Palantir CEO Alex Karp used a 19-minute CNBC Squawk Box appearance on July 1 to declare war on the token-based business model that powers OpenAI and Anthropic. “Something has gone completely wrong,” he said. “The basic view among enterprises in this country is I’m going to chillax and waste my time with tokens.” Shares rose 8%.
🔍 THE BOTTOM LINE
Karp’s rant was part sales pitch, part genuine industry fault line. Enterprises are balking at runaway token costs, and the sovereign AI argument — own your models, own your weights, own your data — is gaining traction globally, including in New Zealand. But the messenger has an obvious stake: Palantir sells the infrastructure that makes sovereign AI work.
The Argument
Karp’s core thesis is straightforward. When you pay OpenAI or Anthropic per token, you’re renting access to a model trained on everyone’s data — including, potentially, yours. You don’t own the weights. You don’t control the compute. Your proprietary data, the thing that gives you a competitive edge, becomes training fuel for a model that also serves your competitors.
“They want to know they own the means of production,” Karp said of enterprise customers. “It’s not being transferred to someone else.”
A day before the CNBC appearance, Palantir posted a 9-point manifesto on AI sovereignty to X. It called “tokenmaxxing” — the pursuit of high token usage as a proxy for AI adoption — something that “hijacks your value orientation and decreases your institutional fortitude.” The manifesto argued that organisations should control their own model weights, comparing them to “the distilled form of hard-won, accumulated institutional knowledge.”
None of this is abstract. CNBC reported that enterprises are shifting away from tokenmaxxing toward ROI-focused deployments, building proprietary models on open-weight foundations. Chinese models are simultaneously closing the capability gap, adding urgency.
The Palantir-Nvidia Play
Karp’s timing wasn’t accidental. Earlier in the week, Palantir announced an expanded partnership with Nvidia to build custom models for US government agencies using Nvidia’s Nemotron open-weight models inside Palantir’s Sovereign AI platform. The pitch: bring your data, we’ll build you a model you actually own, hosted on infrastructure you control.
This positions Palantir-Nvidia as the “own your stack” alternative to the OpenAI/Anthropic “rent tokens” model. Karp framed it as a national security imperative, not just a business preference: “Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane.”
The Delivery
Karp’s delivery was — by any measure — unusual. Aaron Rupar’s tweet sharing the clip called it a “televised nervous breakdown.” It racked up 6,700 likes and 7,700 bookmarks. CNBC anchor Becky Quick told him: “You sound pretty angry.” Karp responded: “No, this is the voice of American business that is being channeled through me.”
He went on tangents about wealth taxes, called himself “the neurodivergent, crazy person that apparently is on drugs, the one thing I don’t do,” and accused the AI industry of being “completely, irresponsibly over-sold.” The performance overshadowed the substance — which is a shame, because the substance matters.
Why It Matters for New Zealand
Karp’s sovereign AI pitch isn’t just American. We’ve been tracking this tension for months. New Zealand faces the same question: do you rent AI capability from US labs and hand over your data, or do you build your own stack and keep control?
We’ve already looked at how the APAC sovereign AI race is leaving NZ behind and how local efforts like Xeroforce and Project Kererū are trying to build domestic AI capability. Karp’s argument — even filtered through a sales pitch — reinforces the case: governments and enterprises that don’t control their own model weights are ceding long-term competitive advantage.
The counterargument is that most organisations can’t afford to build their own models, and open-weight models like those now matching closed-lab performance make sovereignty more accessible than ever. But accessible isn’t the same as easy. It still takes infrastructure, talent, and political will.
The Other Side
OpenAI and Anthropic would argue the token model is precisely why AI adoption has exploded. Pay-per-use means any company can access frontier models without a data centre. The friction is the feature — you get GPT-5.6 or Claude Sonnet 5 without buying GPUs or hiring ML engineers.
They’d also point out that Palantir’s sovereign AI platform is itself expensive, exclusive, and oriented toward defence and intelligence clients with budgets to match. Sovereignty for Fortune 100 enterprises and three-letter agencies is not the same as sovereignty for a Wellington SME.
Karp’s response would be that the token model’s low barrier to entry is the trap. Easy today, dependency tomorrow.
❓ FAQ
What is “tokenmaxxing”? A term for the pursuit of maximum token consumption — using AI models as much as possible — as a proxy for AI adoption. Karp argues it creates the illusion of progress while transferring your proprietary data to model providers.
Does Palantir actually build sovereign AI? Yes. Palantir’s platform lets organisations deploy AI models within their own infrastructure, retaining control over data and model weights. The Nvidia partnership adds open-weight Nemotron models to that stack.
Is Karp right that the token model is broken? For large enterprises with sensitive data and long-term competitive concerns, the argument has merit. For startups and small businesses, pay-per-token remains the most accessible path to frontier AI. Both models coexist — the market is bifurcating.
🔍 THE BOTTOM LINE
Karp’s rant was loud, but the signal under the noise is real. The AI market is splitting. On one side: rent tokens, get frontier capability, accept the dependency. On the other: build your own stack, own your weights, accept the cost. New Zealand sits at the same fork. The question isn’t whether sovereign AI is a good idea — it’s whether we’re willing to pay for it.
