Palantir's Karp Says Token Costs Are Breaking the AI Business Model

Palantir Chief Executive Alex Karp has publicly turned on the pricing models underpinning OpenAI and Anthropic, declaring that "something has gone completely wrong" with the way those companies charge for AI. Karp argues that skyrocketing token costs are now a structural problem, not a rounding error — one that is actively steering enterprise buyers toward cheaper alternatives.

The Commercial Case Against Tokenmaxxing

Karp's central charge is that the token-consumption model, as practiced by the two dominant closed-model providers, has made cost management impossible for companies that want to run AI at any serious scale. He singles out what he calls "tokenmaxxing" — the pattern of maximizing token usage — as a practice that has inverted the incentive structure for AI deployment. Instead of rewarding efficient, outcome-driven use, the current pricing model rewards burning through more tokens.

For Karp, this is not a philosophical disagreement. It is a buying-behavior problem: companies confronted with runaway inference bills are making a rational choice to route workloads elsewhere.

Open-Weight Models as the Pressure Valve

The beneficiaries Karp identifies are open-weight models — AI systems whose parameters are publicly available and can be run without paying per-token fees to a third-party API. As token costs from OpenAI and Anthropic climb, those models become a more attractive baseline for enterprises that need predictable infrastructure economics.

This reframing matters commercially. If Karp is right that cost pressure is already redirecting workloads, it implies that the market share assumptions baked into OpenAI's and Anthropic's growth projections face a structural headwind — one driven not by model quality but by unit economics.

What It Means for Palantir

Karp's attack is not disinterested. Palantir sells AI platforms to governments and large enterprises and positions itself as a provider of AI systems that deliver measurable operational outcomes rather than raw model access. A market narrative in which token-based pricing is seen as extractive rather than efficient is one that advantages Palantir's approach.

Whether enterprise buyers act on the cost-efficiency argument at the scale Karp describes remains to be seen. But his willingness to name OpenAI and Anthropic directly, and to frame their model as a system failure rather than a pricing quirk, signals that the commercial debate over how AI gets paid for is moving from boardrooms to public view.