Alex Karp and the Great AI Trust Recession
Karp’s CNBC critique mirrors the reality of enterprise AI procurement. I analyze his claims, identify their accuracy, and outline the four signals determining who ultimately captures AI’s value.
On the first morning of July, Alex Karp sat down at CNBC’s table and made a promise. “I have so much respect for Nvidia and Jensen Huang,” the Palantir co-founder began. “I’m going to try to keep this more adult than I usually do.”
The promise lasted about ninety seconds. Over the next seven minutes, Karp described the dominant mode of enterprise AI adoption as companies “chillaxing” with tokens. He accused the industry of a plan to “triply oversell” its product. He likened AI spending to “a wealth tax that does not help the poor,” called a piece of Silicon Valley consensus “effing insane,” and described himself as “the neurodivergent crazy person that apparently is on drugs, the one thing I don’t do.” When an anchor observed that he sounded angry, Karp corrected her: “This is the voice of American business that is being channeled through me.”
It is tempting to file the whole thing under Karp Being Karp, the philosopher-CEO with a Frankfurt doctorate on aggression and jargon, whose career has been a longitudinal study of both. The eccentricity is not incidental to Palantir’s brand; it is the brand, and it has been since the years when the company was dismissed as a consultancy in a software costume.
But I want to make a different argument, and I want to make it as a practitioner rather than a spectator. Before I edited this newsletter, I spent a decade as a product manager building and selling B2B AI systems, pre-LLM machine learning, then the generative wave. I have run the pilots Karp is mocking. My honest reaction to his CNBC performance was not amusement. It was recognition. Strip away the theater and Karp said, in public, the thing that frontier-lab CEOs will not say and that every enterprise buyer already believes: “something has gone completely wrong” with how AI is sold, and the wrongness is not about capability. It is about trust.
Capability compounds with scale. Trust does not. You cannot fix a trust deficit with a bigger training run, and that asymmetry, models improving faster than anyone’s willingness to depend on them, is the defining economic fact of this cycle. Call it the AI trust recession: a period in which the technology gets monotonically better while committed adoption stalls, because the buyers being asked to bet their businesses on AI do not believe the sellers are on their side.
Karp made three claims that morning. Each is testable. Each maps onto a mechanism I’ve watched operate from inside the deal. Together they amount to a theory of where the money in AI actually goes, and it is a better theory than the one embedded in current market prices.
Claim one: the pricing is a confession
The frontier labs charge for tokens, metered units of usage, like kilowatt-hours. Karp’s knife: “If it was so valuable, let’s say I can make you a billion dollars right tomorrow, wouldn’t I say, I’ll make you a billion dollars and I want 30 percent? Why are they charging for tokens if it’s so valuable?”
Economists have a dry name for this, revealed preference, but the sales floor version is blunter: watch how something is priced and you learn what its seller privately believes. A litigator certain of victory takes the case on contingency. A gym, certain of nothing except your January optimism, sells you a membership and hopes you never show up. Outcome pricing is the tell of confidence; usage pricing is the tell of doubt. The AI industry talks like the contingency lawyer and bills like the gym.
I want to be fair to the labs here, because I’ve priced AI products and it is genuinely hard: outcomes are contested, attribution is a nightmare, and usage pricing is the only model that scales without a services army. But the buyer doesn’t grade you on difficulty. In every enterprise deal I’ve worked, the sophisticated customer runs exactly the inference Karp describes: if the people who built this thing won’t stake their revenue on its value, why should I stake my company on it? That inference, not model quality, not GPU supply, is the real bottleneck in enterprise AI. Karp says he hears it in “every single enterprise” he deals with. So did I, for ten years, about far less capable systems.
His other two claims explain what those enterprises are actually afraid of, and why the fear is structural rather than paranoid. One is about where AI’s durable profit pools sit, and Karp backs it with an admission against his whole industry’s interest. The other is a fear that, in his telling, Fortune 500 executives will only voice in private, and it ends with the Pentagon’s favorite software salesman reaching, on live television, for Karl Marx.
Below the paywall: the “chillax economy” and the precise mechanism by which enterprise pilots die in committee, from someone who has watched it happen a dozen times; the electricity analogy that reframes the entire capex boom; the strongest case against Karp, part of which he made himself, possibly by accident; and the four signals that will decide, within eighteen months, who captures AI’s value. If this newsletter sharpens how you underwrite the AI trade, this is the piece to upgrade for.