Nadella: you pay for AI twice — with money and your company's data

On Sunday, July 13, 2026, Satya Nadella, CEO of Microsoft, published a post on his personal blog titled "Reverse Information Paradox," making a claim that is uncomfortable for the entire AI industry: companies pay for intelligence twice — once in cash for tokens, and a second time in institutional knowledge they hand to the model in order to make it useful. According to Nadella, the second payment is more expensive and more dangerous.

Key takeaways

Nadella argues every prompt sent to an external AI model carries a hidden payment in the form of company data
Models learn from "exhaust" — prompts, agent tool calls, corrections — absorbing customers' institutional knowledge
Nadella's solution: proprietary learning environments, model distillationdistillation: Training a new, often cheaper model on the outputs of an existing one. rights, and orchestration layers enabling model-switching
In June 2026, open-source models accounted for 29% of traffic through Vercel's AI gateway — a signal of a growing shift toward owned models
The post targets AI lab hypocrisy: they train freely on public data while restricting customers from distilling their own models

"You pay for intelligence twice"

The central metaphor in the post is the "reverse information paradox": the more you want from a model, the more you have to tell it about yourself. What you tell it becomes part of its knowledge — not yours. The post was published by Satya Nadella on his personal blog snscratchpad.com.

Models learn from exhaust — the prompts people write, the tools agents use, and especially the corrections people make when the model is wrong. Every correction is distilled into institutional know-how. This is the kind of knowledge a competitor could never buy. And yet enterprises are handing it over.

Satya Nadella, snscratchpad.com, July 13, 2026

Model distillation and lab hypocrisy

A key part of the post is the argument about distillation — the technique of training a new, often cheaper model on the outputs of an existing one. The Microsoft CEO points out that AI labs train on public data themselves, relying on fair usefair use: A US legal doctrine allowing limited use of protected material without the owner’s permission. doctrine. But they simultaneously attempt to prohibit customers from distilling their models.

In February 2026, Anthropic accused Chinese open-source labs of sending millions of prompts to Claude to absorb its capabilities. Nadella cites this precedent to argue that if model vendors can freely use the world's data, customers should have the right to distill models they pay for.

The solution: own infrastructure, orchestration layers

Nadella's practical recommendation is not neutral. The cloud giant's CEO suggests companies build proprietary learning environments, implement orchestration layers enabling model-switching, and avoid lock-inlock-in: Dependence on a single vendor that makes switching to another costly or hard.. Tools like OpenRouter and Vercel AI Gateway, which enable routing requests across models, are gaining traction.

Idit Levine, CEO of Solo.io — which makes networking and security software for enterprise AI — confirmed this trend to TechCrunch. Her clients, after initial experiments with proprietary models, increasingly ask about open-source models on-prem. Her company serves SAP, T-Mobile, and ADP.

Why this matters

Nadella's voice carries specific weight. He is the CEO of a company that is both one of the largest investors in OpenAI and Anthropic, and a cloud provider those models run on. When he warns companies to be careful about handing institutional knowledge to external models, the warning carries market authority. The on-prem open-source model trend is already visible: in June 2026, open-source models accounted for 29% of traffic through Vercel's AI gateways — and that number is growing. Nadella's post gives additional arguments to companies already considering a move away from closed platforms.

What's next

On-prem model installation by enterprise will accelerate — according to Solo.io and Vercel, this shift is already visible in customer inquiries from Q2 2026 onward
Orchestration tools (OpenRouter, Vercel AI Gateway, LiteLLM) may see accelerated adoption now that lock-in arguments gain voices of authority
Key to watch: whether OpenAI and Anthropic modify their distillation terms in response to growing pressure from customers and regulators