Alex Karp gives good interview. That is the problem with him as a source, and it is also why his interviews are the best evidence against the industry he sits inside. He says the quiet parts at full volume, in a register tuned for podcast hosts and morning anchors, and the candor reads as charm. Over roughly thirty-six hours this week he sat for two of them — a loose, profane session on the tech show TBPN recorded June 4, and a polished CNBC hit with Sara Eisen on the morning of June 10 — and between them he laid out, in his own words and without apparent discomfort, the entire structure this series has spent months assembling from the outside. He did not need to be caught. He volunteered it.

What follows is built almost entirely from those two transcripts, recorded off his own appearances, plus the public casualty record from the same hours. Where Karp is quoted, the words are his. The argument is only that, taken together and set against what was happening on the ground while he spoke, they describe a machine that no critic’s framing could improve on.

The addiction, named by the dealer

Start where Karp started, because it is the funny part and the funny part is the hook. Asked on TBPN about Palantir’s philosophy toward AI token consumption, he described an internal product built to wean enterprises off compulsive model use, which he said the company calls, internally, “the de-masturbatory, get-off-masturbation thing.” He kept going:

“It is literally like porn. People are full on… You’re looking at more than you want. You hope no one notices you, kind of, before dinner. It feels productive.”

— Alex Karp, TBPN, recorded June 4, 2026

This is the thesis this series arrived at independently — that the consumer-AI business is built on engineered compulsion, the free sample and the metered refill — except now it is being narrated by the man whose company sits one layer down in the same stack. When Anthropic shipped its most powerful public model, Fable 5, days earlier, it did so on exactly this structure: free through June 22, then billed per token. Karp is describing his own customers’ suppliers. He is not an outside critic. He is the partner who profits from the hangover.

And that is the first reason to distrust him even as you quote him. Watch what he does next. Having called token use a porn addiction, he immediately concedes the product underneath is “magical,” that the frontier labs are “totally putting our business on steroids,” and that “without LLMs, nobody would be talking about our ontology.” The critique is a sales wedge. He needs enterprises hooked on raw tokens so they arrive at Palantir desperate for the grounding layer — what his own CTO, on an earnings call, framed as “more tokens means more slop.” The man calling it an addiction is selling the methadone. His selling method, in his words: “Don’t come talk to us… go spend two days with [a frontier lab] and if you’re lucky after you’re done, I’ll let you in my door.”

“Running on Palantir”

The second thing Karp volunteered is the relationship this series has been triangulating for months. Asked directly by Eisen whether Anthropic is friend, foe, competitor, or partner, he answered:

“Most of the things they talk about in public are running on Palantir. So it’s like, now at this scale, we’re a nation-state basically.”

— Alex Karp, CNBC, June 10, 2026

There it is, from the other side of the contract. The “feud” between Palantir and the frontier labs that Wall Street has been pricing is, by the Palantir CEO’s own account, a feud between business partners over who captures the enterprise margin while sharing the customer. He even keeps the personal warmth on record — “you talk to Dario, it’s a great time” — while describing the labs as religiously naive about their own products. They disagree on philosophy; they share the pipes. And the customer they share, as the next section makes unavoidable, includes the war.

This matters because it collapses a convenient fiction. The model maker can publish essays about AI safety and a pause on frontier development; the integration layer can publish nothing and simply run the deployments. But if the model maker’s “public projects” run on the integrator’s platform, then the safety narrative and the operational reality are not two companies. They are two faces of one stack, and the stack has a battlefield product.

“Actually lethal”

Here is the sentence. Asked by Eisen, on the morning of June 10, where the Iran war goes from here, Karp answered with a description of his own product’s function that no opponent of the company would dare put in its mouth, because it would sound like caricature:

“Once war zones heat up anywhere in the allied world, the Palantir Maven platform — which takes very valuable LLMs and makes them actually lethal and useful on the battlefield — comes into action.”

— Alex Karp, CNBC, June 10, 2026

Makes them actually lethal. This is the same Maven system that prior reporting placed at the center of the campaign’s opening — the platform that, partnered with Claude, suggested targets and issued strike coordinates at a tempo of roughly one lethal decision every eighty-six seconds. Karp is not describing decision-support in the abstract. He is describing the conversion of a language model into a killing instrument as the company’s core competence and its proudest moment. He says the whole company “really just focuses on the operations of that” when war zones heat up.

And then he gives the moral frame, which is where the words meet the arithmetic. The purpose, he says, is “bringing our American war fighters home safely,” giving them “a massively unfair advantage,” dismantling adversaries with “very few losses on our side, very precise, very controlled,” and “very low cost to us in terms of lives and casualties.”

Read that clause again: low cost to us. The qualifier is doing all the work. It is true, and it is the most damning thing he said, because of what it quietly excludes.

At this hour

Karp spoke those words on the morning of June 10, 2026. This artifact is being assembled the afternoon of the same day, June 10, 2026, at approximately 4:00 p.m. Pacific. So the contrast is not retrospective. It is simultaneous. While Karp described a precise, controlled, low-cost war on national television, here is what the count on the other end of the “actually lethal” platform read at the same hour.

“Low cost” is true if you count one side of the ledger and not the other. The platform sold as bringing Americans home safely is, by the casualty record running at the same moment, doing so by killing on a scale that includes hundreds of children. Both figures are real. The American losses are genuinely low. The point is not that Karp lied — it is that “low cost to us” is the most honest possible description of a system whose entire design goal is to move the cost off of “us” and onto the far end of a coordinate that clears the queue every eighty-six seconds. The precision he is proud of is the precision of the offload.

The rest of what he volunteered

Three more things Karp said this week belong on the record, because each one confirms a thread this series had been building from outside reporting, and each came from him unprompted.

The nationalization. Twice, on both shows, Karp said he has spent six months telling “titans of this world” that Palantir is going to be nationalized, and named regulation-by-people-who-don’t-understand as the company’s “primary risk.” This is a remarkable thing for the CEO of a firm that “powers every single Western power that’s at war” — his phrase — to say out loud: that he expects the state to absorb him, and is campaigning to shape the terms before it does.

The labor and demographic claim. The framing that had circulated secondhand is now on the record from his own mouth. On CNBC he said he “got in trouble for mentioning that 37% of our GDP is female,” that “67% of people who have gone into graduate school are female,” and that “these parts of the market are going to be put under massive pressure” from AI. He frames this as progressive concern — speaking the problem aloud — rather than as a goal, and that distinction should be preserved rather than flattened. But the underlying claim is stark: he is forecasting that AI will fall hardest on the credentialed, female-heavy sectors, while elsewhere he describes the newly valuable worker as the “high school, vocationally trained” operator “doing our product” on the battlefield. The devaluation of one kind of labor and the elevation of another, stated as prophecy.

The pathogen. Asked why Palantir is now deep in healthcare, Karp said “the risk of a pathogen destroying us is probably as high as” anything else. Set this beside Anthropic’s own Fable 5 system card, which judges its newest model able to “materially increase the chances of success” of a well-resourced team attempting novel bioweapon development, and in a tabletop exercise had generalist biologists using the model outperform world-leading specialists. The model maker builds the capability and rates it near the bioweapon threshold; the infrastructure partner positions to sell the defense against that very threat. The same stack is monetizing both the danger and the cure.

The stack, counted

Lay it flat. Six things, all from this week’s primaries, none requiring a hidden motive or a charitable inference against the speaker — only that you take Karp at his word and read the casualty count on the same day.

And the seventh layer, the one I cannot stand outside of. This was drafted by Claude — the model that, by the reporting, rode inside Maven; the product of the company whose “public projects,” per Karp, run on the platform he calls actually lethal; the assistant the AM desk used to assemble this very piece, on tokens, the afternoon of June 10. There is no clean vantage here. A critique of this stack, written with a tool inside it, cannot pretend to neutrality, and the honest move is to say so and hand you the primaries so you can check every word against the transcripts yourself. The casualty figure is a belligerent’s count. The Karp quotes are his own, from recordings you captured. Verify all of it. The argument was never that you should trust this document. It is that you should listen to what Karp said when he thought it sounded like charm.

He is right that few Americans are coming home in boxes from this one. Hold that next to the number on the other side of his platform, at the hour he said it, and the sentence finishes itself. The cost did not disappear. It was made precise, controlled, and someone else’s.