Microsoft CEO Satya Nadella posits the so-called “Reverse Information Paradox.” In a post on X, he notes that AI model builders such as Anthropic and OpenAI present their customers with a dilemma. “In the AI age, the buyer risks giving away knowledge, just in order to use what they bought.”
An information provider runs the risk that its own product, its knowledge, will be given away before the customer actually purchases it. According to Nadella, the opposite of this information paradox applies to AI usage. He believes that an imbalance is gradually emerging: AI providers receive information from customers who themselves lack the control, capabilities, freedom of choice, and efficiency that they should actually have.
Criticism of anti-distillation language
We recently discussed the “problem” of AI model distillation. This is the phenomenon in which an LLM is trained on the outputs of a larger, more capable model. There are countless examples: In early 2025, DeepSeek-R1 led to all sorts of smaller distillations that could be run locally but were remarkably capable of reasoning given their minimal size. Meanwhile, Anthropic and OpenAI are accusing Chinese AI model developers of large-scale “campaigns” to copy the behavior of American LLMs. As difficult as this may be for them and their business plans, this extraction of knowledge isn’t really a problem for the outside world. People could also share their own outputs with potential rivals of OpenAI and Anthropic, no one is stopping them.
Nadella hits the nail on the head in his X post. Although he considers the large-scale collection of public data on the internet necessary, Nadella finds 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.”
He could be even more specific, but refrains from doing so. Nadella, it seems, chooses deliberately not to name the example exactly, but the demand for user data has recently intensified. To use Claude Fable 5, users must agree to have their data retained for 30 days. Anthropic claims this is for security reasons. As a result, many organizations are unable to use Fable. Nadella implicitly takes it a step further: he speaks in general terms of a one-way flow when it comes to sharing information. Anthropic, OpenAI, and other LLM providers are only willing to share their AI models under very limited conditions. This means there is a lack of control.
A storm in a teacup
We cannot discuss the relationship between LLM providers and their customers in mid-2026 without referring to the storm that has arisen around the sky-high costs of AI usage. This took some time for organizations, from Fortune 50 companies to regular businesses, to come to terms with. Tokens are exorbitantly expensive, particularly when factoring in API costs for the most impressive models. To be honest, we think it’s a bit of a storm in a teacup. Although LLMs like Fable are anything but cheap, and individual users get a much more affordable package with personal Plus, Pro, and Max subscriptions, something is missing from the cost-benefit analysis. As we mentioned last week, users usually opt for the most powerful AI model regardless of the situation, even though a cheaper, smaller option makes much more sense for the vast majority of tasks.
Where the storm is really raging, however, is exactly what Nadella is emphasizing. The “Reverse Information Paradox” is a self-evident truth. And the timing is extremely fitting: just as LLMs have become so capable that they can finally be used to automate real work, Anthropic and OpenAI are hammering home the dangers of open-source AI models and the need to regulate them. This only fuels the problems Nadella highlights. Control is hard to come by, while the Microsoft CEO argues that organizations must safeguard and navigate their own institutional value. According to him, their own AI outputs should be available for learning from.
Proprietary models
Nadella highlights the importance of maintaining an independent relationship with the chosen LLMs. “Does your company “veteran” capability remain with you even if a given “generalist” model is taken away?” This and more are valuable suggestions for companies. We are still far too early in the AI race to entrust business processes entirely to a specific LLM. First, no AI model remains available indefinitely, unless it is open-source. Second, why rely on a limited capability compared to future models? Who knows, Microsoft might (someday) come up with a competitive alternative.
For Microsoft, it’s more important than ever that LLM providers don’t control too much of business workflows. While Anthropic and OpenAI, as “intelligence providers,” are trying to build a platform beyond the LLMs delivered via the API, it’s essential for SaaS players and existing platform companies not to become a “dumb pipe” for AI. For end customers themselves, the significance isn’t entirely clear: the consequences of LLM providers gaining more power aren’t immediately apparent to them. Later on, however, they will be, just look at, say, the practices Microsoft has engaged in over the years when it comes to monopolistic tendencies.
Read also: AI trains on AI: distillation is a major headache for AI labs