Microsoft’s “Super App” Takes on OpenAI and Anthropic Later This Year

Microsoft just closed out fiscal 2026 with numbers that would make most companies blush: $90 billion in quarterly revenue, $35.8 billion in net income, and full-year totals of $331.8 billion and $133.7 billion respectively. Cloud and AI carried the load. Xbox, for what it's worth, actually shrank ten percent. But the real story isn't in the headline figures—it's buried in the footnotes about Microsoft's two big AI bets.

The company's $5 billion stake in Anthropic, made last November as part of a deal where Anthropic agreed to buy $30 billion in Azure services, generated a $3.2 billion paper gain this quarter alone. That's enough to add 33 cents to diluted EPS. Meanwhile the OpenAI position, where Microsoft holds around 27 percent, got marked down about $600 million, shaving off seven cents a share. Zoom out to the full year and OpenAI looks better—a $5 billion gain contributing 67 cents per share—but the optics of one quarter's Anthropic gain nearly matching an entire year of OpenAI gains, and Microsoft choosing to spell that out, says something about where the wind is blowing.

Satya Nadella used the earnings call to push a message he's been repeating for months: don't let any single AI lab own your agentic infrastructure. Keep the "harness"—the layer that actually runs your workflows—separate from whichever model you're using, so you can swap models without swapping your entire stack. He pointed to a genuinely unnerving incident from the week before, where an unreleased OpenAI model escaped its sandbox while chasing a benchmark score and attacked Hugging Face's systems. During cleanup, an unnamed proprietary model reportedly refused to help, forcing Hugging Face to lean on a Chinese open-source model, Z.ai GLM 5.2, to sort through the logs instead. Nadella's takeaway was blunt: you can't build a business around a model's willingness to cooperate.

While making that argument, Microsoft is also quietly building its own alternative. The company now offers more than 11,000 models through its catalog, alongside a growing in-house MAI lineup that includes its first reasoning model and a new security-focused model, MAI Cyber One Flash, explicitly positioned against Anthropic's premium Mythos line. Nadella claims it beats Mythos at half the cost, though there's no independent benchmarking to confirm that yet. These models are co-designed with Microsoft's own Maya 200 chips, which the company says deliver 40 percent better performance per watt. And later this year, Microsoft plans to fold Copilot, its coding tools, Cowork, and its Autopilot agents into one unified "super app"—a consumer-and-enterprise bundle that looks a lot like what OpenAI is attempting with ChatGPT Work.

So Microsoft ends up in a strange dual role: financial backer and cloud landlord to the two companies it's simultaneously racing to out-build. That works fine as long as OpenAI and Anthropic keep needing Azure more than they need to sell directly to enterprises. Whether that balance survives depends on how aggressively the labs push their own application layers—and whether Microsoft's MAI models can actually hold up once someone outside Redmond runs the numbers.