Microsoft builds its own AI models to cut inference costs

Microsoft unveiled its own AI models on 2 June with two stated aims: lessen its reliance on OpenAI, and lower costs for developers.

Those aims are the same aim. Buying inference from a partner is a cost line, and building your own is how you stop paying it, CNBC reported.

For the company that made the largest bet on OpenAI of anyone, that’s a meaningful shift in posture even if the models themselves are unremarkable.

Microsoft in-house AI models are a margin decision

Copilot runs at enormous volume across Office, Windows and GitHub. At that scale a few cents per thousand requests decides whether the product is profitable.

You do not need a better model to improve a margin. You need an adequate one you are not renting.

The follow-on move fits that reading. On 27 July Microsoft promoted a cost-saving model for cybersecurity, pitching economics rather than capability.

Two launches, both led with cost. That is what a company does when the capability question is settled enough that price is the remaining variable.

Parity made Microsoft in-house AI models possible

Building your own models only makes sense if adequate is achievable, and the evidence says it now is.

Six labs sit within 79 Elo points at the top of the ratings, which our piece on what parity changes works through. The gap between best and good enough has closed for most tasks.

Open weights make it cheaper still. Meta published a 30B model for consumer hardware in August, described by TechCrunch as an open version of its strongest closed system, and a company with Microsoft’s resources can fine-tune from that base rather than train from scratch.

The consumer side is under the same pressure

Cost discipline showed up in the product line as well. Microsoft merged its Copilot apps and moved Deep Research behind a paywall this month, killing several features while doing so.

The pattern extends across the industry. Google shipped an entry-level model aimed at coding and agents three weeks after the last one, because the cheap tier is where volume lives.

Read those together and one story emerges: reduce what each request costs, and charge for the requests that cost the most. That’s a company optimising a business rather than chasing a frontier.

It also sets up an awkward relationship. Microsoft remains OpenAI’s largest backer while building the thing that would let it buy less, and both facts are true simultaneously.

Watch which model actually serves Copilot requests over the next year, since that routing decision, rather than any announcement, is where the reliance either falls or doesn’t.