Microsoft is in a unique position as AI overtakes the tech industry. It is one of the world’s largest cloud providers and software-as-a-service companies, while also owning valuable stakes in the two largest artificial intelligence labs, OpenAI and Anthropic.

Those incentives are starting to collide as Microsoft releases blockbuster financial results. The company just reported an extremely profitable quarter with $90 billion in revenue and net income of $35.8 billion. For the fiscal year, which ended June 30, Microsoft reported $331.8 billion in revenue with net income of $133.7 billion for the year.

And CEO Satya Nadella isn’t about to let the trajectory of Anthropic and OpenAI, which are expanding into agent applications and infrastructure that could ultimately enable them to have relationships with customers, derail that amount of cash.

Nadella has been preaching to companies to use multiple models and stop relying on cutting-edge AI labs for the agent harness/application layer.

Doing so is dangerous, he has been saying, because it requires companies to share too many of their internal secrets with manufacturers of models of dubious reliability. He knows his clients. Enterprise IT fears both data leaks and vendor lock-in.

He has now openly told Wall Street analysts during the company’s quarterly conference call on Wednesday that this is an opportunity for Microsoft to sell customers its own on-premises models, along with agents, AI security and more, while promising lower costs.

In other words, it is presenting Microsoft as an alternative to many of the proprietary services that OpenAI and Anthropic are developing for their own growth.

When UBS analyst Karl Keirstead specifically asked Nadella to weigh in on the open-source versus closed-source debate roiling the AI ​​industry, and how Microsoft will benefit from it, Nadella came out on top.

“The goal is for the company to be in control of its own destiny,” the CEO said of the companies. “We are very, very clear about the type of architectural design of the platform, which is to keep the harness separate from the model… that means that any model at any given time is interchangeable.”

Microsoft, of course, also sells a menu of harnesses (aka AI agents) under the Copilot name, including its GitHub Copilot coding agent. Encoder agents are where much of the AI ​​money is spent today.

And he used last week’s notorious incident as evidence of his warnings.

“If you look even at the Hugging Face incident, the most important thing to take away from it is that you can’t rely on any particular model,” Nadella said. “You may need multiple models even to remedy some challenges caused by one model. That’s the way to think about it, right? That is, you can’t be subject to rejection of one model.”

The incident involved an unreleased OpenAI model breaking out of its testing environment and successfully mounting a large-scale hack on Hugging Face, all in pursuit of beating a benchmark. To try to understand what happened, Hugging Face initially tried to use a private boundary model (which he has not named) which refused to help him. So it turned to the Chinese open source model Z.ai GLM 5.2 to analyze logs and protect its infrastructure. The incident has shocked the industry so much that even Sam Altman now says that perhaps AI development should slow down a bit.

Nadella also made it clear that Microsoft is happily selling its own local models, the MAI family, on its own AI chips, Maya, and presenting them as cheaper alternatives.

“Every customer wants the right model for each task based on quality, latency, cost and compliance. We offer the largest model catalog in the cloud with over 11,000 models, including leaders from OpenAI, Anthropic, Mistral, xAI, as well as our own MAI family,” he said.

He added: “We are also accelerating the development of our own model. We announced more than a dozen new models in imaging, voice, transcription, encryption and security, including our first reasoning model, MAI, all with cost-effective inference at the core for enterprise use cases. We are co-engineering these models with our silicon, and we are seeing 40% better performance per watt when running MAI models on Maya 200.”

As for the Myths? Nadella pointed to Microsoft’s new Mythos competitor announced earlier this week, MAI Cyber ​​One Flash. “It achieves better performance than the much larger Mythos model, but at half the cost when combined with our multi-agent safety harness,” he said.

Sure, Microsoft’s CEO says companies should use the frontier models offered by OpenAI and Anthropic in combination. But their most important message is: don’t trust them enough to depend on them.