On July 1, 2026, Palantir CEO Alex Karp appeared on CNBC to discuss his company's expanded partnership with Nvidia. The conversation went somewhere else. Over roughly twenty minutes of unscripted commentary, Karp called the AI industry "effing insane," accused OpenAI and Anthropic of charging enterprises for tokens that create no value, and described the prevailing API billing model as a wealth tax on American business.
"I am paying for tokens that create no value," Karp said, channeling what he characterized as unified enterprise sentiment. "These people are stealing the weights and alpha." He framed two questions every enterprise should ask its AI vendor: "Are you keeping the data? Are you going to enter our business?" When the anchor asked if he was angry, Karp corrected the framing. "This is the voice of American business that is being channeled through me."
The voice had a product behind it. In March, Palantir and Nvidia had announced a Sovereign AI Operating System Reference Architecture combining Nvidia's Blackwell Ultra GPUs with Palantir's AIP, Foundry, and Apollo platforms. On June 29, two days before the interview, they extended it: an engine for deploying Nvidia's open-weight Nemotron models inside sovereign environments. The system runs in air-gapped networks. No data leaves the customer's perimeter. No tokens are billed per query. Government agencies and critical infrastructure operators can customize models on their own hardware, train on their own data, and retain full ownership of the resulting weights.
Palantir closed at $125.73 on July 1, up 7.8 percent on 57 million shares, adding roughly $21.7 billion in market value in a single session. But the financial substance underneath the repricing was thin. Every dollar figure disclosed across the week's partnership announcements totaled well under one percent of the single-day gain. The market did not pay for the deal. It paid for the diagnosis.
The diagnosis resonated because the spending crisis is real. The average cost per million tokens fell roughly 75 percent in a single year, from about $10 to $2.50, according to Ramp's enterprise spending data. Enterprise AI budgets exploded anyway. Agentic workflows at 2026 adoption levels consume multiples of what corporate spreadsheets projected. Companies that pushed hardest for AI adoption discovered that falling per-token prices and rising per-engineer consumption do not cancel out. They compound. When GitHub shifted developers from flat-rate Copilot subscriptions to usage-based token billing on June 1, some heavy users reported projected costs jumping from $29 per month to $750 or more.
Karp's contribution was naming the second cost. Not the tokens but the data. Every API call sends context to the lab that bills for it. Enterprise prompts carry customer records, financial projections, internal strategy, competitive intelligence. The labs say they do not train on enterprise data. Karp argued the structure itself is extractive. The vendor learns which workflows enterprises build, which domains consume the most compute, which capabilities command the highest willingness to pay. "Do they get to control the weights," he asked, "or do you get to control the weights?"
In a nine-point manifesto posted to Palantir's corporate X account the day before the interview, the company extended the argument. It coined "tokenmaxxing" to describe an incentive structure that rewards disposable scripts over robust software and transfers institutional knowledge to vendors one API call at a time. Controlling model weights, the manifesto argued, equals controlling institutional fate. The manifesto was the intellectual case. The Nvidia partnership was the commercial one. The CNBC interview was the advertisement.
Palantir's own financials illustrate the economics of the alternative. In Q1 2026, the company reported $1.633 billion in revenue, 85 percent growth year over year, with a 46 percent GAAP operating margin and $871 million in net income. U.S. commercial revenue grew 133 percent. Remaining deal value in U.S. commercial contracts reached $4.92 billion. Karp projected $15 to $18 billion in free cash flow within two years. These are deployment margins, not research margins. Palantir does not train frontier models. It deploys other people's.
That is the tension underneath the thesis. The air-gapped architecture runs on open-weight models. Nemotron is trained by Nvidia. Llama is trained by Meta. LongCat is trained by Meituan. But the frontier models currently driving the enterprise spending Karp criticized are funded by the token revenue his architecture eliminates. Anthropic's annualized revenue surged from $9 billion at the end of 2025 to $47 billion by mid-2026. OpenAI generated $5.7 billion in a single quarter. That money funds the next generation of capabilities. The wealth tax, in Karp's framing, is also the research budget.
After the rally, Palantir was valued at approximately $308 billion, roughly 40 times its forward revenue guidance of $7.65 billion. That multiple prices in a future where enterprise AI migrates from API consumption to on-premises deployment, and Palantir captures the deployment layer. Karp diagnosed a wealth tax and sold an air gap. But the $21.7 billion the market added to his company in one session was not extracted through tokens. It was extracted through narrative. The wealth tax did not disappear. It changed address.