"There are Marxist overtones and undertones to our business," Alex Karp wrote to shareholders on August 3. "Others, including many of those building large language models, intend, knowingly or otherwise, to capture the means of production of their purported partners."
He wrote that in the same letter that booked $1.9 billion in quarterly revenue, up 93% year-over-year, and $1.1 billion in profit — more profit in a single quarter than Palantir made in total revenue a year earlier.
Reportedly, Anthropic's valuation crested $1 trillion on secondary markets, sailing past OpenAI, whose $852 billion mark reportedly now trades at a discount amid slumping share demand.
Read those two stories as one, because they are. Karp booking a record quarter while accusing frontier labs of trying to "capture the means of production" and Anthropic overtaking a discounted OpenAI are the same trade, run from opposite trenches.
We had asked whether Anthropic's private book could hold above its reported $965 billion Series H mark. The secondaries reportedly answered emphatically — buyers are reportedly offering to sell their homes for shares — and Karp appeared to escalate from "effing insane" to Marxism. Both moves point at the same thing.
The thing is a single enterprise fear: handing your IP, your workflows, your "alpha," to a vendor that also sells to your rivals.
Karp names it precisely in the letter. Customers want control over "data, the prompts that the models ingest, and more fundamentally the organizational and business intelligence, their alpha, that the language labs are not only ready and willing, but structurally designed to capture."
That is not a model-quality complaint. It is a value-capture complaint.
And it explains why two companies with almost nothing in common — a model-agnostic, air-gapped defense-software shop and a frontier lab that raises money on refusal — are being rewarded by the market in the same week for selling the same insurance policy.
Palantir sells sovereignty: own your stack, own your weights, own your compute, and the lab can't tax you or migrate your know-how into a competitor's product. Anthropic sells trust: a brand built on the things it won't do, so you can hand it your crown jewels and believe they're safe.
Opposite pitches. Identical customer anxiety. Which tells you where the fight has actually moved.
Opposite pitches. Identical customer anxiety. The fight has moved off the model layer entirely — the war is now over who owns the layer underneath.
It has moved off the model layer entirely. When frontier models converge — and they are converging — nobody's moat is the model anymore. The war is now over who owns the layer underneath: orchestration, specification, and data control.
That is the real repricing signal buried in this week's split-screen. The market is no longer paying for raw capability. It's paying for who owns the customer's alpha.
Look at what the money is chasing. Anthropic's surge is, per one account, being driven on Claude's torrid revenue growth and momentum as a coding assistant — which this article reads as the orchestration layer where the customer's actual work lives, not a benchmark score.
And Palantir's own pitch, in Karp's words from July, is "control over their compute, their models, their data stack and their alpha." He said customers "want to know they own the means of production. It's not being transferred to someone else."
He was saying "means of production" a month before he wrote the Marxist letter. This is a consistent book, not a one-off tantrum.
Which brings us to the uncomfortable part: Karp is talking his book and he is correctly diagnosing the war. Both are true at once.
The self-interest is glaring. Palantir is one of the frontier labs' most important deployment partners — by Karp's own account, a great deal of what Anthropic does in the field runs on Palantir's rails. The man calling the labs "insane" resells them.
His new Nvidia deal, folding open-weight Nemotron models into Palantir's platform, is the product the "Marxism" is selling. The "livid" CEOs he keeps invoking are his prospective customers. When he ranted on CNBC in July, the stock reportedly popped around 8-9% the same day. This is a pitch that moves the price.
And the valuation math makes the self-interest sharper still. Palantir reportedly trades at roughly 40 times forward sales — about double the multiple implied by Anthropic's $965 billion Series H, per Decoding Discontinuity. A diagnosis can be accurate and self-serving simultaneously. This one is both.
But grant him the diagnosis, because the deeper structure supports it. The entire edifice rests on the labs' pricing power — the spread between what they charge per token and the marginal cost of inference. That premium funds everything: the capex, the research, even the safety posture Anthropic markets so effectively.
That premium only holds while customers believe the value justifies it. And two forces are eating the belief.
The first is open weights and sovereignty — the exact alternative Palantir and Nvidia are packaging. Karp has been pushing enterprises toward open-weight models that, per CNBC's reporting, are "capable of performing similar tasks at a fraction of the price," warning too that Chinese models are closing the gap fast. The moment credible alternatives exist, the token premium compresses.
The second force is more fundamental, and it comes from inside the house. Google DeepMind's own strategy chief reportedly says recursive self-improvement is central to the AI industry's spending thesis — and that without it, today's revenue may not justify Google's roughly $200 billion in annual capex, per The Information.
Sit with that. The lab premium the entire structure rests on is a bet on RSI arriving, not a fact. Strip out the recursive-self-improvement assumption and the capex doesn't pencil.
So the market is pricing two labs to hold $242.6 billion of the $305.6 billion AI 50 pool — roughly 80% of the AI 50 venture funding pool, per Forbes — on what this article reads as a spending thesis that its own architects admit is contingent.
That concentration is the tell. Eighty percent of the money is riding on a bet, and Anthropic's $1 trillion secondary mark is that bet at its most euphoric — a valuation the source itself reportedly calls "driven more by sentiment than by market fundamentals," with some buyers reportedly offering to sell their homes to get in.
Here is the read that should make everyone uncomfortable. The market is currently pricing trust and positioning above revenue precisely because the model is no longer a moat. When capability is the moat, you pay for capability. When capability commoditizes, you pay for who won't steal your alpha.
That's why Anthropic's trust and Palantir's sovereignty are both winning trades this week. Both are answers to the same question — who owns the customer's value — and neither answer depends on having the best model.
But the two trades sit on different ground. Palantir monetizes the fear that the premium is a tax. Anthropic is the premium. If open weights and sovereignty become credible at scale — Karp's whole thesis — the lab pricing power compresses, and the money that funds Anthropic's safety brand, its capex, its trillion-dollar mark, compresses with it.
Which means Karp isn't just talking his book against Anthropic. He's describing the mechanism by which Anthropic's book could deflate. If he's right about sovereignty, the $1 trillion secondary is the top, not the trend.
The performance was unhinged. The self-interest was total. It may also have been early.