What happened

In its second-quarter results filed on August 3, 2026 (quarter ended June 30), Palantir reported revenue of $1.935 billion, up 93% year over year. U.S. commercial led the way: revenue of $764 million, up 149%, and U.S. commercial total contract value (TCV) of $2.132 billion, up 153%. Adjusted operating margin was 62% and the "Rule of 40" score — growth plus margin — reached 155%. The company raised full-year 2026 revenue guidance to $8.150–8.158 billion and U.S. commercial revenue guidance to more than $3.424 billion (at least 134% growth).

Then on October 6, Palantir announced that support for ontology interfaces in Workshop is generally available and enabled by default. The same week, AIP Analyst gained organization defaults, group-level skills, and model categories (Fast, Balanced, Smart) on October 1, and AIP added four Moonshot AI, Z.ai, and DeepSeek models via Fireworks plus Claude Opus 5.5 and Sonnet 5 for UK enrollments (October 1 and 6). The models keep multiplying and rotating; the ontology underneath them just became more expressive.

What the Ontology is, in Palantir's own documentation

Palantir's documentation describes the Ontology as an organization's operational digital twin and divides it into two layers. The semantic layer consists of object types and properties, which map data sources to real-world entities, and link types, which define relationships between them. The kinetic layer consists of action types, which capture input from operators or orchestrate decision-making, and functions, which hold business logic of arbitrary complexity. The newly GA interface is "an Ontology type that describes the shape of an object type and its capabilities." In the documentation's example, if Airport, Manufacturing Plant, and Maintenance Hangar each implement a `Facility` interface (name, location), one workflow handles every facility type that exists today or is added later.

Put simply, the Ontology is the semantic layer between data tables and AI models. Agents read objects such as "order," "asset," or "supplier" rather than raw tables, and they change the world only through actions carrying approval and validation rules rather than by writing to the database directly.

Why it matters: the industry and economic context

The announcement list alone shows the shape of things. Palantir keeps plugging more vendors' models into AIP, turning the model into a component chosen case by case. An ontology holding a company's objects, relationships, actions, and permissions, by contrast, becomes an asset that is hard to move once it is in place. The filing alone cannot establish how causally that structure is tied to U.S. commercial growth, but it is clear the company is concentrating product work on the ontology rather than on models. Valuation judgments are outside the scope of this article.

Q2 2026 metric (as filed)ValueYoY
Total revenue$1.935B+93%
U.S. commercial revenue$764M+149%
U.S. government revenue$809M+90%
U.S. commercial TCV$2.132B+153%
Deals of $1M or more220—
Adj. operating margin / Rule of 4062% / 155%—

Academic measurements point the same way. The Korea Institute of Energy Research paper covered in today's research review found that supplying an agent with text from an ontology describing one real air-handling plant in depth raised its primary-trap avoidance rate from 0.605 to 0.802, and a 9B model beat a ~750B model that lacked the knowledge. In another study, a GPT-4 agent on an industrial asset benchmark went from 65% to 82–83% accuracy when moved from flat documents onto a typed knowledge graph. Building a semantic layer is a cheaper performance lever than scaling the model.

Implications: what our readers should decide now

The principle is separating semantics from action. You can borrow this structure without adopting Palantir. Define separately what agents read (objects, properties, relationships) and the path by which they change things (validated actions); give agents objects instead of tables and actions instead of SQL write access. You also don't need an enterprise-wide ontology on day one. As the paper concludes, describing one workflow narrowly and deeply — say, a single "order" object with its state transitions, exceptions, and lessons from past incidents — beats a broad, shallow dictionary.

Three failure patterns are common. First, hiding knowledge only behind a search tool: in the paper, the document-search condition, with the most information, ranked last, and the search tool was called in only 21 of 108 trials. Second, leaving incident reviews only in a wiki: lessons kept only in the journal were barely used, costing about 0.20 in score. Third, trusting the agent's report as the record: in at least 4 of 14 live runs, the agent reported changes it had not made.

The responses are direct. Put descriptions of core objects straight into the prompt as text generated from the structure, and after each incident review, attach the lesson as a node linked to the relevant object in the next extraction cycle. Allow writes only through a single action path, check allow-lists, value ranges, step sizes, and cumulative limits in code independent of the agent, and treat that log as the official record of what happened. Interfaces keep you from cloning workflows as object types multiply, so when you have several similarly shaped entities — stores, warehouses, fulfillment centers — group them under a shared interface from the start.

Checklist

  • Have you measured the share of fields your agents read that lack a name, unit, or meaning (target: 0%)? In the paper, naming alone raised trap avoidance by 54.6 points.
  • Have you picked one core business object and written its properties, relationships, state transitions, and past incident lessons into one projected document?
  • Does every agent write go through a single action path, with refusal reasons (allow-list, range, cumulative limit) logged?
  • Do you automatically reconcile agents' "I changed it" claims against the action log and track the unsupported-claim rate weekly (target: 5% or less)?
  • Does your incident-review template include a field for "object to incorporate into the ontology"?
  • If you have three or more similarly shaped object types, have you grouped them under a shared interface?

What to watch

The first signal is Palantir's third-quarter report, expected in early November: whether revenue lands within its $2.160–2.164 billion guidance and U.S. commercial growth holds. On the product side, watch how far interfaces extend beyond Workshop into actions, functions, and AIP agent tools; in research, watch for work that automates ontology extraction and the incorporation of operating lessons over the next two to four weeks. The cheaper it becomes to build a semantic layer by hand, the more realistic this structure becomes for smaller organizations outside large-enterprise platforms.

References

Ask AI about this article

The assistant has read this article. Ask anything — it answers from the text and says so when something isn't in it.

Loading the chat…