Satya Nadella's 'Reverse Information Paradox': Why Companies Using AI Pay Twice

On July 12, 2026, Microsoft CEO Satya Nadella published an essay titled “The Reverse Information Paradox” on his personal blog, sn scratchpad¹. Its central warning: companies using AI “pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal to make that intelligence useful.”

What makes the essay noteworthy is who wrote it. Microsoft has invested in both OpenAI and Anthropic, and is itself one of the world’s largest sellers of AI through its cloud business². Yet here is its CEO arguing, from the customer’s side of the table, that enterprises should stop letting AI providers absorb their knowledge. TechCrunch called it “a shocking warning to companies using AI”², and The Register ran the headline “Microsoft chief turns hostile on frontier AI labs, warns companies to guard their IP”³. This article walks through the essay’s argument and what it means for organizations adopting AI.

Paying for Intelligence Twice: Arrow’s Paradox, Inverted

The essay starts from a classic piece of economics: Kenneth Arrow’s “information paradox.” The value of information is unknown to a buyer until they have it — but once they have it, they have effectively acquired it without paying. Traditionally, then, it was the seller of information who bore the risk of knowledge leaking away¹.

Nadella argues that AI has flipped this structure. To get good output from an AI model, you have to pour in your own proprietary knowledge — internal procedures, decision criteria, customer context — as prompts and data. The better you want the model to perform, the more you must reveal. The leakage risk has moved from seller to buyer. That is the “reverse” information paradox¹.

And the imbalance compounds with use. As the essay puts it, “Over time, the information asymmetry becomes increasingly skewed. The seller learns more and more about you as you use what you purchased, while you learn very little about what the seller is learning in return.”¹³

“Intelligence Exhaust”: How Know-How Leaks with Every Correction

Where exactly does proprietary knowledge escape? Nadella points to what he calls “exhaust” — the byproducts of everyday use. Models learn from the prompts people write, the tools agents use, and especially the corrections people make when the model gets something wrong¹.

Corrections are the part Nadella weighs most heavily. When an employee fixes a model’s output, a person who knows the right answer is teaching the model their judgment. In the essay’s phrasing, every correction is distilled into institutional know-how¹². Distillation is the technique of extracting knowledge from one model’s outputs to train another — we explain how it works in our guide to model distillation and quantization. The result is a quiet accumulation, outside the company, of exactly the kind of practical expertise a competitor could never simply buy.

The Irony: Training on Public Data Is Fair Use, but Distillation Is Forbidden

The essay also takes a pointed shot at industry practice. Model providers have enjoyed the right to train on public data under fair use. Yet their terms of service impose restrictive conditions when customers try to learn from the models in return — that is, to distill them. Nadella calls this asymmetry ironic¹, and argues: “In consuming intelligence, you are creating intelligence. And what you create should belong to you.”¹²

TechCrunch frames this as a critique of AI companies that “reserve the right to learn from customer usage and interaction data”². If learning flows in only one direction, the essay contends, economic value concentrates with the owners of the infrastructure rather than with the people who created the knowledge¹.

The Prescription: A Trust Boundary and Five C’s

Nadella does not stop at diagnosis. The foundation of his prescription is a “trust boundary” — a hard boundary across which nothing crosses, not even the intelligence exhaust, without consent³. On top of that, the essay organizes the work into five pillars¹.

First, Control: maintain private evaluation standards and retain ownership of organizational memory, traces, and decision data. Second, Capability: build proprietary learning environments inside secure boundaries where models can be trained and tuned on real workflows. Third, Choice: decouple the orchestration layer from any single model, so the business keeps running through model transitions. Fourth, Cost: optimize by combining context, models, and tasks flexibly without sacrificing quality. Fifth, Compound: combine the four above into a continuous learning loop that belongs to the enterprise rather than the vendor, so the value of AI investment compounds over time¹.

Press coverage translates these into concrete implementations: AI gateways that can switch between model providers, and learning environments kept within a company’s own tenant boundary²³.

The Backstory: An AI Seller’s Complicated Position

The essay lands on a trajectory of change in Microsoft’s relationship with frontier AI labs. Microsoft invested billions of dollars in OpenAI, and Azure was initially ChatGPT’s exclusive cloud provider — an exclusivity that was loosened in April 2026³. Since OpenAI’s Stargate project, OpenAI has moved to secure its own infrastructure, and the distance between the two companies has been shifting — a story we have followed on this site.

Meanwhile, Microsoft has been making enterprise AI adoption a core business. On July 2 it announced the Frontier Company, a $2.5 billion unit that embeds 6,000 experts inside customer enterprises, and shortly after announced that GPT-5.6 would become the preferred model in Microsoft 365 Copilot. Embedding other companies’ frontier models in its own products while telling customers not to hand too much knowledge to model providers — the essay can also be read as a strategic statement positioning Microsoft as the guardian of customers’ data boundaries. Which means the warning deserves attention, but so does the question of who is issuing it.

The broader theme — how much trust to place in AI tools at work — has already surfaced as a real incident this month. Alibaba’s ban on employee use of Claude Code runs in the opposite direction (distrust of the tool itself), but arrives at the same question: using AI at work means deciding how much of your company’s information to entrust to someone else.

Counting What You Give Away

AI adoption decisions tend to focus on price and performance. Nadella’s essay adds a third column to the comparison: the knowledge your organization is handing over. The daily prompts, the corrections to model output, the procedures given to agents — it is worth taking stock of what these add up to over time.

In practical terms, the essay’s five C’s suggest actions like: (1) check your AI services’ terms for when input and usage data can be used for training, and what the opt-outs are; (2) reduce dependence on any single model by keeping an orchestration layer that allows switching; (3) manage your evaluation criteria and feedback records as company assets. None of these require a sophisticated AI platform to start.

That said, how much of the essay’s risk materializes depends on how each service actually handles data. Most major AI services promise, in their enterprise plans, not to train on customer data — reading the essay as “use AI and your knowledge will leak” is a leap. The essay itself accuses no specific company of violations; it reasons about the structural consequences if learning stays one-directional. Still, the point about fair-use trainers restricting their own customers’ distillation is a sharp lens on the industry’s power dynamics. The original is a short essay, and worth reading in full if you are involved in AI adoption decisions¹.