Microsoft Chairman and CEO Satya Nadella spent Sunday, July 12, explaining why every company buying AI from closed labs like OpenAI and Anthropic is paying twice: once in dollars, once in the proprietary knowledge those systems quietly absorb. His essay, titled "The Reverse Information Paradox," reached 6.2 million views on X within a day and neared 10 million soon after. His proposed fix asks the labs for something they have never offered customers in return: the right to learn from their own AI usage the way the labs learn from everyone else's.

On Sunday afternoon, Microsoft Chairman and CEO Satya Nadella opened a long essay with a reference to a 64-year-old economics paper. Within a day, according to American Bazaar, the post had been viewed almost 10 million times.

Nadella wasn't writing about Windows updates or an Azure discount. He was describing what he sees as a hidden toll built into every enterprise AI contract, one that grows the more a company actually uses the tool it paid for.

"You essentially 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," Nadella wrote. "The better you want the model to perform, the more of that knowledge you have to feed it!"

He called it the Reverse Information Paradox. The name is a direct rewrite of Arrow's own framework, and it quickly became one of the most argued-about posts in tech.

An Old Paradox, Turned Inside Out

Nadella built his essay on a 1962 paper by Kenneth Arrow, the Nobel Prize-winning economist who wrote it while affiliated with the RAND Corporation. Arrow described a problem facing anyone who sells information: a buyer cannot judge what it's worth without seeing it first, but once they've seen it, they already have it for free. Patents, Arrow argued, were one partial fix, letting an inventor disclose an idea without simply giving it away.

Nadella's essay flips the direction of the risk. "AI creates the reverse problem," he wrote. "In the AI age, the buyer risks giving away knowledge, just in order to use what they bought." Where Arrow's seller had something to lose by proving value, Nadella argues today's enterprise buyer has something to lose simply by using the product correctly.

What "Exhaust" Actually Means

The mechanism, according to Nadella, isn't a single leak. It's an accumulation. He calls it "exhaust," the prompts employees write, the tools an AI agent calls, and above all, the corrections people make when a model gets something wrong.

"Every correction is distilled into institutional know-how," Nadella wrote. "It's the kind of knowledge a competitor could never buy, and the kind that leaks almost imperceptibly: trace by trace, correction by correction, eval by eval."

Multiplied across thousands of daily interactions, that exhaust becomes an archive of how a business actually operates, distinct from anything in its official documents. Nadella linked the idea to the Austrian economist Friedrich Hayek, describing it as "your particular intelligence... the knowledge of time, place, and circumstance that no one else can hold."

A One-Way Street, By Design

Here is where the essay turns into an accusation, even though Nadella never names a company. Distillation is the practice of feeding a model's own answers back into a new, cheaper model to copy its behavior, without paying for the original training run. OpenAI and Anthropic, two of the industry's dominant proprietary labs, both prohibit customers from distilling their models under their terms of service.

Nadella's essay itself never surfaces the words "OpenAI" or "Anthropic." TechCrunch and The Decoder each identified the two labs as the obvious targets anyway, since both train on public data under fair use while banning customers from doing anything similar with their outputs.

"While the great innovation that comes from model providers having fair use rights to train models on public data is needed, I find it ironic that the status quo is to then turn around and impose restrictive terms on distillation, and to reserve the right to learn from customer usage and interaction data," Nadella wrote.

His conclusion: "If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself."

Microsoft ended core pieces of its exclusive alliance with OpenAI on April 27, 2026, and has since shifted parts of its Office suite onto its own in-house models, according to American Bazaar. Nadella's company is also negotiating to sell Anthropic the Maia AI chips that would run Claude's inference workloads, a reminder of how tangled the supplier and critic roles have become.

Five Letters, One Blueprint

Nadella's fix comes down to five principles, each starting with the letter C.

Principle	What Nadella Wants Enterprises to Do
Control	Keep ownership of memory, feedback, evals, and outputs generated from a company's own tasks
Capability	Build private learning environments inside the company's own tenant, not the vendor's
Choice	Avoid depending on one model; stay able to operate if that model disappears
Cost	Separate the orchestration layer from any single model to control spending
Compound	Combine the first four into what Nadella calls a "continuous learning loop"

He reinforced the argument by quoting Palantir CEO Alex Karp: "What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it's not being transferred to someone else."

The Other Side

Model providers have a real defense, and it isn't just profit protection. In February 2026, Google said its Gemini model was hit by a single cloning campaign that fired more than 100,000 prompts at it, extracting its internal logic to build a competing model for free. OpenAI separately told the U.S. Congress that DeepSeek was disguising attempts to copy its models.

Around the same time, Anthropic accused Chinese AI labs of mining Claude through millions of prompts, including a reported 28 million exchanges tied to an Alibaba-linked distillation effort. From that angle, distillation bans look less like customer lock-in and more like defense against actors who never paid for the underlying training at all.

There's a fairness gap on Nadella's side of the ledger too. The New Stack pointed out that Microsoft sells Copilot, a product whose value depends on wide access to enterprise email, documents, and chat logs through the Microsoft Graph. Research from Concentric AI found Copilot accessed nearly 3 million confidential records per organization in the first half of 2025.

EPC Group audits separately found roughly 80% of enterprise Microsoft 365 tenants carried what they called significant oversharing risks. Microsoft maintains that data pulled through Graph is not used to train its models, but the U.S. House of Representatives banned, then later reversed a ban on, staff using Copilot over exactly these concerns. Every recommendation in Nadella's essay also happens to run on cloud infrastructure that Microsoft sells.

The Bottom Line

Strip away the economics lecture and Nadella's essay says something simple: the more useful an AI system becomes to a business, the more of that business ends up inside the AI system, for free, forever.

Enterprises noticed. Idit Levine, founder and CEO of Solo.io, told TechCrunch she's watching customers ask whether an open-source model running on their own servers can match 90% of what a frontier model does at a fraction of the cost, while keeping control besides. Open models already made up 29% of traffic through Vercel's AI gateway last month.

None of that proves Nadella is right that enterprises deserve a mirror right to learn from AI the way the labs learn from them, or that Microsoft's own products would survive the same scrutiny he's applying to OpenAI and Anthropic. But nearly 10 million views in a day means the argument landed somewhere real.

As Nadella put it: "In consuming intelligence, you are creating intelligence. And what you create should belong to you." Whether the labs training that intelligence agree is a different question, one they have every financial reason to leave unanswered.