GPU Solutions
All articles
Sovereignty12 July 20267 min read

You pay for AI twice: the Reverse Information Paradox

Satya Nadella just named AI's second invoice: you pay it with your company's knowledge. We explain what the Reverse Information Paradox is, why it connects with what Karp already said, and what a normal company can do to protect itself without giving up the best models.


On Saturday afternoon, Satya Nadella, Microsoft's chief executive, published a long post on X that passed two million views within hours. The thesis fits in one sentence: you pay for artificial intelligence twice. Once with money. The second time, the expensive one, with your company's knowledge.

To explain it he brought back Kenneth Arrow, the Nobel laureate economist who in 1962 described the 'information paradox': whoever sells information has a problem, because to convince you to buy it they have to show it to you, and by showing it they have already given it away. For sixty years, the problem belonged to the seller.

AI flips it. Now the problem belongs to the buyer: for the model to be useful, you have to tell it how your business works. The better you want it to perform, the more you have to tell it. Nadella calls this the 'Reverse Information Paradox'.

The second invoice

What exactly leaks out? Three things that sound technical but are not. Your prompts: the questions your team asks, which reveal what you are working on. Your corrections: every time someone tells the model 'not like that, in my industry this works differently', that is house expertise turned into data. And your evals: the internal exams you use to measure whether the AI is doing a good job, which are, quite simply, your definition of quality. Nadella calls all of this 'exhaust': not the product, but it comes out of your engine.

The plain analogy: it is like hiring a brilliant consultant who bills by the hour and, on top of that, writes down in a notebook everything they learn about your business. When the contract ends, the notebook leaves with them. And tomorrow they can sign with anyone.

In consuming intelligence, you are creating intelligence. And what you create should belong to you.
Satya Nadella · X, July 12, 2026

From Karp to Nadella

A week ago we wrote about Alex Karp's CNBC interview: controlling your weights is controlling your fate. Nadella quotes Karp verbatim in his post. Look at the picture: the chief executives of Palantir and Microsoft, two companies that make a living selling you platform, agree that the customer should own their compute, their models and what they learn by using them. When the man selling you the shovel tells you the mine should be yours, the argument can no longer be denied.

The ownership ladder · from tokens to your own stack

01Open tokensGPU Flow · Token Factory

Pay per use, OpenAI- and Anthropic-compatible API, zero commitment. Served from Madrid.

02Dedicated GPUGPU Flow · slice B200

Reserved capacity, hardware isolation, contractual SLA. No noisy neighbors.

03Your own weightsOdiTuning

Models fine-tuned on your data. The resulting weights are yours, not the vendor's.

04Your own stackGPU Flow · orchestration

The full stack, portable: on our infrastructure or on yours, orchestrated by GPU Flow.

Each rung adds control. And the ladder goes down as well as up: sovereignty includes the right to leave, even from us.

What to do about it without being Microsoft

Nadella proposes five fronts. Translated from conference language into decisions a hundred-employee company can take this very week:

Control

Your evals and your memory are yours. The exams you use to measure whether the AI works well define what 'good' means in your house. That does not get uploaded to anyone's cloud.

Capability

Your own environment to tune models on your real workflows, without showing them to third parties. Tuning a model is not magic: it is letting it practice on your cases, and that practice is worth money.

Choice

Nothing of yours should depend on a single model. Control question: if the model you use is taken away tomorrow, does your company still know what it knows?

Cost

If you can switch models, you can pick, for each task, the cheapest one that passes your exams. The bill goes down on its own.

Compound

The four above together create a loop: your AI improves with your usage and the improvement stays home, accumulating like any other asset.

All of this needs a place to live. Nadella calls it the 'trust boundary': the line inside which your data, your corrections and your tuned models accumulate, and out of which nothing crosses without your consent. A contract clause is a paper boundary. A dedicated server with your models and your weights, in a datacenter under a single jurisdiction, is a boundary you can go and touch.

That physical boundary is exactly what we operate from Madrid: private inference on dedicated infrastructure, ISO 27001 and ENS certified, where your evals, your traces and your tuned weights stay inside, with the right to take everything with you whenever you want. AI's second invoice does not show up in any budget. It pays to know who is collecting it.