Palantir's software starts where most data projects stall: the data is spread across dozens of systems, and joining it is only half the job; the other half is acting on it. Its platforms do both.

Foundry pulls data from a company's systems into pipelines and maps it onto the Ontology, a model of the business as objects (plants, parts, orders, patients), the links between them and the actions people may take on them. Teams build apps on that model that write decisions back into the systems they came from. Gotham does the same for defense and intelligence: it fuses sensor and intelligence data into one picture for analysts and commanders, and passes their findings to the people who act on them. AIP, launched in 2023, connects large language models to the Ontology, so AI agents work on live data and take actions under the same permissions as a person. Apollo ships and updates all of it everywhere, from public clouds to classified networks and disconnected hardware.

The way Palantir sells is as distinctive as the software. It targets large, hard problems at large institutions, the projects where in-house builds and consultancies have already failed. Its engineers work on site with the customer (forward deployment), and it opens doors with free or low-cost pilots and AIP bootcamps that build a working workflow on the customer's own data in days. A contract starts with one use case and spreads across divisions; Palantir had 954 customers at the end of 2025 and grows mainly by widening what each one uses.

The customers are split between Government (US defense, intelligence and civilian agencies, and allies) and Commercial (manufacturers, energy companies, insurers, hospitals, pharmaceutical companies and others), the two segments Palantir reports. They pick it because it works sooner than the alternative, which is usually building a data platform in-house from Snowflake or Databricks, an orchestrator, a BI tool and a consultancy's time, and because its security model was built for classified data from the start.

Once a customer's operations run on the Ontology, Palantir is hard to remove. The pipelines, the object model, the apps people use every day, the actions that write back to SAP or a mission system, and now the AI agents all live inside Palantir's platform. Moving off means rebuilding each of them on another stack, retraining the users, and re-earning the security approvals, which on classified networks can take years.

The work done together adds to it. Palantir's engineers often spend years inside a customer's operations, and the knowledge of how the plant, the fleet or the unit actually works is encoded in the Ontology they built. Government customers add procurement: a new system means a new competition, and enterprise agreements across a military service consolidate many programs onto Palantir at once.

The alternatives a customer would move to are a lakehouse such as Databricks or Snowflake with its own apps and agents built on top, a defense contractor's or Anduril's systems, or the AI built into the software it already runs. Those are cheaper per unit and more open, and a customer that only ever used Palantir for analytics can leave more easily. The ones that stay are those whose daily operations run through it.

Palantir's software, once built, costs little to deliver, and its figures since 2022 show it.

Revenue was $4.48B in 2025, up 56.2%, after growth of 16.7% in 2023 and 28.8% in 2024. Over the five years to 2025 it grew about 32.6% a year, from $595M in 2018. Gross margin rose from 72.2% in 2018 to 82.4% in 2025: the cost of hosting and of the engineers on site grew more slowly than the contracts.

The operating margin tells the turn. It was negative 107.4% in 2020, the year equity pay reached $1.27B, more than the $1.09B of revenue. Palantir reached an operating profit in 2023 (5.4%), then 10.8% in 2024 and 31.6% in 2025, as revenue grew much faster than sales, research and administration costs. Net income was $1.63B in 2025, a 36.3% net margin.

It needs almost no capital: capital spending was $33.9M in 2025, under 1% of revenue, because the hardware is the cloud providers' and the customers'. Operating cash flow was $2.13B and free cash flow $2.10B (46.9% of revenue). Equity pay is still a large cost, $684M in 2025 or 15.3% of revenue; after it, free cash flow was $1.42B, 31.7% of revenue. Palantir ended 2025 with $1.42B of cash and $5.75B of short-term investments.

Palantir is building the next decade of its products around AI agents running on its Ontology, in both of its markets.

In companies, AIP turns Foundry from a data and apps platform into the place where agents do operational work: rescheduling production, updating suppliers, drafting claims decisions, with people approving the steps that need them. Bootcamps and the Developer Tier, which gives developers limited free access to Foundry and AIP, are meant to bring in customers far faster than the old pilot-led sales did. Palantir is also building industry-wide versions, shared platforms for airlines, shipbuilding, insurance, healthcare and telecoms that many companies in one industry run on.

In government, it is moving from programs to enterprise agreements that standardize a whole military service on its software, and from headquarters to the edge: Gotham and AIP on vehicles, aircraft and disconnected hardware, kept current by Apollo. Allied governments are the next market, where US software faces questions of sovereignty.

Apollo, through FedStart, makes Palantir a route into government for other software companies as well as a delivery system for its own. The direction across all of it is the same: from software people use to look at data, to software that runs operations, with AI doing more of the work.

• Proving AI agents at scale. Bootcamps show a working workflow in days; running thousands of agents reliably inside a manufacturer's or an army's daily operations is a harder engineering problem, and failures there are costly and public.
• Competing with the lakehouses and the clouds. Databricks, Snowflake, Microsoft, Google and AWS are all building agent platforms on customers' data, priced by use and open to many tools; Palantir's proprietary stack has to stay far enough ahead to justify its price.
• Depending on governments. Much of the business rests on defense and intelligence budgets and procurement decisions, in the US and abroad, that can change with politics.
• Controversy. Work for immigration enforcement, police and intelligence agencies draws criticism that can cost it customers, especially in Europe, and staff.
• Scaling forward deployment. On-site engineers win accounts but limit how many customers Palantir can serve; the Developer Tier, bootcamps and partners have to carry more of the load.
• Equity pay, $684M or 15.3% of revenue in 2025, remains a large cost of keeping its engineers.