Palantir's Karp bashes OpenAI, Anthropic token model: 'Something has gone completely wrong'

Palantir CEO Alex Karp warned that U.S. AI labs’ “token model” has “gone completely wrong,” arguing enterprises are unlikely to keep “tokenmaxxing” as AI costs rise and newer models get pricier. He says the shift is toward ROI-focused approaches, including open-weight models at a fraction of the cost and more proprietary in-house tools, while faster China-based model progress adds competitive pressure. The comment follows Palantir’s expanded partnership with Nvidia to build custom models for U.S. government agencies.

Analysis

The important market shift is not “AI is expensive” — it is that buyers are moving from novelty spend to procurement discipline. That changes the revenue mix for the model layer: broad, metered API consumption is more vulnerable than embedded, workflow-specific deployments where the customer owns the stack and the vendor earns on software, integration, and inference infrastructure. In that regime, PLTR is structurally better positioned than the pure frontier-model names because it monetizes control points inside the enterprise rather than raw token throughput.

For NVDA, the risk is less about demand destruction than about mix. If customers move from frontier calls to smaller open-weight or custom models, GPU demand may persist but pricing power can compress at the margin, especially in inference-heavy workloads where efficiency matters more than peak benchmark performance. The second-order winner is the on-prem / sovereign AI ecosystem: systems integrators, deployment layers, and government contractors should see more budget share as CIOs try to own data, models, and governance.

The contrarian view is that the market may be underestimating how bullish cheaper models are for total workload adoption. Lower unit costs usually expand the number of use cases that clear ROI hurdles, so any near-term bearish read-through on AI spend can reverse once enterprise pilots convert into production. China’s faster progress raises a real competitive overhang, but for public-market positioning the bigger signal is that model differentiation is shortening — which favors firms with data access, workflow lock-in, or distribution moats over standalone model branding.

The key falsifier is whether enterprise AI budgets actually reaccelerate despite lower token prices. If NVDA commentary, hyperscaler capex, or PLTR government/enterprise bookings show no slowdown over the next 1-2 quarters, the “AI spend is peaking” thesis should be retired; if open-weight adoption rises but compute demand still grows, the apparent margin pressure on the ecosystem is likely a rotation rather than a contraction.

Key Decisions for Investors

• Maintain/add to PLTR on pullbacks over the next 1-3 months: the stock benefits from the shift toward owned, customized models and sovereign/government deployments. Risk/reward improves if the market keeps rewarding ‘AI software with control points’ while discounting generic model access.
• Hold NVDA but avoid chasing momentum here; use weakness to add rather than shorting. The near-term risk is multiple compression if investors extrapolate lower token pricing into lower GPU demand, but the 6-18 month setup remains intact if cheaper models expand total inference volume.
• Pair trade idea: long PLTR / short a basket of pure model-exposure names or AI-API proxies on any further token-cost scare. Thesis: procurement shifts toward workflow owners while standalone model economics become more commoditized; invalidated if model vendors show sustained enterprise monetization despite price pressure.
• Set a watch item on NVDA commentary around inference mix and enterprise on-prem demand. If management indicates customer demand is shifting toward smaller/custom models without a corresponding rise in unit volumes, that would be the first sign of margin pressure in the ecosystem.
• If a broad AI selloff emerges, treat it as a rotation signal rather than a de-risking signal: rotate capital toward infrastructure and application/control-layer winners rather than exiting the theme entirely.