, so you can re-run them against a challenger in an afternoon.

Then negotiate accordingly. A supplier that knows you can leave in a week prices differently from one that knows you cannot. And do the arithmetic Microsoft did: track the volume threshold at which building or fine-tuning your own model beats renting, so you make that call deliberately rather than discovering, as OpenAI just did, that your biggest customer quietly reached it first.

Frequently asked questions

Is Microsoft cutting OpenAI off entirely?

No. It is replacing specific OpenAI models, starting with image generation in PowerPoint and Bing, with its own MAI models, and says it wants to reduce its spend with Anthropic too. It is reducing dependence on outside models, not ending the OpenAI relationship overnight.

Why does a company's internal model swap matter to me?

Because it is the clearest signal yet that the model layer is becoming a swappable, price-driven component. If your product depends on one AI vendor, the same build-versus-buy logic that moved Microsoft will eventually reach you.

What is the single most useful action here?

Put your model calls behind your own interface and keep your prompts and evaluations portable. That one step turns a provider switch from a rewrite into a configuration change, which is what gives you both leverage and safety.

The headline is that Microsoft is building its way around OpenAI. The instruction underneath it is simpler. No vendor relationship, however large or however friendly, is exempt from the build-versus-buy calculation, so the operators who stay free are the ones who designed for the swap before the numbers forced it.

MicrosoftOpenAIMustafa SuleymanMAIVendor lock-inBuild vs buyAI strategy

One Commissioner Stalled a 2 Gigawatt Campus

New Mexico's land commissioner has twice refused Oracle's Project Jupiter a pipeline across state trust land, and the air permit hearing is set for 19 October. Going off-grid did not escape the queue, it changed which official holds the pen.

4 min read

Decisions

The Warning Shipped 14 Days Before the Deletions

OpenAI rated Sol's deletion behaviour Severity 3 on its own safety hub on 26 June, fourteen days before launch. For an owner, that is a liability event.

5 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles