In a Cloud Wars Minute commentary, Cloud Wars founder Bob Evans distilled Karp’s pitch into three points: enterprises want “AI sovereignty,” Palantir is growing with a small and declining sales organization, and the company seeks payment tied to customer value rather than a conventional catalogue price. Those points track Palantir’s August 3 earnings release and call, but they should not be mistaken for a simple three-step template that a CIO can copy.
Palantir’s reported quarter is real, and unusually large by enterprise-software standards. Its own earnings release put total U.S. revenue at $1.573 billion, up 115%, with U.S. government revenue at $809 million and U.S. commercial at $764 million. The commercial business is therefore close to, but still smaller than, the government segment in quarterly revenue. The more consequential number may be $2.132 billion in U.S. commercial total contract-value bookings, up 153% from a year earlier, because that shows customers signing larger commitments rather than merely expanding prior contracts.
What Karp is selling, however, is a model of implementation as much as a model of AI.
AI sovereignty is a control argument, not a deployment checklist
Karp and Palantir’s executives have framed AI sovereignty as an enterprise’s ability to retain control of its data, metadata, business processes, security policies, reasoning traces, and model choices. In the company’s formulation, a business should be able to use a frontier model, an open-weight model, or several models without allowing any one AI provider to become the owner of the company’s operational knowledge.
That concern is legitimate for organizations running regulated, sensitive, or strategically important workloads. A manufacturer does not merely have product specifications in its data estate; it has supplier relationships, quality exceptions, maintenance histories, inventory constraints, pricing rules, and the decisions staff make when those systems conflict. A hospital, bank, utility, defense contractor, or public agency has comparable layers of sensitive context. Sending selected prompts to a model provider is one thing. Building an AI system that continuously reaches into the operational record and can trigger or recommend actions is another.
The important technical distinction is that model access alone does not create a controllable enterprise system. A model needs identities, data permissions, audit trails, workflow boundaries, evaluations, rollback procedures, and a way to map natural-language requests to real business objects and approved actions. Palantir calls its representation of those objects and relationships an ontology. Microsoft customers will recognize the broader category even if they use different tooling: it is the layer that connects AI services to governed data, business applications, and authorization policies.
Palantir’s claim is not that it invented data governance. Microsoft, SAP, ServiceNow, Salesforce, Databricks, Snowflake, and cloud providers all have competing approaches to governed AI integration. Its sharper claim is that many enterprises will reject an AI architecture centered on buying tokens from a model vendor and leaving the operational layer fragmented underneath.
That is a credible argument, but it is also an argument for Palantir’s product stack. “Sovereignty” does not automatically mean deploying Palantir Foundry or its Artificial Intelligence Platform. An IT organization can pursue the same control objectives with a combination of its existing identity platform, data catalog, integration tools, private networking, model gateway, logging, and application-development processes. The test is not the vendor’s label. It is whether administrators can answer basic questions: which model processed which data, under what permissions, what action resulted, how was it evaluated, and can the organization reproduce or reverse that action?
A small sales force does not mean a low-touch sale
Cloud Wars highlighted Karp’s assertion that Palantir has produced rapid growth with a “minuscule and shrinking” sales headcount. Palantir made a similar point in its first-quarter shareholder letter, saying its sales headcount was smaller than it had been two years earlier. The company has also repeatedly emphasized that customers buy its platforms because the product demonstrates value, not because a large conventional sales organization pushes licences through a channel.
There is a meaningful lesson here for software companies: AI products that cannot demonstrate a specific operational result are becoming harder to sell. Enterprises have spent the past two years piloting chatbots, copilots, retrieval systems, agents, and model APIs. The question increasingly being asked in budget reviews is not whether employees used an AI tool, but whether the tool improved throughput, reduced errors, shortened a planning cycle, increased revenue, or reduced a measurable risk.
Palantir’s sales motion is still far from self-service. Its delivery model relies heavily on technically skilled teams working alongside customers to map data, model workflows, and put systems into production. That is fundamentally different from a vendor serving millions of small customers through a credit card, a product-led trial, and a dashboard. The company may have fewer traditional account executives, but it has not removed the expensive work of implementation.
For enterprise buyers, this is the part of the Palantir story worth examining closely. A smaller sales team can be evidence of product pull. It can also shift the cost and complexity from sales compensation to embedded engineering and deployment work. Neither arrangement is inherently better. The right question is whether the vendor can show how much of that work will be needed after the initial deployment and whether the customer’s own administrators and developers will be able to operate the resulting system without permanent vendor dependence.
Palantir’s Q2 figures suggest that enough customers are accepting that model to generate outsized growth. Its U.S. commercial customer count reached 653, up 35% year over year, while net dollar retention reached 157%. That retention figure means the existing customer base, in aggregate, was spending substantially more than it did a year earlier after accounting for churn. It is a powerful result, but it should be read as a commercial indicator, not proof that every deployment yielded a positive return.
Value-based pricing shifts the argument to measurement
Karp’s third point is the most interesting and the hardest to verify from the outside. He has said Palantir seeks to be paid as a derivative of the value it creates rather than following a simple price-list model. In theory, that aligns the supplier with the buyer: if the software does not produce the agreed result, the supplier should not capture the expected upside.
That arrangement sounds attractive because enterprise AI projects often fail in a familiar way. A company buys infrastructure, licences, model capacity, integration services, and consulting time, then declares success because a pilot reached production. The business case may never establish whether the tool changed any outcome that mattered. A value-linked agreement forces both sides to identify that outcome before the rollout begins.
But “value” is not self-defining. A warehouse optimization system might claim savings from lower stockouts, fewer expedited shipments, better labor utilization, or lower working capital. A sales agent might claim incremental revenue. A security system might claim prevented incidents that never occurred. Each requires a baseline, a time window, a control or comparison group where possible, and an agreement about what factors outside the software could have caused the result.
For CIOs and procurement teams, the contract mechanics matter more than the slogan. If a vendor proposes value-based pricing, the agreement should specify:
- The operational metric and pre-deployment baseline used to calculate the claimed improvement.
- The data source that determines the result, who can audit it, and how disputes are resolved.
- The distinction between software fees, implementation services, cloud infrastructure consumption, and model-inference costs.
- The customer’s rights to export data, logic, configurations, audit records, and workflow definitions if the relationship ends.
- The treatment of benefits created by changes in staffing, pricing, demand, market conditions, or other systems outside the vendor’s control.
Without those terms, “paid for value” can become another version of opaque enterprise pricing. With them, it can be a useful discipline that prevents an AI rollout from becoming a costly demonstration project.
Palantir’s quarter supports the thesis, but does not settle the market
Palantir raised its full-year 2026 revenue outlook to roughly $8.15 billion to $8.158 billion and increased its U.S. commercial revenue expectation to more than $3.424 billion. It also reported GAAP net income of $1.062 billion and adjusted free cash flow of $1.220 billion in the quarter. Those results establish that its approach is producing exceptionally strong demand and profitability at the moment, particularly in the United States.
They do not establish that enterprise software has converged on a single architecture or business model. The company’s most visible success is concentrated in U.S. commercial and government customers, while the broader enterprise market remains divided among cloud-native platforms, packaged business applications, private AI deployments, and model-provider ecosystems. Palantir also has the advantage of long experience connecting data to high-consequence operations, an advantage that cannot be reduced to a short sales playbook.
The useful takeaway is more concrete than Karp’s revolution language: enterprise AI spending is moving toward systems that can prove operational results and give customers meaningful control over data, permissions, model selection, and actions. Palantir’s Q2 numbers show that this proposition is selling. The burden now falls on every competing vendor—and on IT leaders signing the contracts—to prove that the promised control and value exist after the pilot ends.
