Alex Karp: “Tokens That Create No Value” — What He Said

Alex Karp’s July 1, 2026 CNBC interview — the one where the Palantir CEO said enterprises are paying for “tokens that create no value” — is being quoted everywhere and sourced almost nowhere. Most coverage blurs two separate documents into one blob of “Karp said”: a live Squawk Box exchange on July 1 and a written nine-point “AI sovereignty” manifesto Palantir posted on X the day before.

The distinction is not pedantry. The spoken remarks and the written manifesto make related but different arguments, carry different evidentiary weight, and were published a day apart in different formats. And both landed inside a commercial frame that most write-ups skipped: two days before the interview, Palantir announced an expanded partnership with NVIDIA to deploy open-weight Nemotron models in sovereign environments — the product that Karp’s critique of token pricing conveniently sells. Palantir shares rose 8–9% on the day of the interview.

This guide reconstructs the record with source discipline: what Karp actually said on air, what Palantir actually wrote on X, a claim-by-claim table separating the two, a framework for reading each claim against Palantir’s commercial interest, and what the episode means for anyone paying a metered AI bill right now.

The interview did not arrive out of nowhere. It was the third beat of a tightly sequenced week. On Monday, June 29, Palantir and NVIDIA announced an expanded partnership to deploy NVIDIA’s open Nemotron models inside Palantir’s Sovereign AI environments for U.S. government agencies and critical infrastructure. On Tuesday, June 30, Palantir’s official X account posted a nine-point written manifesto on “AI sovereignty.” On Wednesday, July 1, Karp appeared on CNBC’s Squawk Box — ostensibly to discuss the NVIDIA deal — and delivered the remarks that dominated the day’s AI coverage.

CNBC’s own “Key Points” framing captured the substance in three beats: Karp “blasted the token structure used by Anthropic and OpenAI”; he positioned the Palantir–NVIDIA open model as “the solution for enterprises and governments seeking to own their data”; and he said CEOs are frustrated by “skyrocketing AI token costs.” CNBC also drew the manifesto connection explicitly in its write-up: “Ahead of Karp’s interview, Palantir on Tuesday released a 9-point manifesto on the importance of ‘AI sovereignty’ on social media platform X. The post criticized tokenmaxxing as a business model and encouraged companies to maintain ownership of their data.”

The most-quoted line of the day was also the most carefully hedged. Karp went out of his way to say he was not attacking the frontier-lab CEOs personally — “These people are, [OpenAI CEO] Sam [Altman] and [Anthropic CEO] Dario [Amodei] — there’s nothing more fun than debating Dario in private, so I’m not throwing shade at them — but something has gone completely wrong, and the basic view among enterprises in this country is I’m going to chillax and waste my time with tokens, I’m gonna get no value, and they’re gonna get my IP.”

The affirmative case came when he explained the NVIDIA alignment: “What aligns me with Nvidia, and I think is what the technical customers want, is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, it’s not being transferred to someone else. They’re not interested in some fake deploy co that somehow is deploying tokens that transfers the alpha to a third party. And the jig is up.”

The buyer’s checklist, spoken on air

The most practically useful spoken passage was a checklist framed as questions any AI customer should be able to answer about its vendor: “The clients have to be able to ask and answer very basic questions. Are you keeping the data? Are you going to enter our business?” He then escalated to the government context: “In the classified context, when the Department of War goes to you and says, ‘I need this application’ — do they get to control the weights to do it, or do you get to control the weights? Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley? That is effing insane.”

The “wealth tax” passage — read the framing carefully

The “tokens that create no value” line itself arrived as reported speech, not a first-person claim about Palantir’s own bills. Karp presented it as what enterprise customers tell him privately — that they are livid, and that their complaint runs: “I am paying for tokens that create no value. These people are stealing the weights and alpha of my business, and they’re creating a wealth tax that does not help the poor, it just punishes.” He attributed the sentiment to enterprises across the country and closed the thought with his own diagnosis: the models “have been completely, irresponsibly over-sold.” Flattening this into “Karp says Palantir is being taxed” misreads the transcript — he cast himself as the channel for customer anger, not the customer.

Note what is absent from the spoken record. Across every outlet that covered the interview, no third-party ROI figure, failure-rate statistic, or independent enterprise survey was sourced to Karp. The only financial number he personally offered was Palantir’s own projected free cash flow of $15–18 billion two years out — a self-referential claim about his own company, cited as validation of its non-token operating model.

The manifesto is a different document — written, numbered, deliberate, and posted on X by the company account rather than spoken by its CEO under live questioning. Its nine points build a single argument: institutions that hand their data, weights, and accumulated know-how to third-party model providers are surrendering the source of their competitive advantage.

Point 1 sets the frame: sovereignty is “the precondition for choice,” and relinquishing it transfers your institution’s future choices to others “likely to exploit it for their gain and your loss.” Point 2 is the one most often misread: “Data retention is your treasure. Transfer it at your own peril.” Read in full, it is an abstract argument about institutions guarding their own proprietary data — the compounding asset behind their “winning plays” — before handing it to any third-party model provider. It names no vendor and makes no claim about any specific company’s retention practices.

Point 3 coins the document’s most quotable term. “Tokenmaxxing hijacks your value orientation and decreases your institutional fortitude and intelligence,” it argues — the pursuit of high token usage “incentivizes disposable scripts over robust software — with the addictive feeling of false progress.” Point 4 extends the argument from data to weights: “Controlling your weights is controlling your fate. Weights are the distilled form of hard-won, accumulated institutional knowledge. If you let others control your weights, you are allowing them to migrate the alpha of your business to theirs.” Point 5 closes the economic loop, insisting “There is no contradiction between sovereignty and alpha.” Points 6 through 9 drift toward politics and track-record epistemology — listen to institutions with “a proven record of being right” — and are less central to the AI-economics argument.

One correction worth making explicit, because sloppier summaries have implied otherwise: neither the manifesto nor the interview contains a claim that Anthropic enforces or mandates data retention on enterprise customers. The manifesto’s retention point is advice to customers about their own data discipline. On air, Karp posed “Are you keeping the data?” as a question buyers should put to vendors — he did not assert an answer for any named company. If the episode prompts you to actually ask that question of your own vendors, the current factual landscape around frontier-model data-retention terms and zero-data-retention options is worth reading alongside the rhetoric.

No outlet we reviewed cross-references the manifesto and the interview as two separate documents — most quote one, or blur both into a single stream of attribution. The table below is that missing reference: each core claim, the format it actually appeared in, the primary record for it, and the date. If you quote Karp or Palantir on this episode, this is the provenance to preserve.

Why the discipline matters: the two documents differ in register and in reviewability. The manifesto is a considered corporate text; the interview is a CEO speaking live under pushback, in a roughly three-minute monologue an anchor eventually interrupted. Composite quotes that splice a written manifesto line into a spoken sentence manufacture a statement nobody made — and several of the day’s viral summaries did exactly that.

Everything Karp said is consistent with what Palantir sells. That does not make it false — but it does mean each claim deserves an explicit look at who benefits if you believe it. The interview existed because of the NVIDIA deal; the company that booked the airtime is the company offering the alternative; and the market rewarded the performance with an 8–9% single-day move. The table below runs each core claim through that filter.

The honest interpretation sits between stenography and mockery. Karp is describing a real anxiety — enterprises genuinely are scrutinizing what metered token spend returns, and the pricing debate he attacked is live across the industry — but he is describing it from the seat of the vendor selling the cure, days after announcing that cure, to an audience that bid his stock up 8–9% for the performance. Both facts are load-bearing. Palantir’s own data ambitions are also not hypothetical: the company’s platform strategy has been expanding into commercial data work, as Palantir’s agentic-data push into marketing shows. The messenger has a portfolio; the message still names a real problem.

This was not a friendly platform delivering a monologue. After roughly three minutes of uninterrupted Karp, co-anchor Becky Quick cut in: “You sound pretty angry.” Andrew Ross Sorkin, hearing the repeated disclaimers about Altman and Amodei, observed: “That sounds like shade” — to which Karp replied, “No, no. This is reporting.” He ended the segment self-deprecatingly, telling the desk “I feel like I’m gonna be kicked out of the room” as Sorkin cued him to wrap.

The next day’s coverage then split into three recognizably different products — worth knowing, because which one you happened to read determines what you think happened.

The interview also wandered into tangents that the mockery tier leaned on heavily; none of them bear on the token-economics argument, which is why the demeanor coverage and the substance coverage read like two different events. For a business reader, the practical rule is simple: take quotes from the transcript-grade sources, take the temperature check as context, and let neither substitute for the other.

Strip the theater and the self-interest, and the episode leaves a genuinely useful residue: a set of questions worth asking about any AI spend, regardless of whether Palantir’s answer is the right one for you. Karp’s on-air checklist — are you keeping the data, are you going to enter our business, who controls the weights — is a reasonable procurement standard even if you never buy anything from Palantir or NVIDIA.

Two of those threads have deeper treatments on this blog. The pricing-model tension Karp attacked — subscriptions versus metered usage — is mapped in our breakdown of the subscription-to-usage-credits shift at OpenAI and Anthropic, and the most concrete near-term instance is Anthropic’s announced early-July usage-credit transition. The ownership question underneath the manifesto is the build-versus-buy calculus we work through in custom AI tools versus branded SaaS. And if the honest answer is that nobody in your organization can say what the token bill bought, start with a framework for measuring what your AI spend actually returns.

Looking forward, the safest projection is not that enterprises abandon token-metered AI — the announced pricing changes point the other way — but that the burden of proof shifts. Vendors selling metered access will increasingly be asked Karp’s questions in procurement, in writing; vendors selling sovereignty will increasingly be asked for independent audits of the isolation and portability claims they currently self-attest. Buyers who can answer “what did the tokens buy” with a number, per workload, will be the ones this argument cannot whipsaw. That instrumentation work is exactly what our AI transformation engagements build first.