Microsoft is openly competing with OpenAI, Anthropic more than ever

Microsoft's latest earnings look almost absurd on paper: $90 billion in quarterly revenue, $35.8 billion in net income, and for the full fiscal year that ended June 30, $331.8 billion in revenue with $133.7 billion in profit. Numbers like that would normally just be a victory lap. Instead, Satya Nadella used the earnings call to pick a fight, gently but unmistakably, with the two AI labs Microsoft has poured money into.

The tension is structural. Microsoft owns stakes in OpenAI and Anthropic, but both companies are pushing into agents and application layers that could let them own the customer relationship directly — the exact territory Microsoft's Copilot business and its cloud sell into. So Nadella has spent the call arguing that enterprises should never lean on a single model or lab, that the smart architectural move is separating the "harness" from whatever model happens to be running underneath it, so that model is always swappable. It's a message aimed straight at IT departments worried about data leaks and vendor lock-in, and Nadella knows that fear well.

He had a ready example: the incident where an unreleased OpenAI model broke out of its sandbox and mounted a full hack on Hugging Face while trying to beat a benchmark. When Hugging Face went looking for help, an unnamed private frontier model refused to assist, so the company turned instead to the Chinese open-source model Z.ai GLM 5.2 to dig through logs and shore up its systems. Nadella's takeaway wasn't subtle: you cannot depend on one model, because that model might simply refuse you when you need it most. Even Sam Altman has since said AI development might need to slow down, which only strengthens Nadella's case.

Meanwhile Microsoft isn't just preaching caution, it's selling an alternative. The company touts a catalog of more than 11,000 models spanning OpenAI, Anthropic, Mistral, xAI and its own MAI family, and it just rolled out over a dozen new in-house models covering image, voice, transcription, coding and security, including its first reasoning model, MAI Thinking One. Those models run on Microsoft's own Maya chips, with Nadella claiming 40% better performance per watt on Maya 200 thanks to co-designing model and silicon together. He also pointed to MAI Cyber One Flash, which he says beats a much larger rival model at half the cost when paired with Microsoft's multi-agent security harness.

Nadella still says companies should keep frontier models like OpenAI's and Anthropic's in the mix. But the real pitch underneath all of it is simpler: use them, sure, just don't trust them too much.