Alex Karp says every enterprise using frontier AI labs is fed up with tokenmax

Palantir CEO Alex Karp believes the biggest enterprise AI problem is not model quality. It is deployment reality. In a Wednesday interview with CNBC’s Sara Eisen, Karp said that “every single enterprise customer” Palantir has is unhappy with frontier AI labs like Anthropic and OpenAI.

Karp’s core argument is blunt: enterprises are “fed up” because these labs seem to chase a different scoreboard, one he describes as a “hyper religion of hyper optimism.” In his telling, frontier companies act as if the problems they create, and the ones they do not acknowledge, will magically be solved by their systems. “Enterprises are fed up because they know this doesn’t actually work this way, and isn’t working,” he said. And that frustration, he argues, is exactly why businesses gravitate toward Palantir Foundry, which Karp describes as AI-agnostic data integration plus “cognizing” (his word) with whichever LLMs a customer chooses to deploy.

What makes this matter is that Karp is not arguing from abstract vibes. He is pointing at a real pattern in enterprise AI outcomes: many projects do not produce the ROI executives are being asked to justify. The source cites a Gartner estimate that only 28 percent of AI use cases fully meet ROI expectations, and that many do not even escape the pilot phase. If you are a CFO, that is the part you cannot wave away with a demo day deck. If you are a CTO, it is the part you can feel in your backlog. And if you are an investor, it is the part that changes how fast “frontier progress” converts into recurring enterprise value.

Karp’s solution is also not subtle. He claims that infrastructure is where the value lives, at least over the next several years, because simply deploying an LLM on a business problem is not enough. He describes Foundry as an AI-agnostic data integration platform intended to unify disparate data sources and then apply cognizing capabilities with whatever models a customer deploys. He also suggests Palantir has considered a sales tactic that would expose prospective customers to frontier labs early, then compare outcomes before contracts are signed. Karp said the idea came up in Palantir leadership debates about paying potential customers to “go talk to frontier labs themselves.” The reported reaction he shared is memorable: customers come out “screaming,” saying frontier labs “don’t understand the enterprise” and “don’t care about my enterprise.”

This framing also connects to a more specific incentive problem: token economics. Karp alleges frontier labs want customers to “tokenmax,” meaning they view token consumption as productivity and usefulness. That charge is not out of left field, at least according to the source. It points to Google CEO Sundar Pichai acknowledging the phenomenon at I/O last month. The logic is straightforward: burning more tokens can become expensive for companies, especially when workflows are redesigned around usage rather than outcomes. The source adds that OpenAI is reportedly considering reducing its per-token charge to attract more customers, in what it describes as a growing competitive war with Anthropic. And Karp calls Anthropic the “leading frontier firm” in his interview, while still accusing the broader category of driving the wrong behavior.

Karp goes further and targets strategy as well as pricing. He says OpenAI’s recent agreement to acquire UK-based AI consulting firm Tomoro, and to use it as part of a newly launched OpenAI Deployment Company aimed at helping customers generate returns from ChatGPT investments, is an attempt to replicate Palantir’s success. Karp’s verdict is harsh: he calls it “a complete farce” and says they do not understand how “unlikeable” they are. Importantly, he also tries to soften the tone by separating people from product. Karp says he is friends with some frontier lab leaders and that they are great to chat with. But he argues the product is expensive and does not actually work in the way enterprise deployments require.

He also brings the conversation back to implementation. Karp says that many things frontier labs brag about publicly are successful because they are “running on Palantir.” He insists that LLMs are still “crucial for the world,” but he argues the implementation layer is where the value will show up, especially in the next 7 years. That is the bet he wants the enterprise market to make: not that frontier models are irrelevant, but that model deployment is the real bottleneck, and infrastructure determines whether AI becomes an operational system or a permanent pilot.

For executives, this is the strategic stake. Enterprise AI budgets are not just competing with other IT priorities, they are competing with the credibility gap between demos and measurable returns. When Karp claims only 28 percent of AI use cases hit ROI expectations, he is effectively challenging boards to ask a tougher question: are you buying capability, or are you buying outcomes? And when he frames tokenmax incentives and “largely religious” assumptions as the root cause, he is urging peers to evaluate whether vendors build for business constraints, deployment friction, and data reality, or for model usage metrics.

For boards and investors, the second-order implication is that differentiation may shift from “who has the best model” to “who can make models usable, governable, and financially defensible in messy enterprises.” Karp is betting that when companies get fed up with token-based value narratives, the winners will be the platforms that absorb complexity first, then let enterprises plug in whatever frontier models they choose.