The most powerful accusation in the AI economy cannot be proven. Neither can its denial. That is the actual crisis.
Alex Karp went on CNBC's Squawk Box to announce a partnership with Nvidia. Within minutes he was doing something else entirely: relaying, in public, what enterprise executives say to him in private about OpenAI and Anthropic.
"I am paying for tokens that create no value. These people are stealing the weights and alpha of my business." Alex Karp, relaying enterprise sentiment, CNBC Squawk Box, July 1, 2026
He said the enterprises are "livid." He called the arrangement "a wealth tax that does not help the poor, it just punishes." He said "something has gone completely wrong" with how AI is sold. Palantir's stock rose nine percent before the market closed.
Start with the obvious. Karp delivered this verdict while announcing a competing product, one whose entire pitch is that customers keep their own weights, their own data, and their own alpha. He is not a neutral witness. He is the most motivated witness available, and the market rewarded the performance the same day.
So discount the messenger. What survives the discount is still valuable, because Karp was not making a claim about the technology. He was reporting the private sentiment of the buyers. And an executive class that is furious in private and silent in public is not a rumor. It is a market condition. Nobody wants to be the chief executive who accused the labs on the record and then discovered the accusation could not be backed. So the anger circulates in boardrooms and dinners, unpriced and unresolved, while the contracts keep renewing.
That silence is the tell. People go quiet in exactly one situation: when they believe something they cannot prove.
Here is the part the coverage skipped. The frontier labs' enterprise contracts generally state that customer data is not used to train their models. Taken at face value, the theft Karp described is barred on paper for exactly the customers he was describing. If the paper is honored, the alpha never enters the training pipeline at all.
But notice what the reassurance rests on. Paper. A representation in a contract, backed by the vendor's internal controls, which the customer has never seen and cannot inspect. The enterprise cannot verify what actually crossed its boundary. It usually cannot even enumerate the surfaces where its own people are feeding material into models: the sanctioned deployment, the unsanctioned browser tab, the vendor tool with a model quietly embedded in it, the employee's personal account on a consumer tier where the training terms are different. The institution is asked to trust, and it cannot check.
Now turn it around, because the labs are in the same trap. Accused on national television of taking the alpha of their customers, what can they produce in their own defense? A policy page. A contract clause. An assertion. They cannot hand an angry customer a verifiable record proving what was retained, what was discarded, and what never touched a training run, because no independent record of that kind exists.
The accusation is unfalsifiable. The denial is unfalsifiable. Both sides of the largest commercial relationship in technology are standing on assertion. That is not a scandal about one vendor's conduct. It is a structural fact about the entire market, and it means the fear compounds no matter what the truth is, because nothing exists that could settle it.
Karp's sharpest instinct was about pricing. Why does an industry that claims transformative value sell it by the metered token, the way a utility sells kilowatt hours?
The fashionable answer is that token pricing is a confession that the value is not real. I do not think that is right, and the truth is more useful. Token pricing is a confession that the proof is not real.
Think about what it would take for a lab to price the way Karp says a confident vendor would price: as a share of the outcome. Outcome pricing requires attribution. Attribution requires a record both sides trust: what the system actually did, what it produced, what decisions it shaped, what it touched and what it never touched. That record has to be neutral, tamper evident, and independently checkable, because money moves on it and disputes end on it. Every mature industry that prices on outcomes has an instrument like this underneath it. Securities have the audit. Insurance has the actuarial record. Construction has inspection. Medicine has the chart.
Autonomous AI has nothing. No neutral ledger of the act exists between the lab and the enterprise. So the industry prices the only thing it can measure honestly: the meter. Compute in, tokens out. The token is not the business model the labs chose because the value is missing. It is the only business model available when the evidence is missing.
This is also why the anger has no exit. The enterprise cannot prove the value is absent. The lab cannot prove the value is present. Both are reading the same meter, and the meter measures consumption, not consequence.
Follow both problems down and they land in the same place. The alpha question, what left my boundary and where did it go, and the pricing question, what did this system actually do for me, are the same question wearing different clothes. Both are answered by a record of the act itself. Neither can be answered by anything else. Not by policy, which states intent. Not by audit, which arrives after the fact and samples. Not by the vendor's own logs, which are the defendant's diary.
The answer is a control layer that sits beneath the act while it happens: one that decides in real time whether an action is allowed to run, holds the boundary on what may leave, and signs a record of every call it makes that no one, vendor or customer, can quietly rewrite. The record that proves the institution controlled its AI is the same record that proves what the AI actually did, what it touched, and what it never touched. Governing the act and proving the act are one job.
This is what ALEETH builds, and I will state it plainly rather than grandly. The Institutional Control Architecture runs today. It returns a signed yes or no on an action before the action happens. It halts an agent mid-task and quarantines it. It writes sealed, independently verifiable evidence of the whole sequence, and its coverage maps to thirteen governance frameworks, re-sealed twice a day. It does not yet settle revenue shares between frontier labs and Fortune 500 customers. But it is the category of instrument that debate will eventually require, because a market cannot price outcomes it cannot prove, and it cannot resolve accusations it cannot test.
Beneath the act. Above the machine. Impossible to bypass, across the surfaces it governs.
The token is not a confession that the value is missing. It is a confession that the proof is missing. A trillion dollar market is running on mutual assertion, and the first side that can replace assertion with sealed evidence sets the terms for everyone else.
Karp named the anger, and the anger is real. But the resolution will not come from louder accusations or better contracts. It will come from the ledger neither side has: a neutral, signed record of the act. Institutions that hold one will negotiate from evidence. Institutions that do not will keep paying the meter and wondering.
The full standard, Institutional Control Architecture, is published at publications.aleeth.com/standard. The instrument is in production today.
Not pitched. Not promised. Proven.