Palantir CEO Alex Karp is not mincing words. In remarks reported by Quartz, Karp said something has “gone completely wrong” with how OpenAI and Anthropic charge for AI. His specific target is the token-based model, which he argues has left enterprise customers frustrated and “empty-handed,” meaning buyers feel like they are paying but not getting the kind of usable, confident results they need.

That critique matters because token pricing is not a small tweak to a menu. Tokens are how most usage-based AI contracts translate real compute into billable units, and in enterprise environments, billing mechanics often become the gating factor for adoption. Karp’s message is essentially that the incentives embedded in token-based charging can misalign with how organizations actually deploy AI at scale: teams want predictable budgets, clear success criteria, and output that maps to business value. When customers feel frustrated and empty-handed, the deployment slows down, internal champions get squeezed, and procurement turns into a cycle of “can we afford this?” instead of “can we ship this?”

To understand why Karp is sounding the alarm, you have to look at what token pricing does to buyer psychology. In theory, charging for usage is fair. In practice, it can make costs feel like a black box, especially when prompt length, retries, tool calls, and workflow complexity vary from one use case to another. Enterprises do not just buy AI for a demo. They buy it for workflows that change week to week. If the way usage translates to cost is hard to forecast, finance teams will get conservative, and product teams will be forced to ration experimentation. That is one path to the “frustrated and empty-handed” outcome Karp is pointing to.

There is also a procurement and governance angle. Enterprise buyers often need approval from multiple stakeholders: technical leads want performance, legal teams want contract clarity, and CFOs want cost controls. Token-based models can be challenging to compare across vendors because pricing terms may be explained differently, bundled with different packaging, or affected by different rate limits and tooling. Even when the underlying pricing is transparent, the operational reality is not always intuitive. So an enterprise can end up feeling like it is paying for consumption rather than measurable deliverables, which is exactly the kind of mismatch that turns AI rollouts into contested internal budgets.

Karp’s comments land in a broader moment where the AI market is still “forming,” and the winners are not only the best models but also the best business models. OpenAI and Anthropic are both trying to scale enterprise adoption while maintaining revenue that is tightly connected to usage. Palantir, meanwhile, positions itself around turning AI into operational outcomes, which naturally puts pricing and value translation under the microscope. When an executive like Karp complains that something has “gone completely wrong,” it is a sign that pricing is becoming a competitive lever, not a back-office detail.

Second-order implications for boards are also real. When AI spend becomes contested due to unclear unit economics, it can trigger governance questions: Who owns the ROI? Is there sufficient measurement? Are pilots turning into permanent programs or dissolving under cost pressure? A board does not need to take a position on whether token pricing is “wrong” in absolute terms. But it does need to ask whether enterprise customers can scale adoption without predictable financial guardrails. If Karp’s claim resonates in customer circles, it suggests that pricing strategy could influence retention, contract renewals, and how quickly enterprise deals close.

There is also a regulatory shadow over all of this, even if Karp is not naming regulators in the Quartz piece. In general, regulators have been attentive to issues like consumer protection, transparency, and fair dealing in digital services. Pricing models are part of that transparency story. For enterprises, unclear cost drivers can prompt internal compliance worries and vendor risk assessments. Meanwhile, for AI vendors, the more usage-based billing becomes central to enterprise procurement, the more vendors will face pressure to make pricing understandable not just mathematically, but operationally.

For decision-makers evaluating AI vendors today, the takeaway is simple but uncomfortable: pricing mechanics shape adoption outcomes. If token-based charging drives customers to feel frustrated and empty-handed, then even strong model performance may not convert into successful enterprise deployment. And if Palantir’s CEO is hearing this in the market enough to declare that the system has “gone completely wrong,” other leaders should assume procurement friction is spreading, not shrinking.