Social Capital quotes Microsoft CEO Satya Nadella: companies renting AI models essentially pay for intelligence twice. Once in money, and once in proprietary knowledge that must be surrendered to make the model useful.

The mechanism is written as a learning loop. Each prompt and correction takes away a part of what the company knows how to do. After knowledge is distilled, the provider of the closed model retains this information. Nadella referred to this as the reverse information paradox in July. Kenneth Arrow originally stated that sellers must disclose information to prove its value. AI shifts the risk back to the buyer. The more the seller uses it, the better they understand you, while you know almost nothing about what the other party has learned.

Leaks do not occur through obvious intrusions but through what he calls waste gas: the prompts you write, the tools called by agents, and the corrections you make when the model errs. Every modification teaches the model. He stated this is knowledge that competitors cannot buy, and it is also a nearly imperceptible leak, occurring through each trace, correction, and evaluation.

Social Capital extrapolated from this. If using data is more valuable than the cost of servicing these tokens, the provider may pay you to use it. A precedent is OpenAI's opt-in program: eligible accounts relinquish the training rights of their interface traffic, with certain models able to receive up to 1 million free tokens daily. The official help page states that levels 3 to 5 are flagship models with 1 million daily and small models with 10 million. Levels 1 to 2 are 250,000 and 2.5 million. The default is not mandatory, with a 30-day notice before termination. Fine-tuning training is not included in the quota.

Chips can be rented, algorithms can leak or be distilled. Prompts and corrections will compound at the holder's end. Nadella's solution consists of five parts: mastering evaluation, memory, and feedback, training within one's tenant boundaries, and separating orchestration from single models. If learning only flows in one direction, the value will accumulate to the owner of the learning infrastructure, rather than the creator of the knowledge.

This is data rent, not a computing bill. The buyer is the company that feeds processes and corrections into the closed model, while the seller is the model provider that leaves the distilled results. The driving force is deeper usage and more surrendered knowledge. Money is first paid for tokens, and knowledge is then retained by the other party. The beneficiary is the closed model that receives correction signals, while the burden falls on the renter who embeds proprietary practices into prompts and cannot reclaim learning results.

Source: Public Information