Palantir CEO Alex Karp is wrong about the threat Anthropic and OpenAI pose to most enterprises. That doesn’t mean he doesn’t have something to lose

The article centers on Palantir CEO Alex Karp’s CNBC critique of frontier AI labs, arguing enterprises are “chillaxing” on token costs without getting value and that frontier vendors risk taking enterprise IP. The piece pushes back that ROI is real but uneven—often tied to use-case selection and workflow redesign—and notes that major model providers say they do not train on enterprise prompts/outputs unless customers opt in (especially via secure cloud routes like Azure/Bedrock/Vertex). It also flags limited scenarios where “design partner” access could enable competitive outcomes (e.g., Anthropic/Figma), while broader AI safety standards are reported to be slipping and China is considering restricting foreign access to leading AI models.

Analysis

The near-term read-through is mostly sentiment, not earnings: the market is more likely to punish names priced on AI narrative purity than those with clear utility capture. The cleanest mechanism is multiple compression in PLTR if investors conclude its moat is less about exclusive “data control” and more about workflow integration, which is harder to defend at premium valuation. By contrast, the secure-cloud distributors of enterprise AI access — MSFT, AMZN, and GOOGL — should keep absorbing demand because they own the compliance boundary customers actually trust.

Second-order, the article reinforces a bifurcation in AI spend: model usage may stay scrutinized, but infrastructure and governance budgets should keep rising. That favors NVDA at the hardware layer and the hyperscalers at the deployment layer, while pressuring any vendor whose upside depends on customers believing it can both sell and safely sandbox their strategic data. The biggest risk is that this turns into a procurement delay rather than a spend cancellation; if CIOs freeze pilots for 1-2 quarters, the pain shows up first in high-multiple software and only later in cloud consumption.

The contrarian miss is that “IP theft” is probably the wrong lens; the real issue is vendor substitution. Enterprises may not fear model training on their prompts, but they do fear becoming dependent on a single app layer that can be displaced by the model provider or hyperscaler over 6-18 months. That’s why the most durable winners are likely the control points — cloud, identity, security, and GPUs — not the branded frontier model vendors or application-layer names with the richest narrative premium.

Key Decisions for Investors

• Short PLTR vs long NVDA on a 1-3 month horizon: use any post-rally strength in PLTR to lean into multiple compression risk, while NVDA retains direct monetization from sovereign AI and enterprise compute demand. Falsify if PLTR prints accelerating AIP revenue or NVDA commentary shows order deceleration.
• Add MSFT/AMZN/GOOGL on pullbacks as the preferred AI deployment trade for 3-6 months: these names own the secure distribution layer that enterprises will route through if they want AI without exposure to model-vendor data concerns. Risk/reward is better than owning standalone model narratives.
• Treat FIG as a tactical short watchlist only if liquid exposure is available elsewhere in your book: the design-partner conflict risk is a warning that vertical AI collaborators can become competitors. Invalidates if management demonstrates partner-friendly product boundaries and no evidence of lost pipeline.
• Reduce overexposure to high-multiple AI software names that depend on “AI premium” rather than verifiable workflow ROI; reallocate toward infrastructure beneficiaries. The catalyst is the next round of enterprise budget reviews over the next 1-2 quarters.