Microsoft's AI chief Mustafa Suleyman gathered staff at an internal town hall this week and made a remarkable claim: the company had quietly trained a model that beats the one powering OpenAI's GPT-5.6, while using roughly a quarter of the computing power OpenAI spent to get there. For a company that is OpenAI's largest backer and biggest cloud landlord, that is an awkward thing to boast about. The market's reaction told its own story. Microsoft shares ticked up about one percent, then slipped after hours. For a four-trillion-dollar company, that is a shrug.
The model is MAI-Base-2, a mixture-of-experts design that activates only part of its network for any given task, with 80 billion active parameters out of 2.8 trillion total. Microsoft says it was trained on 16,000 of Nvidia's GB200 chips over 38 days in August and September. The company put hard numbers on efficiency rather than quality: it claims the model is four to twelve times more efficient in pre-training computation than DeepSeek's best open base model. The eye-catching comparison to OpenAI, though, rests entirely on Suleyman's reported remarks and private tests. There are no public benchmarks, no release date, no pricing. OpenAI has neither confirmed nor challenged it, and the man making the claim runs the unit being judged.
So treat the OpenAI line with caution, because the efficiency figure is the part that actually matters. Microsoft spent nearly 116 billion dollars on capital projects in its last fiscal year, most of it on AI infrastructure. If it can reach frontier-grade quality on a fraction of the compute, every one of those dollars stretches further, buying more training runs and more customers. There is a second, quieter saving inside Copilot: every query Microsoft answers with its own model is one it does not pay a partner to handle. On a July earnings call, finance chief Amy Hood called model diversification "a margin improvement opportunity," which is corporate for exactly this.
That framing comes straight from the top. Satya Nadella set the direction in July when he said "every model is substitutable," and MAI-Base-2 is that philosophy made concrete. Microsoft has spent years as OpenAI's dependent partner, reselling its models on Azure and wiring them into Copilot. Building a credible in-house alternative does not end that relationship, but it changes the balance of power inside it. Microsoft still holds a stake in OpenAI reportedly worth 135 billion dollars and has rights to its technology through 2032. It no longer needs OpenAI to be the only option on the shelf.
The plan is to license MAI-Base-2's weights to "trusted partners" through Azure's Foundry platform, so Microsoft collects revenue whichever model a customer picks. That is the logic of a company hedging: keep the lucrative OpenAI partnership, keep selling rival models to anyone who wants them, and keep a home-grown option in reserve that costs less to run. The claim that it beat OpenAI may or may not hold up once the model finishes post-training and faces public scrutiny. The strategy underneath it, owning the floor instead of betting on any single model, is the part Microsoft clearly means.