Palantir's 96% Growth vs AI's Token Illusion: Ontology-Driven Enterprise Value

Palantir's 96% Revenue Growth and the Token Economy Critique

In Q3 2026, Palantir reported revenue of $2.313 billion, up 96% year-over-year and 20% quarter-over-quarter, with GAAP net income of $1.128 billion. U.S. commercial revenue reached $984 million, up 148% year-over-year, surpassing U.S. government revenue ($953 million) for the first time. Full-year revenue guidance was raised to approximately $8.53 billion.

Despite the strong results, CEO Alex Karp used the earnings call to criticize a segment of the AI industry. He argued that some software vendors rely on raw token usage, misleading deployment metrics, and reselling tokens and booking them as revenue to fuel growth. Karp contends that enterprise customers are growing weary of AI products that cannot demonstrate tangible business outcomes. He does not dispute the contributions of foundation model companies; rather, he questions whether application-layer vendors can create sufficient commercial value beyond model capabilities.

The critique highlights a gap in enterprise AI adoption: increased model calls and higher employee usage frequency indicate adoption, but these metrics remain distant from actual business revenue . An agent processing tens of thousands of requests daily may only accelerate existing tasks; if purchase approvals still take days, inventory overstock persists, and customer issues remain unresolved, AI scale does not directly translate to operational efficiency.

As foundation model quality improves and access costs drop, merely wrapping a large model in a chat interface or reselling model inference becomes an increasingly weak competitive moat. Palantir's differentiation strategy is to embed AI deeply into existing business systems so it participates in complete workflows , proving value through measurable business results.

Why the Enterprise AI Unit Should Be the Workflow, Not the Knowledge Worker

CTO Shyam Sankar stated: "The fundamental unit of the enterprise is not the knowledge worker, but the workflow." This challenges the common approach of giving every employee an AI assistant (e.g., sales writing emails, finance analyzing spreadsheets). While such tools reduce repetitive work, business processes typically span multiple departments, systems, and approval steps. Efficiency gains at a single role do not guarantee end-to-end process improvement.

Example: In supply-chain procurement, AI can quickly identify material shortages and generate replenishment recommendations. However, execution still requires cross-checking inventory, production schedules, supplier lead times, budgets, and approval authorities. Even if analysis time drops from two hours to ten minutes, the overall procurement cycle barely changes if subsequent steps remain manual across disconnected systems. The enterprise needs to solve the continuous process from risk detection, solution formulation, to approval and execution .

Therefore, enterprises need an environment that simultaneously orchestrates models, business software, and people . Traditional agent harnesses manage prompts, context, tool calls, and execution state; in an enterprise setting they must also handle cross-system data access, business rules, human approvals, and operational permissions. Palantir AIP is built along this line, embedding AI capabilities into existing business processes.

Figure 1: Onyx Enterprise OS layered architecture: data sources, logic sources, and execution systems converge at the Ontology layer, supporting analytical workflows, agent automation, and product SDKs

AIG's commercial insurance underwriting illustrates the approach. Complex commercial policies involve document review, risk analysis, rule validation, and human judgment — beyond a single model Q&A. Palantir and AIG integrated AI, business software, and human decisions into an AIP workflow, so the model participates in concrete underwriting tasks. AI value is no longer limited to helping underwriters read faster; it depends on whether the entire underwriting process increases throughput, shortens cycle time, and maintains necessary risk controls .

When AI enters cross-departmental workflows, new challenges arise: how the model understands internal data relationships, knows which operations comply with business rules, and ensures execution safety and reliability. These cannot be solved by simply adding prompts or tool interfaces.

How Ontology Lets AI Truly Understand Enterprise Business

Enterprises have abundant data — ERP for orders and inventory, CRM for customers, manufacturing systems for equipment and work orders, finance for budgets and settlements — but these systems were built for departmental needs, with differing data structures, governance rules, and access permissions. Even if a model can access all data, it lacks a unified business context to interpret relationships accurately.

Example: A single order links a customer, product, inventory, supplier, and delivery plan. When the order changes, the business must determine affected production tasks, whether procurement quantities need adjustment, and who approves related actions. These rules are scattered across databases, application code, and employee experience. Without a unified business context, a model cannot reliably perform cross-system analysis, let alone participate in execution.

Palantir's Ontology addresses this by building a unified business representation on top of existing data assets, using Object, Property, and Link to describe business entities, attributes, and relationships. Entities like orders, customers, equipment, and suppliers become objects with explicit business meaning; data from different systems can be joined around these objects. When an agent queries an order, it can traverse to related inventory, suppliers, and production plans, understanding the current business state and its impact radius.

Ontology goes beyond data linking. Its architecture organizes Data, Logic, Action, and Security together: Data maps heterogeneous system information into business objects; Logic encapsulates calculation rules, optimization algorithms, and model capabilities; Action defines executable business operations; Security enforces data access and operation permissions.

Figure 2: Palantir Ontology system architecture: Data, Logic, Action, and Security layers connecting upward to applications and agents, downward to data services, logic services, and workflow services

In the procurement scenario, when the system detects a material that may affect production delivery, the agent uses Ontology object relationships to locate relevant orders, inventory, and suppliers, then invokes business logic to evaluate replenishment options. If a purchase quantity change or approval is needed, the agent must execute via the corresponding Action under established permission constraints . Analysis, decision, and execution revolve around the same business objects, reducing repeated information translation across systems.

Figure 3: Ontology Actions open to both humans and agents: Data, Logic, Actions via functions, rules, and write-back operations to ERP, SCM, MES systems

Another key role: enabling continuous reuse of data generated during business execution . The system records not only order status and execution results, but also human corrections, business feedback, and decision rationale associated with tasks. After curation and governance, this data can evaluate model performance and serve as training material for subsequent fine-tuning. Ontology thus provides both business context management and a data foundation for continuous model adaptation to enterprise tasks.

Models Need Not Grow Ever Larger; Business Feedback Becomes the New Training Resource

Following the Ontology-and-workflow path, Palantir advances model post-training. Shyam Sankar argues that many economically valuable enterprise tasks do not require the most capable frontier models . For tasks with well-defined scope and stable evaluation criteria, fine-tuning smaller open-weight models on proprietary enterprise data can yield a better cost-performance trade-off.

L3Harris case study: The U.S. defense contractor previously integrated 30+ ERP systems via Palantir and built numerous business applications around Ontology. In an internal pilot, L3Harris fine-tuned an open model on its own business data and, in under 48 hours, achieved superior performance on a specific task compared to the previously used frontier model, with 95% cost reduction . This result comes from vendor-disclosed preliminary testing, scoped to a concrete business task, and does not imply small models surpass frontier models in general capabilities.

McCarthy Building Companies case study: The construction firm applied post-training to engineering document processing. Requests for Information (RFIs) involve extensive construction requirements, project constraints, and risk data; subcontract drafting also demands precise understanding of site conditions. McCarthy used over 100,000 actual project RFIs for model development and leveraged Ontology-annotated data to fine-tune NVIDIA's Nemotron 3 Nano 4B model for engineering risk identification and contract clause drafting. Per Palantir's disclosed expectations, the solution could save roughly 30,000 hours of related work annually.

NVIDIA case study: NVIDIA uses a similar method to optimize supply-chain tasks. After organizing business processes and data on the Palantir platform, NVIDIA evaluates and post-trains open models. According to earnings-call disclosures, on specific internal tasks the optimized smaller model outperformed the previously used larger model and can complete a new training round on a single GPU in hours. Compared to relying on continuous general-model upgrades, this approach lets enterprises adjust model capabilities in step with their own business changes.

These cases share a common technical loop: enterprises generate data and feedback during business execution; Ontology organizes state and records tied to business objects; after filtering, annotation, and governance, data enters the post-training pipeline; the new model is evaluated and deployed back into the workflow, where fresh business feedback validates the improvement. Palantir is increasing automation of this cycle through tools like AIP Evolve, enabling the system to continuously adjust model selection, prompts, and workflow configuration around explicit task objectives .

Figure 4: Unified model-instance replacement in AIP: switching to GPT-5.4 nano simultaneously updates sentiment analysis, translation, and summarization logic nodes

The mechanism's impact depends heavily on enterprise data quality and evaluation rigor. Business feedback may contain errors; human corrections are not always correct. Model optimization must balance quality, cost, and latency while avoiding local task gains that degrade overall flow. For enterprises, post-training's value lies in converting long-accumulated business experience into reusable model capabilities while retaining control over training data, evaluation standards, and deployment processes.

From $1M to $15M: Why Enterprises Keep Expanding AI Investment

Technical architecture must ultimately prove commercial value, visible in customer willingness to keep investing. Palantir's Q3 disclosures on several contract expansions offer a window into commercialization progress.

A global appliance company started a $1M AI pilot in Q2, expanded to a $3M annual contract in the same quarter, and grew to $15M in Q3, covering commercial, finance, and supply-chain workflows.

An industrial materials company grew a near-$3M Q2 pilot into a $29M annual contract in Q3.

Such projects typically begin with a concrete business problem. Once a workflow demonstrates results, the existing data connections, business objects, and execution mechanisms can be reused to extend AI to other departments. Compared to rebuilding systems for each new scenario, a unified Ontology and AIP platform can reduce redundant integration and development costs , making it easier to scale AI from isolated pilots to multiple business domains.

This expansion pattern appears in Palantir's operating metrics: Q3 closed 212 deals worth at least $1M, 64 of them $10M+. Net dollar retention reached 165% , reflecting sustained growth within existing accounts. However, contract value does not equal realized economic benefit; AI's actual value creation must still be measured by cost savings, revenue growth, throughput improvements, and risk reduction.

Strong growth coexists with cautionary signals. Q3 total new contract value was $2.918B, up 6% year-over-year; U.S. commercial new contract value was $1.463B, up 11%, significantly below the prior quarter's ~153% year-over-year growth. Revenue, customer expansion, and new bookings are not moving in lockstep, indicating that while enterprise AI demand remains robust, project expansion pace and future revenue growth cannot be extrapolated solely from current results.

Palantir's technical approach also faces practical challenges. Ontology requires ongoing maintenance of accurate business objects, relationships, and operational rules; legacy systems and organizational boundaries in complex enterprises do not vanish simply because agents are introduced. Post-training likewise depends on stable data feedback, reliable evaluation frameworks, and sustained engineering investment. As cloud providers, data platforms, and model companies strengthen their enterprise agent offerings, Palantir must still demonstrate that its technology stack can reliably deliver tangible business outcomes across more industries and scenarios .

Conclusion

Karp's criticism of the token business ultimately points to an increasingly realistic problem in enterprise AI commercialization: model usage scale and the economic value enterprises ultimately capture do not have a simple correspondence .

Palantir is attempting to bridge that gap by using Ontology to connect data with business rules, using workflows to put models into actual operations, and then using business feedback to continuously improve models. This method still needs more long-running case studies for validation, but it already reveals an important direction in enterprise AI technology evolution.

As large models become commoditized foundational capabilities that enterprises can procure on demand, competition shifts deeper into business systems. Whether an enterprise's long-accumulated business knowledge, process rules, and execution experience can be organized into understandable, invocable, and continuously improvable intelligent capabilities will directly determine the return on AI investment.

Future enterprise AI competition may not occur only on model leaderboards, but in order processing, production scheduling, risk control, and every other real-world business workflow that actually runs .