Satya Nadella, Microsoft's chief executive, published a blog post this week warning enterprises that they are effectively paying for proprietary AI models twice: once upfront in costs, and again by giving away their own institutional knowledge to the vendor TechCrunch. The post appeared on Nadella's personal website, snscratchpad.com, under the title "The Reverse Information Paradox" TechCrunch.

His argument hinges on what he calls model "exhaust" — the data that accumulates when a company uses an AI model. Every time an organization runs a workload against a proprietary model, it generates a trail: the questions asked (prompts), the instructions given to the AI (agentic tool calls), and corrections made when the model gets something wrong. According to Nadella, this exhaust is not thrown away. Instead, vendors use it to improve their next version of the model. That means a company's internal pricing logic, workflow shortcuts, and domain expertise — the things that made it competitive — end up in a form that other customers, including direct competitors, can benefit from TechCrunch.

Nadella frames this as structurally unfair. Proprietary model makers train their systems on public internet data with minimal restrictions, yet turn around and impose strict licensing terms on their own models that forbid customers from training their own competing versions. The result, he argues, flows value in one direction only: toward the model maker TechCrunch.

His remedy has two parts. First, companies should retain ownership of all their own interaction data — prompts, corrections, feedback — rather than letting it vanish into a vendor's training process. Second, they should build what he calls proprietary learning environments on cloud infrastructure, complete with orchestration layers (essentially, software that sits between a company and multiple AI models) that allow an organization to switch between model providers without rebuilding its entire stack or losing the value of its own accumulated data.

Nadella also published a piece in the Wall Street Journal titled "We Can't Let AI Giants Eat the Economy," which takes the argument into the policy realm, framing control of the most advanced AI models as a risk to competitive markets WSJ.

It is worth flagging that Nadella runs Microsoft, a company that sells exactly the kind of cloud infrastructure and orchestration tools he is recommending enterprises build. Microsoft also holds a major commercial partnership with OpenAI, one of the proprietary model vendors his post implicitly criticizes. The post should be read as both a genuine technical warning and a strategic move to position Microsoft as the platform for customers who want to avoid dependence on a single model vendor. Microsoft stands to benefit commercially if large enterprises heed this advice and build their switching infrastructure on Azure rather than locking themselves into one provider.

That caveat said, the technical concern underneath the positioning is real and worth taking seriously. Companies running complex AI workflows at scale do generate enormous amounts of interaction data — tool invocations, corrections, retrieval feedback — and most commercial AI contracts have been vague about where that data lives, who owns it, and whether it quietly improves the vendor's next model. Data rights and training obligations have lagged far behind the speed at which enterprises have adopted AI, and Nadella's framing gives procurement and legal teams language for a negotiation they have already been having informally for the past two years.

The proposed solution — orchestration layers that allow companies to swap between models — is not a new architectural idea. Enterprises built similar multi-cloud abstraction layers a decade ago to avoid being locked into a single cloud provider like AWS, Azure, or Google Cloud. Whether the same approach works smoothly for AI models is less clear. Switching between different AI models can mean losing months of fine-tuning work (customization specific to one model's behavior), abandoning carefully engineered prompts optimized for that model's quirks, and rewriting test suites built around a particular vendor's output. Orchestration layers reduce contractual lock-in but do not eliminate the technical lock-in that comes from sustained iteration against one model's specific behavior.

What Nadella's post does accomplish is to hand enterprise buyers a new phrase — "pay twice" — for a conversation they were already having with AI vendors about data ownership and whether they can opt out of being included in model training. Whether that phrase shifts contract terms across the industry, or remains a rhetorical move in the ongoing competition between cloud providers and frontier-model companies, will depend on how the largest enterprise customers — banks, hospitals, manufacturers — use it in their next vendor negotiations.