Palantir CEO Alex Karp told CNBC on Wednesday that every enterprise customer his company works with is privately frustrated with the frontier AI labs, and the complaint is specific: the labs optimise for token consumption as a signal of progress rather than for outcomes the business can measure.
“It’s not just the man and woman on the street that is unhappy with the frontier labs, it’s in private, every single enterprise we deal with,” Karp said in an interview with Sara Eisen. He used the term “tokenmaxxing” to describe what customers believe the labs are doing: burning tokens to demonstrate productivity, not to deliver it.
The self-interest flag is worth raising immediately. Palantir sells the application and integration layer that sits between frontier models and enterprise workflows. Karp’s thesis that “the implementation is where the value is, certainly in the next seven years” is also, word for word, Palantir’s sales pitch. That does not mean the customer frustration he describes is wrong. It means you should read his diagnosis alongside his financial incentive.
The underlying signal is worth taking seriously on its own terms. Karp is voicing something the supply side of the AI market has been reluctant to say out loud: the metrics that labs report to investors (token volume, API calls, usage growth) and the metrics that enterprise buyers care about (cost per outcome, measurable workflow improvement) have started to diverge. Enterprises are paying fast-rising AI bills, and the people approving those budgets are increasingly skeptical that spend equals value.
This maps to data that appeared elsewhere this week. Vercel’s routing data showed enterprises shifting token volume toward cheaper models, with DeepSeek absorbing share that frontier providers once held. That is not a capability story. That is a cost-versus-outcome story. If the output quality is adequate and the price is lower, the enterprise buyer moves. The labs, measuring success in tokens consumed, register this as growth. The CFO signing the invoices registers it as a negotiation.
Karp’s comments also land as Anthropic and OpenAI are each moving toward public markets. Both companies will need to demonstrate that enterprise token volume translates to durable revenue, and ultimately to the kind of sticky workflow integration that justifies frontier-model pricing against a backdrop of faster commoditisation. Karp told CNBC that most of Anthropic’s public projects are “running on Palantir.” Whether that claim is precise or promotional, it points to the same structural question: who in the enterprise AI stack captures the margin?
The buy-side correction he describes is not rhetorical. It is a procurement cycle. Enterprises that signed AI spending commitments in 2024 and early 2025 are now at renewal or review stages. The teams running those evaluations are comparing token cost against measurable output, and a growing share of them are finding the ratio uncomfortable. That is what Karp means by “accelerating costs raising alarm.”
The counterpoint the labs would offer is that the value will compound: that today’s token spend is training data, workflow integration, and user behaviour that compounds into defensible moats. That argument is coherent. It is also the argument every SaaS company made about CAC in 2021. The labs that can show outcome metrics alongside consumption metrics will have an easier time making it.
Enterprises currently renewing or renegotiating frontier AI contracts should benchmark their cost-per-outcome numbers now, before the next pricing cycle, because Karp’s observation suggests they have more negotiating leverage than their current spend levels imply.
Reported by CNBC (cnbc.com), 2026-06-10.
