Palantir CEO Alex Karp attacks OpenAI/Anthropic token-based pricing model as IP extraction

Karp's core argument is not that tokens are simply expensive - it is that the closed-model transaction quietly moves value in the wrong direction. In the CNBC interview he compressed the enterprise grievance into a caricature: 'I'm gonna chillax and waste my time with tokens, I'm gonna get no value... they're gonna get my IP' ^[1]. The load-bearing concept is what he calls 'weights and alpha.' When a business runs a proprietary workflow through a closed model, its prompts, its data, and the patterns that encode its competitive edge pass through infrastructure it does not control, so Karp frames data retention itself as the asset at risk: 'Data retention is your treasure. Transfer it at your own peril.' ^[2]. He paired that with the claim that the underlying models have been 'completely, irresponsibly over-sold' to buyers ^[3]. The proposed fix is ownership of the full stack - 'control over their compute, their models, their data stack, and their alpha' ^[4]- which he casts as owning the means of production: 'They want to know they own the means of production, it's not being transferred' ^[5]. Reframed this way, the pricing complaint becomes a data-sovereignty argument, and open-weight models become the structural answer rather than a cost hack.

Every claim Karp made routes back to a product Palantir sells. He confirmed Palantir now ships a product that lets customers switch between models, positioning directly against lock-in to a single closed lab ^[1]. His open-weight advocacy is not abstract either: it is tied to a Palantir-Nvidia alignment that brings Nvidia's open Nemotron models into US government agencies and critical infrastructure in classified and air-gapped deployments ^[6]. That alliance gives the sovereignty pitch a concrete distribution channel and a hardware partner whose CEO, Jensen Huang, independently frames open-source AI as foundational to national security ^[7]. The commercial logic is clean: undermining closed per-token pricing directly favors a model-agnostic, customer-owned stack, and Forbes reported the market reading the incentive correctly, with VCs including Marc Andreessen amplifying the 'labs are stealing enterprise value' thesis and a projected sovereign AI market of roughly USD 600 billion by 2030 as the prize ^[7]. The move landed for shareholders too - PLTR stock rose roughly 8 to 9 percent on the interview day ^[3]. None of this makes the critique wrong, but it means the messenger profits precisely when the audience believes him.

The reason Karp's monologue is more than a sales pitch is that the enterprise spending shift was already documented before he sat down. CNBC reported enterprises pivoting from 'tokenmaxxing' to efficiency, slowing the forecast revenue ramps for OpenAI and Anthropic, with concrete defections: AI startup Lindy moved 100 percent of its traffic off Anthropic's Claude to the cheaper open-weight DeepSeek, expecting millions in savings, and Uber introduced monthly AI tiers after burning through its annual AI budget in about four months ^[8]. Coinbase reportedly cut internal AI spending by nearly 50 percent by defaulting engineers to Chinese open-weight models ^[6]. Pricing pressure is compounding the effect - Google's Gemini 3.5 Flash was offered at roughly half to one-third the price of comparable frontier models ^[8]. D.A. Davidson analyst Gil Luria warned that 'some of their largest enterprise customers may start limiting their out-of-control token spend' ^[8], and CNBC had separately flagged model routing as a structural threat to per-token revenue weeks earlier ^[9]. Whatever one thinks of Karp, these are independent signals that the per-token revenue model is under real strain.

The most revealing pushback comes from people who accept the premise but reject the vendor. Across social discussion, the token-pricing critique drew broad agreement, yet a recurring objection was that Palantir - itself a data business - is a self-interested narrator, with skeptics framing the fight as two data operations competing over the same toll booth and dismissing the interview as incoherent or angry. On the supportive side, investor David Sacks reframed the 'crash-out' as Karp articulating what real enterprise AI safety looks like - customer control over data rather than abstract alignment - which is the strongest steelman of the argument. But even the Palantir-friendly corners surfaced a takeaway that cuts against Palantir: if there is no moat around the model and the moat is how you use your data, the logical endpoint is self-hosting local open-weight models, not buying anyone's managed stack. That is the unresolved tension. The market data validates the diagnosis, the coverage framed it as a meltdown ^[10], and the cleanest response to Karp's own argument may be to route around both the frontier labs and Palantir alike.