Here’s something I didn’t expect to write this week. The CEO of Microsoft, the company that has poured billions into OpenAI and Anthropic, just publicly warned businesses to be careful about using proprietary AI models. Not a competitor. Not a regulator. The guy whose company sells more enterprise AI than anyone on the planet.
Satya Nadella published a blog post on Sunday arguing that companies using AI are paying twice. Once with money for tokens, and again with something more valuable: the proprietary knowledge they must feed the model to make it useful. That second payment is the one nobody notices, and it’s the one that should worry you.
I’ve been thinking about this since I read it, and honestly, the timing makes it even stranger. Allow me to explain.
The Trojan Horse Fear, Explained
There’s a worry that has been building in Silicon Valley for months. The fear is that big AI labs selling proprietary models act like Trojan horses. As startups and enterprises pump their sensitive business data into models from OpenAI and Anthropic, the labs gain deeper and deeper access to how those businesses actually work. And what stops them from eventually competing with their customers?
People like VC Jason Calacanis and Palantir CEO Alex Karp have been raising this concern for some time. Now Nadella has joined them, which changes the weight of the argument entirely.
His core claim, laid out in his post on what he calls the reverse information paradox, is that the better you want a model to perform, the more institutional knowledge you have to reveal to it. Prompts, tool usage, and especially corrections. Every time an employee fixes the model’s wrong answer, that correction distills years of company know-how into a training signal. It’s the kind of knowledge a competitor could never buy. And enterprises are handing it over for free.
The Distillation Double Standard
This is the part of Nadella’s argument I find most compelling, and it’s a genuinely spicy point coming from him.
AI labs trained their models by scraping the public internet under fair use. Fine. But those same labs then impose restrictive terms preventing anyone from “distilling” their models, meaning studying a model’s outputs to train a cheaper competing one. Nadella calls out the irony directly. You can’t build your business on freely learning from the world’s data and then forbid the world from learning from you.
For context, Anthropic accused Chinese open-source models back in February of sending millions of prompts to Claude to improve their models and urged the US government to tighten export controls. So the labs clearly consider distillation a threat when it’s done to them.
Nadella’s position: if scraping is fair, distillation should be too. I don’t think the labs will ever agree, but as a logical argument, it’s hard to knock down. It really is.
The Timing Is the Real Story
Now here’s where it gets interesting, at least to me. Nadella published this warning days after Microsoft launched its own $2.5 billion Frontier consulting arm to help enterprises deploy AI, which I covered in my breakdown of the AI implementation gold rush. Anthropic has Ode. OpenAI has The Deployment Company. Everyone wants to get deeper inside your business.
So read the two moves together. Microsoft is simultaneously selling enterprises deployment help and warning them not to trust the model makers with their data. Contradiction? Not really. It’s positioning.
Nadella’s proposed solution is for companies to retain ownership of their data by building proprietary learning environments in the cloud, and to add orchestration layers that let them switch between models instead of locking into one provider. Guess who sells cloud infrastructure and model-switching tooling. Azure benefits either way, whether you pick OpenAI, Anthropic, or an open source model, as long as you run it on Microsoft’s cloud.
Sure, the advice is self-serving. There’s no denying that. However, self-serving advice can still be correct.
The Open Source Shift Is Already Happening
Whether or not enterprises listen to Nadella, the market is already moving in the direction he’s pointing.
Large companies are increasingly taking open-source models and running them on their infrastructure. Idit Levine, CEO of Solo.io, whose customers include T-Mobile, ADP, and SAP, told TechCrunch her clients keep reaching the same conclusion: an open model running on-prem does almost 90% of what the big proprietary ones do, at way less cost, with full control. TechCrunch’s full report has more of her thinking, and it lines up with what the traffic data shows.
Open models accounted for 29% of all traffic routed through Vercel’s AI gateway last month, and OpenRouter is seeing a similar surge. That’s not a niche experiment anymore. That’s a real share of production workloads.
The pattern reminds me of the early cloud era in reverse. Back then everyone rushed off-prem. Now, for AI specifically, the smart money is quietly building the option to come back.
What I’d Actually Do With This
If you run a business using AI, my honest take is that you don’t need to panic, but you do need a plan. Three things worth doing now:
Read your provider’s data terms. Nadella is especially concerned about clauses where model makers reserve the right to learn from customer usage and interaction data. Find out what yours says. Most people never have, the same way most companies never audited their AI in cybersecurity posture until something forced them to.
Avoid single-vendor lock-in. Orchestration layers and AI gateways are cheap insurance. Even if you never switch models, the ability to switch changes your negotiating position.
Treat your corrections as an asset. The feedback your team gives an AI model is institutional knowledge. Know where it’s going and who owns it.
Nadella ended his post with a line that stuck with me: “In consuming intelligence, you are creating intelligence.” Whatever you think of his motives, the question of who owns what you create is about to define the next phase of enterprise AI. Better to answer it on your own terms.
