Satya Nadella says companies that trust one AI for everything may not survive.

On Sunday, Microsoft CEO Satya Nadella doubled down on the shocking warning he issued earlier this month about companies using AI, this time going one step further. He predicts that companies that rely entirely on proprietary AI labs for their AI needs will ultimately not survive.

He said this on CNN’s ‘Fareed Zakaria GPS’: When Zakaria asked Nadella to explain why companies share so much with AI model providers, Nadella said companies need to be careful about everything they hand over, from data to prompts.

Nadella called for a setup where “every time you use a model, all the metadata around it is maintained, so you can use all of that to train your own weights or an open model.” (Weights are the trained parameters of the model. They are essentially the brains of the model. Nadella’s point: Companies need to have their own usage data so they can ultimately build their own models.)

He added: “We argue that a company without this control cannot remain a company because it has essentially outsourced its thinking.”

Simply put, companies that don’t have their own models or an AI infrastructure layer known as an AI gateway that separates the prompts from the models themselves will be in trouble, Nadella says.

He specifically wants companies to stop relying on the AI ​​lab’s built-in coding tools, known as Harness. (Anthropic’s Claude Code and OpenAI’s ChatGPT Codex are examples of this.)

“By separating the harness from the model and the context and memory from the model, you can use multiple models for the things they are good at, while still being able to maintain control of your own destiny without one model at the same time,” Nadella said.

Microsoft is investing in the two largest AI labs, Anthropic and OpenAI. Coding agents are a particularly popular way for businesses to use AI models, and by all accounts are making model creators a ton of money.

But Nadella is telling companies not to rely too much on it. Microsoft will certainly benefit from that warning. That’s because Microsoft’s cloud business now also sells the kind of alternative infrastructure he recommends.

Despite the obvious selfish scare tactics, he’s not wrong. Companies are increasingly realizing that they need many model options, especially cheaper ones, and are turning to open models (models where the underlying code is publicly available) that can be fine-tuned and run on their own hardware. This ultimately means that we need a way to manage multiple models as well as coding agents that are not tied to a specific model provider.

But Nadella’s observations aren’t limited to runaway budgets. He predicts that once companies “outsource their thinking to models,” there will be little to prevent AI labs from eventually offering competing services of their own. These risks grow as companies adopt AI agents and give them access to the company’s inner workings.

This is the kind of warning that the startup world has been dreading for years. How can we prevent model makers from imitating and competing with startups, thereby killing them?

For example, last May, when OpenAI CEO Sam Altman offered to invest in all Y Combinator startups in the latest cohort by offering AI credits, seed investor Jason Calacanis issued a similar buyer beware message, posting: He posted.

Now Nadella is making the same case for businesses.

One caveat: Nadella’s concerns about oversharing with AI models only apply to businesses, not individuals. When Zakaria asked Nadella specifically how everyday people can protect themselves, Nadella shrugged, saying data sharing is just a price consumers pay to use services, especially free ones.

“There has to be some value exchange in the consumer space, where you’re probably getting something for free for your data. That’s how the advertising business model has worked,” Nadella said.