Palantir CEO Alex Karp faults how major AI labs price tokens, saying customers pay for compute that does not translate into value

Palantir CEO Alex Karp escalated his public critique of how leading AI companies price access to their models, arguing that the current “token” billing structure can push enterprise customers to pay for usage that fails to generate corresponding business value.

According to remarks reported by Yahoo Finance, Karp said he has heard privately from enterprise leaders that they are effectively paying for tokens, the units that measure the amount of text or data processed by an AI system, even when that token consumption does not yield clear returns. His concern was less about whether AI works and more about whether the commercial packaging aligns with how organizations evaluate outcomes such as productivity gains or cost control.

Karp’s comments targeted the pricing strategies associated with OpenAI and Anthropic, both of which sell or license access to their AI models based on token usage. In token-based pricing, customer costs generally rise with the volume of input and output the model processes, regardless of whether the resulting content or decisions improve a specific workflow. Karp’s framing suggested that customers may end up paying for experimentation, low-quality intermediate outputs, or iterative prompts without seeing direct value from those compute cycles.

The criticism landed as AI spending continues to shift from prototype deployments toward enterprise-scale adoption. For large buyers, procurement teams and finance leaders often demand measurable results, including tighter integration, predictable unit economics, and clearer accountability for which parts of a system drive value. Token pricing, while straightforward from a vendor’s perspective, can appear opaque to buyers when it is not tightly connected to business metrics, such as reduced cycle time, fewer errors, or improved revenue.

Palantir, whose platform is designed to integrate data, software workflows, and operational decision-making, has positioned itself as an enterprise layer for using AI in applied settings. The company’s approach depends on consistent deployment practices and repeatable performance, which makes the cost of AI usage a core operational consideration rather than a purely experimental expense.

Still, the remarks reported in the Yahoo Finance account did not specify which deals, customer programs, or contracts Karp was referring to, nor did he describe a concrete alternative pricing model. It was also not disclosed whether Karp believes token pricing is inherently flawed, whether he objected to specific rate cards, or whether the issue is primarily the difficulty of measuring “value per token” across different AI tasks.

What remains unclear is how broad the problem is across Palantir’s own customer base and whether the company has changed any procurement guidance or billing assumptions as a result of the public critique. Karp did not provide detailed examples, such as the size of token bills, the particular use cases in question, or the duration over which customers experienced the mismatch between cost and benefit.

Going forward, investors and customers may look for whether Palantir’s leadership extends this critique into concrete product or commercial proposals, including how it measures AI usage efficiency inside deployed systems. They may also watch for responses from the AI labs he named, particularly around whether token pricing can be paired with stronger assurances, usage reporting, or outcome-oriented pricing for enterprise buyers.