A Deliberate Provocation
Alex Karp does not do subtle, and his recent CNBC appearance was no exception. For roughly twenty minutes, the Palantir chief executive laid into the business model of the frontier AI labs, arguing that token based pricing from OpenAI and Anthropic is, in his words, "completely wrong." "Something has gone completely wrong," he said, and he meant it as an indictment of the entire economic arrangement between labs and enterprise customers.
We take Karp's provocations with the appropriate grain of salt, since he is talking his own book and Palantir competes for the same enterprise budgets. But dismissing him entirely would be a mistake. He is articulating a discontent that many enterprise buyers feel but rarely voice so bluntly, and the market noticed: Palantir shares climbed about 8 percent following the interview.
The Three Way Squeeze
The heart of Karp's argument is that frontier labs extract value from customers three ways at once. First, they charge for tokens, the metered consumption of model output. Second, they gain access to the enterprise's proprietary data and workflows as those systems get wired into the models. Third, and most provocatively, he contends they will eventually commoditize the very competitive advantages their customers hand them.
That third point is the sharpest. Karp is warning that when an enterprise pours its unique knowledge into a general purpose model, it may be training the tool that later erodes its edge. Whether that fear is fully justified is debatable, and the labs would strongly contest it. But the underlying anxiety, that dependence on a third party model means surrendering both margin and differentiation, is a real one that CIOs are increasingly voicing in private.
The End of Tokenmaxxing
Karp coined, or at least amplified, a useful term: tokenmaxxing. It describes the phase of AI adoption where organizations optimized for maximum model consumption, treating token throughput as a proxy for progress. He argues that era is ending, and that enterprises are pivoting hard toward return on investment, asking not how much AI they can consume but what measurable value it produces.
This tracks with what we are seeing across the market. The initial land grab, where companies raced to deploy AI everywhere regardless of cost, is giving way to a more disciplined phase. Budget owners are demanding proof, and some are adopting cheaper open weight models for workloads that do not require a frontier lab's most expensive tokens. The efficiency reckoning Karp describes is already underway.
Own the Means of Production
Karp framed Palantir's partnership with Nvidia in explicitly ideological terms. "What aligns me with Nvidia, and I think is what the technical customers want, is control over their compute, their models, their data stack and their alpha," he said. "They want to know they own the means of production. It's not being transferred to someone else." The Marxist borrowing is deliberate and characteristically Karp.
Stripped of the rhetoric, the pitch is about control. Palantir and Nvidia are selling the idea that enterprises should run capable models on infrastructure they govern, with their data staying inside their perimeter, rather than renting intelligence from a lab that also sees everything they do. For regulated industries and security conscious buyers, that message resonates regardless of what one thinks of the messenger.
The Self Interest Is the Point
It would be naive to treat this as disinterested analysis. Palantir sells a platform premised on customers keeping control of their data and models, so an argument that frontier labs quietly expropriate value is, conveniently, an argument for buying Palantir. The Nvidia alliance he touts is a commercial product he wants enterprises to purchase.
Yet self interest and insight are not mutually exclusive. Karp is making a case that happens to sell his product, but the case itself, that token pricing misaligns incentives and that dependence on frontier labs carries strategic risk, deserves engagement on its merits. The labs would offer a strong rebuttal about the value and safety their models provide, and enterprises should weigh both sides rather than accept either camp's framing wholesale.
The Debate Worth Having
Whatever one makes of Karp's theatrics, he has forced a debate the industry needed. The comfortable narrative of the past two years held that frontier labs and their enterprise customers were simple partners in value creation. Karp is insisting the relationship is more adversarial than that, and even those who reject his conclusion have to engage with the underlying tensions he names around cost, data, and dependence.
We suspect the truth sits between the extremes. Frontier models deliver capability that most enterprises genuinely cannot build themselves, and the labs are not the predatory expropriators Karp implies. But the concerns about lock in, margin, and strategic differentiation are legitimate and under discussed. The healthiest outcome is not that enterprises abandon the labs, but that they negotiate and architect from a position of clear eyed self interest rather than uncritical enthusiasm.
What CIOs Should Take From It
The practical lesson is to interrogate the true cost and strategic exposure of frontier AI dependence. Token bills are the visible cost, but the harder questions concern data, lock in, and differentiation. What proprietary knowledge is flowing into a third party model, and what happens to your advantage if that capability becomes universally available?
We are not endorsing Karp's conclusion that enterprises should route around the labs entirely. Frontier models deliver genuine capability that is hard to replicate. But his provocation is a useful forcing function. Every AI architecture decision should be examined not just for what it costs per token, but for what it means for control, and Karp has at least made that conversation impossible to avoid.
