Palantir’s Alex Karp says AI “token” pricing has “gone completely wrong,” pointing to open-weight models

Palantir CEO Alex Karp delivered a blunt critique of how the AI industry charges for its products, telling CNBC that “something has gone completely wrong” with the token-based model used by OpenAI and Anthropic. Karp said the token structure has left enterprise leaders focused on “chillax[ing] and wast[ing] my time with tokens,” rather than building systems that deliver business value efficiently.

In Karp’s description, token consumption is becoming a cost driver as companies adopt larger, more expensive AI models. As those costs rise, he said, the market is shifting away from what he characterized as “tokenmaxxing,” and toward evaluating returns on investment. That shift, he argued, is pushing some organizations to reconsider which model types they rely on and how they plan to deploy them.

Karp also tied his criticism to the growing appeal of “open-weight” models, which are released so that customers can run and adapt model parameters rather than relying solely on closed systems. He said open-weight approaches can perform similar tasks for a fraction of the price, positioning them as a more practical option for enterprises and governments managing budgets and data constraints.

The Palantir CEO framed the debate as both economic and strategic. He said business leaders in the U.S. are growing frustrated with the pricing mechanics of closed labs, and that open-weight models may offer a path forward. In the same interview, Karp warned that China’s pace in improving AI model capabilities should not be underestimated, suggesting that cost and performance pressures will keep intensifying.

Karp’s comments also intersect with Palantir’s go-to-market messaging around “AI sovereignty,” a term the company uses to describe the ability for organizations to control their data and systems rather than being locked into vendor-managed tooling. While Karp’s CNBC remarks focused on pricing and model structure, they echoed Palantir’s broader push for customers to deploy AI in ways that better align with internal governance and measurable outcomes.

The backdrop to the remarks includes Palantir’s renewed work with Nvidia on AI tooling for government users. In the CNBC report, it was noted that Palantir had announced an expanded partnership with Nvidia earlier in the week to use Nvidia’s AI tools to build custom models for U.S. government agencies. The company did not, in the text provided, offer specific details on model performance, pricing, or deployment timelines tied directly to that expanded partnership.

Palantir’s public stance is unlikely to change the broader industry’s direction in the near term, but it does underscore a growing tension between how AI is sold and how organizations actually budget for it. Token-based pricing can align incentives for usage, yet Karp’s argument is that enterprises may end up paying for consumption patterns rather than for outcomes that justify the spend.

What Palantir did not disclose in the cited interview is how it would quantify “waste” or translate those concerns into a specific cost model for customers. The company also did not provide granular comparisons in the provided text between OpenAI or Anthropic unit pricing and the cost of open-weight alternatives, leaving the practical magnitude of the savings unclear.

For investors and business buyers watching Palantir, the next announcement to monitor is whether Palantir expands its customer deployments that depend on open-weight approaches and custom model workflows, and whether the company links those deployments to concrete procurement outcomes. The market will also look for additional commentary from other large enterprise AI buyers, especially as AI model costs remain a central operational question for both public sector and regulated industries.