Palantir Technologies is the pre-eminent Western supplier of decision-layer software to institutions whose failure modes are measured in lives, sovereignty or systemic loss rather than merely in revenue. Founded in 2003 to serve the United States intelligence community, it has built a vertically integrated stack — Gotham for defence and intelligence, Foundry for the enterprise, Apollo for continuous delivery into any environment, and AIP for governed generative and agentic artificial intelligence — bound together by the Palantir Ontology, a semantic model of an organisation's data, logic and permitted actions. That architecture is Palantir's structural differentiator: it converts frontier-model capability into auditable operational output while leaving proprietary customer data inside the customer's perimeter. The commercial consequence is unusual: in FY2025 Palantir grew revenue 56% to $4.475 billion while generating a 50% adjusted operating margin, and it accelerated to 93% year-on-year growth in Q2 2026. The offsetting reality is a valuation that discounts a decade of flawless execution.
| Metric | FY2021 | FY2022 | FY2023 | FY2024 | FY2025 |
|---|---|---|---|---|---|
| Full-time employees (headcount) | 2920 | 3838 | 3735 | 3936 | 4429 |
| Share of employees outside the US (%) | not disclosed | 37.4 | 35 | 31 | 28 |
| Revenue per full-time employee (USD thousands, derived) | 528 | 497 | 596 | 728 | 1010 |
*Source: Forms 10-K FY2022 (3,838 employees, of which 1,436 outside the US), FY2023 (3,735, 35% outside the US), FY2024 (3,936, 31% outside the US), FY2025 (4,429, 28% outside the US). FY2021 headcount of 2,920 is from aggregator data and is not directly quoted from the FY2021 10-K in this research pass. Revenue per employee is a derived calculation using year-end headcount.*
---
### The company's own characterisation
In Item 1 of the FY2025 Form 10-K, Palantir describes itself in a single line: it builds software that lets organisations integrate their data, decisions and operations at scale. The filing recounts that the company was founded in 2003 and began by building software for the US intelligence community to support counterterrorism investigations and operations, later extending to commercial enterprises facing structurally similar data problems. The 10-K states that Palantir has built four principal software platforms — Gotham, Foundry, Apollo and AIP — and that Foundry is the foundational data-operations platform providing data management, logic authoring, systemic mapping through the Palantir Ontology, analytics and workflow development. AIP is described as the generative-AI platform providing secure connectivity to third-party large language models, a toolchain for building AI agents and automations, AI-enabled end-user applications and an evaluations framework for governing AI workflows in production. Apollo is characterised as a cloud-agnostic continuous-delivery control layer, and Gotham as the platform that integrates with the broader defence offerings to power allied defence and intelligence missions.
The 10-K is explicit and repeated on a point the company evidently regards as commercially load-bearing: Palantir does not collect, mine or sell data. It licenses software platforms that let customers integrate data they already own, and the customer remains the legal owner of that data. This positioning is restated at length in the Board's 2026 proxy statement opposition to shareholder Proposals 4 and 5, where the Board states plainly that Palantir is a software company, not a data or surveillance company, and that its business model is not based on monetising personal data.
### Independent characterisation
Palantir is best understood not as an analytics vendor but as a supplier of institutional operating systems sold on a productised licence, with an unusually heavy forward-deployed services wrapper that is priced into the software rather than billed separately. Three structural features distinguish it from the enterprise-software comparison set:
**First, the unit of sale is the organisation, not the seat.** Palantir does not compete for departmental budgets against best-of-breed point tools; the 10-K states that the company is fundamentally competing with the internal software-development efforts of its potential customers, and that it deliberately targets large-scale, hard-to-execute opportunities where high installation costs, high failure risk, complex data environments and long sales cycles raise barriers to entry. The filing goes further and states the counterintuitive rule that the larger, more complex and more technologically demanding the problem, the more likely Palantir is to succeed. This is an explicit strategy of adverse selection into difficulty.
**Second, the revenue model is subscription-like but contractually softer than SaaS.** Contract terms are generally one to five years; revenue is generally recognised rateably over the term. Critically, the 10-K discloses that the majority of contracts contain termination-for-convenience provisions, and that the US federal government is prohibited from exercising contract options more than one year in advance. This is why Palantir reports three separate backlog measures with materially different quality: total contract value (TCV, the potential lifetime value at execution, assuming all options exercised), remaining deal value (RDV, the remaining potential value of live contracts) and remaining performance obligations (RPO, the non-cancellable contracted revenue not yet recognised). At 30 June 2026 these stood at $13.1 billion RDV and $4.9 billion RPO — a roughly 2.7x gap that quantifies how much of the headline backlog is optional rather than committed.
**Third, the delivery model is engineer-led rather than partner-led, though this is changing.** Palantir's research and development function is described in the 10-K as embedding directly with users; the same forward-deployed engineers who deliver also generate the product roadmap. Recent R&D emphasis is on deploying software, models and critical assets at the edge, including integrations with complex hardware and operation in disconnected environments. The FY2025 evidence that this is working is stark: revenue grew 56% while headcount grew 12.5%, lifting revenue per employee from roughly $728,000 to roughly $1.01 million.
### Revenue model composition
Palantir does not disaggregate revenue into licence, subscription, hosting and services line items in its income statement; it reports a single revenue line with a single cost of revenue. The disclosed composition drivers are:
**Platform subscriptions:** access to Gotham, Foundry, AIP and Apollo, generally recognised rateably.
**Operations and maintenance (O&M) services:** described in the 10-K as the ongoing service the customer depends on post-deployment; sold alongside the platform rather than as a separate labour contract.
**Professional services / deployment:** bundled into productised pricing. The 10-K contains a dedicated risk factor on the possibility that customers may be unwilling to accept a product-based model in place of the labour-contract model with which they are historically familiar.
**Non-cash consideration:** Palantir has entered arrangements involving receipt of equity or other non-cash consideration, primarily in the historical "Strategic Commercial Contracts" cohort (special purpose acquisition companies and related counterparties). FY2025 non-cash consideration was $37.2 million, down from $52.5 million in FY2024, and this cohort is now immaterial to growth.
**Developer Tier:** a limited-access self-service tier for Foundry and AIP available in the United States and selected countries, designed to widen top-of-funnel access without enterprise-level upfront cost.
Pricing is not published. The 10-K carries an explicit risk factor stating that Palantir has changed pricing models in the past and expects to do so again, and that larger organisations may demand different pricing structures or substantial concessions.
### Customers and end markets
As of 31 December 2025 Palantir had 954 customers, up from 711 at 31 December 2024 — a 34% increase. By 30 June 2026 the total had reached 1,049, up 24% year-on-year, of which 653 were US commercial customers (up 35% year-on-year and 6% sequentially).
Concentration remains material but is improving. The top three customers together accounted for 16% of FY2025 revenue, versus 17% in FY2024 and 18% in FY2023. The average trailing-twelve-month revenue from the top twenty customers was $93.9 million in FY2025, up 45% from $64.6 million in FY2024 and $54.6 million in FY2023; by Q2 2026 this had reached $124 million per top-twenty customer, up 67% year-on-year. Palantir's top three customers by FY2025 revenue had been customers for an average of ten years.
End markets served, as enumerated in the FY2025 10-K and in AIPCon customer rosters, include: defence and intelligence; law enforcement and public safety; healthcare and life sciences; energy and utilities; oil and gas; automotive and industrial manufacturing; aerospace and shipbuilding; airlines and aviation; insurance; banking and capital markets; telecommunications; logistics and supply chain; agriculture; retail; media; and civil government (health, welfare, immigration, tax, disaster relief). The company states it is building "industry operating systems" in airline, space, shipbuilding, insurance, healthcare, telecommunications, automotive, security and risk management, and government.
### Value-chain position
Palantir occupies the integration and decision layer between (a) the underlying compute and storage substrate — hyperscale clouds, sovereign clouds, on-premise infrastructure and increasingly Nebius-supplied sovereign capacity — and (b) the frontier and open-weight model layer supplied by third parties. It is deliberately model-agnostic: AIP provides unified access to open-source, self-hosted and commercial large language models. This position is defensible precisely because it is not the compute layer and not the model layer; Palantir monetises the translation of model capability into governed operational action, which is the layer where the customer's proprietary process knowledge lives and where switching costs are highest. Management has articulated this as its central competitive thesis: at Q2 2026 the company argued that customers' competitive advantage should not become training data for future models — the sharpest available statement of why Palantir believes the ontology layer, not the model layer, captures the durable economics.