Palantir sells the layer above the model

The company draws its own boundary

Palantir's FY2025 annual report describes four platforms, and its description of the AI one is unusually precise about what is not inside it. AIP, the filing says, provides "secure connectivity to third-party-provided large language models", alongside a development toolchain, end-user applications, and "a broad evaluations framework for governing AI workflows in production" ^[1]. The model belongs to someone else. What Palantir sells is the connection, the governance, and the place the connection lands — the Ontology, which the same filing describes as the "systematic mapping of data to meaningful context" and sets at the center of the whole product line: "These platforms, backed by Apollo and with Ontology at their heart" ^[1].

The chief executive gave the commercial logic of that boundary in the Q2 2026 release: "Their competitive advantage should never become the training data for future models" ^[3]. Read as positioning rather than as rhetoric, that is a claim about which layer of the stack holds the customer's trust — and it only works if the model layer belongs to a third party.

What the position earns

The financial statements in that same annual filing show what selling above the model looks like ^[2]. FY2025 revenue of $4,475.4 million produced gross profit of $3,686.3 million — an 82% gross margin — operating income of $1,414.0 million, and $2,134.5 million of cash from operations, or 48% of revenue. Two years earlier, FY2023 revenue was $2,225.0 million against operating income of $120.0 million. Revenue has doubled; operating income is roughly twelve times larger.

The acceleration continued into 2026. Q2 revenue reached $1.935 billion, up 93% year over year, with GAAP net income of $1.062 billion at a 55% margin and $1.216 billion of cash from operations. Full-year guidance was raised to between $8.150 billion and $8.158 billion ^[3].

The half of the business the story leaves out

Beneath that headline, the mix is not the one the commercial narrative implies. U.S. commercial revenue grew 149% to $764 million in the quarter — and U.S. government revenue, growing at a merely excellent 90%, was still the larger of the two at $809 million ^[3]. The fastest-growing line is not yet the biggest line. A reader who has filed Palantir under enterprise AI is at least one quarter early.

Where the moat sits, and the ceiling the filing puts on it

The 10-K names its own competitor in one sentence: "We are fundamentally competing with the internal software development efforts of our potential customers" ^[1]. That is a statement about substitution, and it locates the moat exactly where the product description does — not in a model, and not in patents, but in whether an organization would rather model its operations inside someone else's Ontology than build that layer itself.

The evidence that customers stay is in the risk factors: "Our top three customers by revenue for the year ended December 31, 2025, have been with us for an average of ten years as of December 31, 2025" ^[1]. Ten years is a long time to remain inside a system you were always free to replace.

The same section supplies the other half of the picture. Those three customers were 16% of FY2025 revenue ^[1]. Many of the company's customer contracts "may be terminated by the customer at any time for convenience" ^[1]. And the competition discussion concedes that some rivals have "larger and more mature intellectual property portfolios" and "substantially greater financial, technical, and other resources" ^[1]. A ten-year deployment that can be cancelled at will is a strong position resting on continued preference rather than on contract.

The counterparties arrive as suppliers of capability

Two of the relationships recorded against Palantir in this site's graph are partnerships with companies that own precisely what AIP says it does not. NVIDIA describes embedding its CUDA-X libraries, Nemotron models and GPU computing into Palantir's Ontology and AIP ^[4]. Anthropic and AWS made Claude models available inside AIP for U.S. government intelligence and defense work ^[5]. In both, the capability flows inward and the customer relationship stays where it was — which is the arrangement the boundary was drawn to produce.

What would change the answer

The position depends on the model layer remaining a commodity input that someone else funds. Nothing in the cited record dates, or even addresses, the point at which a model provider might move up into the ontology and governance layer itself. The filing does not discuss it, and this article will not guess at it.

What the record does support watching is narrower and duller: whether U.S. commercial revenue passes U.S. government revenue, and whether ten-year tenure holds among customers whose contracts can end at any time for convenience ^[1].