Microsoft CEO Satya Nadella has launched a rare and pointed broadside against the AI industry's leading players, accusing major model providers of a glaring double standard. On July 12, Nadella published a lengthy essay on X titled "The Reverse Information Paradox," sending shockwaves through the industry. The essay effectively takes aim at the asymmetric practices of major model providers such as OpenAI and Anthropic. Anthropic has previously and repeatedly publicly condemned other AI companies for replicating Claude's capabilities through distillation and has consistently called for stronger protections around model capabilities.
Nadella invoked the classic "information paradox" articulated by Nobel laureate economist Kenneth Arrow. Arrow observed that a buyer of information cannot know its value before acquiring it, yet once acquired, they effectively possess it at zero cost. Nadella argues that the AI era has spawned an entirely opposite problem. "You effectively pay for knowledge twice: first in money, and then in something far more precious, the proprietary knowledge you must feed to the intelligence to make it truly useful. And the better you want the model to perform, the more of this knowledge you must supply," he wrote.
He coined the term "Reverse Information Paradox" and warned that information asymmetry worsens over time. The seller learns more about you through your use of the product, but you rarely know what the seller is learning from you.
The Core Double Standard Nadella Is Calling Out
Nadella's most striking criticism targeted the industry's status quo directly. He wrote: "While the fair-use right of model providers to train on public data has been essential to the enormous innovation we've seen, the irony of the current industry landscape is this: on one hand, model providers can learn from public data; on the other, they impose restrictive terms on model distillation and reserve the right to learn from customer usage data and interaction data."
The argument goes beyond simple data protection. AI models learn from all manner of experiences, including the prompts people enter, the tools AI agents use, and the corrections people make when models err. Every correction gets distilled into organizational-level knowledge by the model. This kind of knowledge is something competitors can never buy, and it is the easiest content to leak without even realizing it.
Nadella stressed: "In the process of using intelligence, you are also creating intelligence. And what you create should belong to you." He warned that if the learning process flows in only one direction, economic value will ultimately concentrate continuously among the firms that control the learning infrastructure, rather than flowing to those who actually create the knowledge.
He proposed that enterprises need to establish a new "Trust Boundary" to ensure that data, feedback, evaluation systems, and organizational knowledge are retained internally, forming a self-sustaining capability for continuous learning and intelligence accumulation. Within this boundary, nothing should cross it without enterprise authorization, including so-called intelligence exhaust.
Nadella prescribed four remedies for enterprises. First, create private evaluation systems to define what good outcomes look like. Second, retain ownership of organizational memory, information trails, feedback, and decisions. Third, build proprietary learning environments within tenant boundaries to train or fine-tune models. Fourth, ensure the orchestration layer is not tied to any single model. He asked a crucial litmus test: "If a particular model you're using is removed, do you still have the ability to use other models to run and optimize your evaluations?"
Palantir's Alex Karp Escalates the Attack on Token Economics
The debate ignited by Nadella is far from an isolated event. Shortly before this, Palantir Technologies CEO Alex Karp delivered an even more blistering critique on CNBC. He stated bluntly: "The basic view in corporate America is, I'm going to waste time on tokens, I'm getting no value, and they're going to get my intellectual property." When the host pressed whether this was a rant, Karp responded: "No, no. It's a statement of fact."
Palantir subsequently released a nine-point "AI Sovereignty" manifesto, warning enterprises and government agencies that surrendering sovereignty means handing over your organization's future choices to others who will use it for their benefit and your loss. The manifesto's sharpest accusation targeted the core business model of AI model providers. Palantir stated: "Tokenmaxxing gives you the illusion of progress. Those who sell tokens stubbornly refuse value-based pricing for a reason." The implication is clear: if AI models truly create value for customers, they should adopt outcome-based pricing rather than charging by consumption volume.
Karp's criticism has real-world grounding. According to 24/7 Wall St., ride-hailing giant Uber saw its annual AI budget exhausted in just four months after aggressively rolling out AI coding tools, forcing the company to impose a $1,500 monthly cap per employee. Inside Meta, leaderboards spontaneously emerged tracking which employees consumed the most tokens. Token costs are ballooning, but the return on investment is nearly impossible to measure. This frustration is spreading rapidly through executive ranks.
David Sacks and the Observe-Copy-Expand Pattern
Former White House AI czar David Sacks added his voice to the chorus, directing his fire specifically at Anthropic. He wrote: "Anthropic has rolled out Claude Science, Claude Security, Claude Legal, and Claude Code. Each product directly enters domains previously served by companies building applications on top of its models. The pattern is consistent: observe where value is being created, then move in directly. Dominate the model layer first, then leverage that position to capture the most profitable vertical markets."
This "observe-copy-expand" trajectory is unsettling for the vast number of companies that rely on large model APIs to build commercial applications. For these firms, contributing data and use cases to AI labs may be providing ammunition for future competitors. The anxiety is not just about cost. It is about whether every prompt you send, every correction you make, and every workflow you build becomes part of the model provider's moat instead of your own.
However, the debate has also drawn skeptical voices. Insiders at some AI labs have dismissed the criticism, with one saying that responding to Karp-style theatrics would be foolish as he is simply advocating for his own interests. Critics point out that behind the rhetoric of both Nadella and Karp lies clear commercial logic. Nadella calls for enterprises to establish trust boundaries and learning loops, with the ideal hosting platform naturally being Microsoft's Azure cloud services and its AI development platform Foundry.
The same applies to Palantir. The solution depicted in its AI Sovereignty manifesto essentially positions its core product, Ontology, as the control layer above models. Sarcastic responses quickly flooded X: "So in the end, you're just telling us to buy Palantir, right?"
What This Means for Founders Building on Top of AI Models
Even if commercial motives exist, the problems these warnings highlight are undeniably real. Ballooning token costs, fears that business knowledge may be absorbed by model providers, and uncertainty over who ultimately captures AI's value are all genuine dilemmas facing the corporate world today. Starbucks was recently revealed to be using AI to replace software previously purchased from Microsoft and IBM, putting pressure on both companies' stock prices. This case is seen as a microcosm of how AI is accelerating the reshaping of the enterprise software landscape.
Meanwhile, Meta CEO Mark Zuckerberg has also entered the pricing debate. According to Bloomberg News, Zuckerberg explicitly stated he sees an opportunity to compete on price: "Some other labs have extreme pricing with very high margins. We believe it's entirely possible to offer frontier or high-caliber intelligence services at much more affordable prices."
This debate over AI value distribution, data sovereignty, and model pricing is spreading from tech circles into broader business and policy domains. As AI agents are deployed at scale across enterprises, token consumption is growing exponentially, and corporate scrutiny of costs and value returns will only intensify. The voices of Nadella and Karp may be just the opening salvo in a much larger industry reckoning.
For founders building on top of AI models, the lesson is clear. Build your own evaluation systems. Retain your organizational memory. Keep your feedback loops and interaction data in your own control. Never become dependent on a single model provider for the orchestration layer. As Nadella put it, if the model you are using is removed tomorrow, you need to know that your business can keep running. Your sovereignty in the AI era depends on it.
As of press time, neither OpenAI nor Anthropic had publicly responded to the criticisms. Both companies' current policies state that enterprise customer data is not used to train their models.
