Palantir's 96% Revenue Surge: Why Its CEO Attacks AI's Token Economy

Palantir's 96% Revenue Growth and Critique of Token-Based AI Models

In Q3 2026, Palantir reported revenue of $2.313 billion, up 96% year-over-year, with GAAP net profit of $1.128 billion. U.S. commercial revenue reached $984 million, up 148%, surpassing U.S. government revenue of $953 million for the first time. Full-year revenue guidance was raised to approximately $8.53 billion.

Despite the strong results, CEO Alex Karp criticized AI vendors that rely on raw token usage, superficial deployment metrics, and reselling token capacity to drive growth. He argued that customers are growing tired of AI products that cannot demonstrate tangible business value. The article notes that increased model calls and employee AI usage indicate adoption but do not directly translate to operational efficiency if core processes like procurement approvals or inventory management remain unchanged.

The Workflow as the Fundamental Unit of Enterprise AI

CTO Shyam Sankar stated that the enterprise's basic unit is not the knowledge worker but the workflow. The article illustrates this with a supply-chain procurement example: AI can shorten risk analysis from two hours to ten minutes, but the overall procurement cycle won't improve if subsequent steps still require manual cross-system queries and approvals. Enterprises need an environment that coordinates models, business software, and people simultaneously.

Palantir's AIP (Artificial Intelligence Platform) embeds AI into existing business processes. In a case with AIG's commercial insurance underwriting, AI, business software, and human judgment are integrated into an AIP workflow. The model participates in specific underwriting tasks, and value is measured by the entire process's throughput, cycle time, and risk control — not just faster document reading.

Ontology: Unifying Data, Logic, Action, and Security

Enterprise data resides in siloed systems (ERP, CRM, MES, finance) with different structures and rules. Palantir's Ontology creates a unified business representation using Objects, Properties, and Links. It organizes four layers:

Data : Maps disparate system information into business objects (orders, customers, equipment, suppliers).

Logic : Encapsulates calculation rules, optimization algorithms, and model capabilities.

Action : Defines executable business operations (e.g., modify purchase order, trigger approval).

Security : Enforces data access and operation permissions.

In the procurement scenario, an agent uses Ontology to locate related orders, inventory, and suppliers, evaluates replenishment options via Logic, and executes changes through Actions under permission constraints. This reduces context-switching across systems.

Ontology also captures execution data — human corrections, business feedback, decisions — which after curation becomes training data for continuous model improvement.

Post-Training: Smaller Models Tuned on Business Feedback

Sankar argues that many high-value enterprise tasks do not require the most capable frontier models. For well-scoped tasks with stable evaluation criteria, fine-tuning smaller open-weight models on proprietary business data can yield better cost-performance.

L3Harris : Integrated 30+ ERPs via Palantir, built Ontology-based apps. In an internal test, fine-tuning an open model on business data took <48 hours and outperformed the previous frontier model on a specific task, with 95% cost reduction . Results are preliminary and task-specific.

McCarthy Building Companies : Used 100k+ real project RFIs (Requests for Information) to fine-tune NVIDIA Nemotron 3 Nano 4B for engineering risk identification and contract drafting. Expected to save ~30,000 hours annually.

NVIDIA : Organized supply-chain processes and data on Palantir, evaluated and post-trained open models. On specific internal tasks, optimized smaller models beat larger predecessors and can be retrained on a single GPU in hours.

The common loop: business execution generates data and feedback → Ontology organizes state and records → curated data feeds post-training → new model evaluated and deployed back into workflow → new feedback validates improvement. Palantir's AIP Evolve automates model selection, prompt tuning, and workflow configuration.

Success depends on data quality, evaluation rigor, and balancing quality, cost, and latency. The value lies in converting accumulated business experience into reusable model capabilities while retaining control over data, evaluation, and deployment.

Customer Expansion: From $1M Pilots to $15M+ Contracts

Palantir's commercial traction is reflected in contract expansions:

A global appliance company: $1M pilot (Q2) → $3M annual contract (Q2) → $15M (Q3), covering commercial, finance, and supply-chain workflows.

An industrial materials company: ~$3M pilot (Q2) → $29M annual contract (Q3).

Q3 closed 212 deals worth ≥$1M, 64 deals ≥$10M. Net dollar retention reached 165%, indicating existing customers are significantly expanding spend. The unified Ontology and AIP platform reduces redundant integration costs, enabling AI to spread from initial pilots to multiple domains.

However, new bookings growth slowed: total new contract value $2.918B (+6% YoY), U.S. commercial new contract value $1.463B (+11% YoY), far below the prior quarter's ~153% growth. Revenue, expansion, and new bookings are not perfectly synchronized, suggesting future growth cannot be extrapolated linearly.

Challenges remain: Ontology requires ongoing maintenance of accurate business objects and rules; legacy systems and organizational boundaries persist. Post-training needs stable feedback loops, reliable evaluation, and sustained engineering investment. As cloud providers, data platforms, and model companies enhance enterprise agent capabilities, Palantir must prove its stack delivers consistent real-world results across more industries.

Conclusion

Karp's critique highlights a core issue: model usage volume does not linearly map to economic value. Palantir's approach — connecting data and business rules via Ontology, embedding models in workflows, and using business feedback for continuous post-training — represents a significant direction for enterprise AI evolution. As foundation models become commoditized infrastructure, competition will shift from leaderboards into the operational heart of businesses: order processing, production scheduling, risk control, and every real running workflow.