Alex Karp walked onto CNBC and said out loud what enterprise buyers have been muttering in private: the frontier AI industry, in his telling, is "effing insane," and the people getting fabulously wealthy are the ones selling the tools, not the ones using them.
In a July 1 appearance on CNBC's Squawk Box — ostensibly to promote Palantir's new sovereign-AI partnership with Nvidia — the Palantir co-founder accused leading model makers of "completely, irresponsibly" overselling their systems while quietly absorbing the proprietary data of the companies paying for them. Reporters covering the interview characterized his broader argument as a claim that frontier labs impose a de facto "wealth tax" on American business: enterprises pay steep per-token fees, get little measurable return, and hand over the very data that could sharpen a competitor's model. Palantir's stock jumped roughly 9% in the hours after the segment aired.
The grievance, in Karp's words
Karp's central complaint is that enterprises are "paying for tokens that create no value" while surrendering their operational edge. "The basic view among enterprises in this country," he said, is: "I'm going to chillax and waste my time with tokens, I'm going to get no value, and they're going to get my IP." He describes this as transferring a company's "alpha" — its defensible, proprietary advantage — straight to a third-party lab, because standard closed-model deployments route prompts, customer data and workflows through external infrastructure on every query.
On the national-security side he was blunter still. Asked about relying on Silicon Valley's frontier labs for defense work, Karp called the prospect of outsourcing "the battlefield of this country to the consensus view in Silicon Valley" a security failure waiting to happen — "effing insane," in his words.
The pricing contrast
The numbers underpinning the "wealth tax" framing are real and steep. Frontier Western models command premium rates: GPT-5.5 lists around $15 per million output tokens, and Anthropic's Claude Sonnet 5 launched at roughly $10 per million output tokens on an introductory basis. Against that, Karp is pushing Nvidia's open-weight Nemotron models — the June 29 Palantir-Nvidia "Sovereign AI Operating System Reference Architecture" pairs Nvidia's Nemotron family with Palantir's AIP, Ontology, Foundry and Apollo software. The pitch: agencies and regulated enterprises run capable models on their own air-gapped hardware, owning the weights and the data outright, at a lower cost per unit of capability than renting frontier intelligence by the token.
It is, as one Forbes contributor noted, a very good product pitch dressed as an industry critique. Palantir's Q1 2026 results give the pitch teeth: revenue of $1.63 billion, up 85% year over year, with U.S. commercial revenue up 133% to $595 million.
The inference economics — and the China angle
Strip away the rhetoric and Karp is surfing a genuine shift in how enterprises think about inference. Through 2025 the fashion was "tokenmaxxing" — spending aggressively on AI usage as a proxy for productivity. That has reversed. Uber capped employee AI spending at $1,500 a month after burning through its 2026 AI budget in four months, with its operating chief openly questioning whether the returns justified the bill. The roughly 25-person startup Lindy reportedly cut inference costs by about 90% by switching to DeepSeek, and Walmart, Amazon, Microsoft and Cisco have all introduced spending controls. CNBC reported in late June that OpenAI and Anthropic now face a market pivoting from raw consumption toward measurable efficiency.
That efficiency turn is exactly what makes open-weight models dangerous to frontier pricing power. Chinese open-weight releases — DeepSeek most prominently — have repeatedly demonstrated near-frontier capability at a fraction of the cost, and their permissive licensing lets buyers self-host and avoid per-token metering entirely. Karp's Nemotron push is the American, security-cleared answer to the same economic logic: if the marginal value of a token is falling while its price stays high, the moat migrates from the model to whoever controls deployment, governance and data ownership. That is the layer Palantir sells.
The counterargument is straightforward. Frontier labs still lead on the hardest reasoning tasks, and "cheaper per capability" only holds where an open model is good enough for the job. For frontier coding, agentic workflows and cutting-edge research, GPT-5.5 and Claude Sonnet 5 remain, for many buyers, worth the premium — which is why the labs can charge it.
What to watch
Three things will tell whether Karp's "wealth tax" framing sticks or reads as self-interested spin. First, whether frontier labs cut list prices or introduce cheaper tiers to blunt the open-weight threat. Second, whether Nemotron-and-AIP style sovereign deployments convert into signed government and regulated-enterprise contracts, or stay press releases. And third, whether the tokenmaxxing crackdown deepens into a durable procurement discipline — the moment enterprises start routing routine work to open models by default, the pricing power Karp is attacking begins to erode on its own.
"I'm going to chillax and waste my time with tokens, I'm going to get no value, and they're going to get my IP."— Alex Karp, CEO, Palantir