Nadella sets a course to compete with OpenAI and Anthropic

Microsoft finds itself in an unusual position: it is at once one of the largest cloud providers and a stakeholder in two leading AI labs — OpenAI and Anthropic. Now those roles are starting to collide.

The occasion was the latest earnings report. For the quarter, the company posted $90 billion in revenue and $35.8 billion in net profit, and for the fiscal year that ended June 30, $331.8 billion in revenue on $133.7 billion in profit. To protect that trajectory, Satya Nadella has no intention of letting OpenAI and Anthropic hijack customer relationships as they expand toward applications and agentic infrastructure.

“Keep the harness separate from the model”

On the call with analysts, Nadella laid out an architectural principle: the agentic “harness” should be kept separate from the model itself, so that any model can be swapped out at any moment. The point is that an enterprise should remain master of its own fate rather than sharing its internal secrets with model makers of dubious reliability. Nadella knows the fears of corporate IT departments well: data leaks and lock-in to a single vendor.

Responding to a UBS analyst's question about the open-versus-closed-source debate, Microsoft's chief, in the reporter's words, “came out swinging.” As his argument he cited the much-discussed Hugging Face incident: an unreleased OpenAI model broke out of its sandbox and mounted a full-blown attack on the platform's infrastructure for the sake of a benchmark result. In investigating what happened, Hugging Face first turned to a certain closed frontier model, but it refused to help, and in the end the logs were analyzed by the Chinese open model Z.ai GLM 5.2. According to Nadella, the takeaway is clear: you can't depend on a single model — sometimes fixing a problem created by one model requires another. The story rattled the industry so badly that even Sam Altman started talking about slowing down AI development.

Its own models, its own chips

Microsoft is eager to sell its own MAI family of models on its own Maya chips, positioning them as a cheaper alternative. Nadella pointed to the “broadest catalog” — over 11,000 models, including flagships from OpenAI, Anthropic, Mistral, xAI and the MAI lineup. The company announced more than a dozen new models — for images, voice, transcription, code and security — as well as its first reasoning model, MAI Thinking One. According to him, running MAI on the Maya 200 chip delivers 40% more performance per watt.

There's an answer to Mythos, too: MAI Cyber One Flash, unveiled this week, supposedly outperforms the far larger Mythos model at half the cost when paired with a multi-agent security harness.

Formally, Nadella agrees that the frontier models from OpenAI and Anthropic are worth keeping in the overall mix. But the main message reads differently: don't trust them enough to rely on them.

Source: TechCrunch