The obvious AI bill is the one with tokens on it.
The more important bill is quieter. It is the prompts, corrections, workflows, evals, tool traces, edge cases, and institutional judgment a company feeds into a system so the system can become useful.
That is why Satya Nadella’s “reverse information paradox” is worth paying attention to.
The idea is simple. In classic information markets, the seller risks giving away knowledge in order to sell it. In AI, the buyer risks giving away knowledge in order to use what they bought. TechCrunch picked up the argument, and it lands directly on the enterprise AI fault line: the better you want a model to understand your business, the more of your business you must reveal to it.
That is not just a privacy issue.
It is a compounding issue.
Where does the learning accumulate?
Every serious AI workflow produces a trail. Someone writes the prompt. The model gets it partly wrong. A person corrects it. A tool is added. An eval catches the failure. A memory is updated. A routing rule changes. Over time, that trail becomes a map of how the organization thinks, decides, tolerates risk, and defines quality.
That trail is not exhaust in the casual sense.
It is learning capital.
If that learning accumulates inside a vendor boundary you do not control, the company may still get useful work today. But the durable intelligence created by the workflow is not fully theirs. If that learning accumulates inside their own boundary, the same daily work can improve their private evals, internal memory, routing layer, and operational playbook.
This is why model routing and open-weight models are not just technical preferences. They are strategic pressure valves.
The question is not whether closed models are bad. They are often excellent, and pretending otherwise is childish. The question is whether a company has a way to keep its own learning loop from becoming permanently dependent on one provider’s infrastructure, terms, limits, prices, and product direction.
That is the part founders and operators should care about.
Use the strongest model when the work justifies it. Use cheaper or local models when the task is repetitive, private, or good enough with a smaller system. Keep the orchestration layer outside any single vendor. Keep your evals. Keep your traces. Keep your corrections. Keep the memory of how the company actually works.
The model is rented intelligence.
The learning loop should be owned.
Sources: sn scratchpad, TechCrunch