Enterprise

Microsoft's $90B quarter comes with a warning shot at OpenAI and Anthropic

Nadella used FY26 Q4 earnings to pitch Microsoft as the neutral harness layer for enterprise AI — and pointed at MAI-Cyber-1-Flash as proof it can undercut its own partners.

By Asher Glick · Enterprise software desk · 2026-08-01 · 3 min read

Photo: Wikimedia Commons / Microsoft, CC BY-SA 4.0 — Microsoft CEO Satya Nadella in a portrait photograph

Microsoft closed fiscal Q4 on July 29, 2026 with $90 billion in revenue, up 18% year over year, and $35.8 billion in net income, comfortably clearing the Wall Street consensus of $87.72 billion. The full year finished at $331.8 billion in revenue and $133.7 billion in net income. But the numbers weren’t the story on the call. The story was Satya Nadella spending an earnings hour explaining, in careful language, why Microsoft’s most important AI partners are also its competitors.

Azure crossed $100 billion in annual revenue for the first time. Microsoft 365 Copilot passed 30 million paid seats, up from more than 20 million as of April. GitHub Copilot reached 50 million users. Commercial remaining performance obligations grew 8% sequentially to $678 billion. Capex and finance leases hit $41 billion for the quarter, a 69% jump year over year, the physical footprint of a company betting that inference demand doesn’t taper.

Nadella’s framing was the interesting artifact. He pitched Microsoft’s cloud as the broadest model catalog going, over 11,000 models, with more than a dozen new ones announced across image, voice, transcription, coding and security. Customers should keep the harness separate from the model, he argued, so any model stays swappable at any time. “The goal is to have the firm be in control of their own destiny,” he said. UBS analyst Karl Keirstead flagged the shift: Microsoft is now openly competing with OpenAI and Anthropic, alongside Mistral and xAI, rather than routing enterprise demand primarily through them.

The proof point Microsoft chose was MAI-Cyber-1-Flash, unveiled the same week inside MDASH. It’s designed to handle up to 90% of tasks in the security harness, with GPT-5.4 reserved for the remaining 10%, delivering roughly 50% cost savings versus the current best configuration and 40% better performance per watt when run on Maia 200 silicon. Benchmarked on CyberGym against Anthropic’s Mythos, Gemini and GPT, it’s positioned as a direct substitute for the frontier labs Microsoft still writes checks to.

The Redmond doctrine is now legible. Sell the harness, own the silicon, keep the models interchangeable, and quietly build your own for the workloads where margin matters most. It’s the same playbook IBM ran against its own software partners in the 1990s mainframe-to-middleware transition, just executed at cloud speed and with a $678 billion backlog behind it.

Sources