Palantir CEO Alex Karp walked onto CNBC on July 1, 2026, and said what a lot of enterprise technology leaders have been thinking but not saying out loud.

“Something has gone completely wrong,” Karp told the anchor. He was talking about the token-based pricing model used by OpenAI and Anthropic, and he did not hold back. He called their models “completely, irresponsibly, oversold,” and said enterprises deploying them are “livid” about how much proprietary business value flows straight to the AI labs rather than staying inside the companies that paid for the service.

“The enterprises are just tired of it,” Karp said.

What Is He Actually Complaining About?

The token pricing model works like this: every time you send text to an AI model or receive a response, you pay per token. For simple chat interactions this felt manageable. But as businesses have moved toward agentic AI, where models are planning, reasoning, calling tools, and executing tasks in long autonomous chains, the token count per task has multiplied by 10x to 30x compared to basic chatbot use.

The result is that enterprises running production AI workflows are paying enormous fees to AI labs, often for capabilities that use data and context that the business itself provided. Karp argues that this arrangement hands proprietary business value to the model providers, who then own the intelligence built on top of their customers’ own operational knowledge.

Karp framed the alternative as data sovereignty. In Palantir’s model, enterprises deploy AI in environments where they own the data, the model weights, and the inference infrastructure. The intelligence stays inside the business. Nothing flows back to a lab.

He also announced a new Palantir partnership with NVIDIA as a practical implementation of this approach, designed for enterprises and governments that want to run capable AI without the data exposure and cost escalation that come with the token model.

This Is Not an Isolated Complaint

Karp is saying what a growing number of enterprise technology buyers are feeling. Over the past 12 months, AI lab pricing has created a new category of executive frustration at the intersection of cost control and data governance.

The problem compounds as AI deployments scale. A business running a handful of AI pilots at low token volumes may not notice the cost dynamics. The same business running enterprise-wide agentic workflows across sales, operations, finance, and customer service can end up paying more to AI vendors per quarter than it spent on its entire technology stack 18 months ago.

On top of costs, there is the control question. When an AI model processes your proprietary customer data, your internal documents, your pricing logic, and your operational workflows, who learns from that interaction? Closed model providers retain the right to use anonymised interaction data for model training. Enterprises are beginning to ask whether they want to hand that knowledge to companies who are also selling to their competitors.

What This Means for Business

Karp’s outburst signals a real inflection point in how enterprises think about AI vendor selection.

For years, the dominant story was simply “which model is best.” That question is being replaced by a harder one: “what are the total costs and risks of depending on this provider at scale?”

The right answer is not automatically “avoid the big labs.” Frontier models from Anthropic, OpenAI, and Google offer genuinely powerful capabilities, and for many tasks they remain the most practical option. But deploying them without a deliberate strategy, clear cost controls, and a data governance framework is how businesses end up in the situation Karp is describing, where they are furious but locked in.

This is exactly the kind of decision that benefits from independent guidance before you are committed to a vendor architecture. What are you actually building? How does your data flow? What happens to your AI spend when you scale from 50 users to 5,000? Which capabilities genuinely require a frontier closed model, and which can run on open weights at a fraction of the cost?

These are not questions most business leaders have the background to answer quickly. They are also not questions your AI vendor has an incentive to answer honestly.

If you want the playbook other teams are using with Claude and Codex right now, grab the free Working With Claude field guide. Download it here.