What Palantir actually sells

Palantir sells software that turns a large institution into something closer to a live strategy game. Every plane, patient, part, shipment, transformer, case file, and target becomes an object on one shared map of the organization. People can search it, filter it, graph it, alert on it, and increasingly ask an AI to work on top of it. That is useful when you are running an airline or mine. It is also useful when you are running a military targeting program or a deportation system. The same basic product pattern shows up in both worlds.

That is the cleanest answer to the question people have been asking about Palantir for twenty years.

The company now says it has four main platforms: Gotham, Foundry, Apollo, and AIP. By the end of 2025 it had 954 customers, up 34% year over year, and its top 20 customers averaged $93.9 million in annual revenue. So this is no longer a niche spy-tech company with a few unusual contracts. It is a big software vendor. But the interesting part is still what the software actually does.

The stack

Foundry is the main commercial product. Palantir describes it as a data operations platform for data management, logic, analytics, and workflows.

Gotham is the government and defense side. Palantir is less explicit in public docs, but in practice it is the part most associated with intelligence, military, and police analysis. Palantir’s own filing says Gotham powers defense and intelligence operations and integrates with the rest of the stack.

Apollo is the least glamorous but maybe the most important reason Palantir can win high-stakes deployments. It is a deployment and operations system that pushes software across public cloud, on-prem, and disconnected or air-gapped environments. That matters if your customer is a hospital, a military network, or a police force that cannot just use normal SaaS.

AIP is the newer AI layer. Palantir’s own docs describe it as LLM connectivity, agents, automations, evals, and governed AI workflows. The key point is that AIP is not supposed to be a raw chatbot glued onto company data. It is meant to sit on top of Palantir’s existing model of the organization, with permissions, audit trails, and human review checkpoints already in place.

If you want the short version, Palantir is selling an operating system for institutions.

The core trick is the Ontology

The heart of Palantir is what it calls the Ontology. In plain English, that is a shared model of the organization.

Palantir takes raw inputs from existing systems, then maps them into objects, properties, links, and actions.

An object might be:

• a patient
• a shipping container
• a locomotive part
• a transformer
• a mine sensor
• a deportation target
• a military point of interest

A property is something about that object, like location, health, owner, urgency, warranty status, or risk score.

A link connects objects to each other: this patient is in this ward, this order depends on this supplier, this part belongs to this engine, this person is associated with this address, this target came from this sensor report.

An action is what a user can do from inside the system: assign a task, reschedule a shipment, approve a maintenance plan, mark an alert as resolved, write back to another system, or trigger a workflow.

That sounds abstract until you see the examples. Palantir’s own docs say the Ontology can represent plants, equipment, products, customer orders, and financial transactions. The backend docs show it is built to index, query, search, aggregate, and write back across those objects.

This is the main reason Palantir is not just “a big database.” A database stores records. Palantir tries to build a live model of the organization, then lets people operate through that model.

What data goes in

This is where Palantir gets concrete.

In a consumer goods supply chain case study, Palantir says it integrated seven ERP systems and modeled plants, SKUs, customers, bills of material, inventory, raw materials, and forecast demand. The goal was not a prettier dashboard. It was to let managers ask questions like: where can we swap materials, reduce waste, and raise margins on specific products?

In a shipping case study, the inputs included booking systems, container tracking systems, CRM data, and task queues for thousands of agents in more than 100 countries.

In a rail warranty case study, the data covered each locomotive part’s lifecycle from purchasing to storage to installation to failure to removal, plus warranty terms and claims workflows.

At Ferrari, Foundry ingests Grand Prix data, test bench results, and part information for power unit engineers. Ferrari says an F1 season can generate as much as 1.5 trillion data points.

At Airbus Skywise, the data includes work orders, spares consumption, component data, fleet configuration, onboard sensor data, flight schedules, operational interruption history, pilot reports, technical requests, and service bulletins. Airbus now says more than 12,300 aircraft and 48,000 users are connected to the platform.

At PG&E, the claimed input scale is 8 to 10 billion data points per day. The company has described using Foundry with smart meter data, geospatial data, grid topology, and wildfire risk information across 25,000 miles of wire.

At Rio Tinto, the data ranges from hundreds of equipment units and rail systems in Pilbara to thousands of underground mine sensors at Oyu Tolgoi in Mongolia.

In healthcare, NHS England says its Federated Data Platform is meant to work with bed counts, waiting lists, staff rosters, supplies, discharge data, and social care capacity. Tampa General Hospital has described integrating nursing schedules, patient information, and acuity levels.

In semiconductors, Athinia uses Palantir Foundry to let fabs and materials suppliers share codified or anonymized process and materials data. The pitch is better quality control, supply chain transparency, and faster root-cause analysis.

In government and enforcement, the data gets more sensitive.

Reuters says the Pentagon’s Maven Smart System takes in data from satellites, drones, radars, sensors, and intelligence reports to identify military points of interest and speed analyst work.

The ACLU’s roundup and 404 Media’s reporting say Palantir systems used by ICE can pull together immigration history, family relationships, employment information, phone records, biometrics, criminal records, home and work addresses, and data from sources like HHS, USCIS, and commercial data products.

That range is the story. Palantir can work with almost any data, but it is strongest when the data describes an operation with lots of moving parts and a high cost of being wrong.

What users actually see

A lot of confusion about Palantir comes from people imagining some giant black box. The user-facing layer is actually more mundane.

Palantir’s Object Explorer is basically search for the Ontology. Users can run keyword queries or property filters, see tables, maps, charts, compare object sets, export results, and take bulk actions.

Quiver is the analysis surface. It uses cards on a canvas. Some cards filter or join data. Others make charts, time-series views, anomaly detection, or dashboards.

Workshop is the app builder. Palantir’s own getting-started example is a “Flight Alert Inbox” with an object table, filter list, object view, and a resolve button that writes back through an action. That sounds trivial, but it is a good picture of what many Palantir apps really are: operator screens for people who need to see the current state of the world and do something about it.

This matters because Palantir is not mostly selling executive dashboards. It is trying to build software that working teams sit inside all day.

Why companies buy it

The value proposition is simple: most big organizations already have the data they need, but it is trapped in systems that do not line up with the actual work.

A manufacturer has ERP, MES, procurement systems, quality systems, spreadsheets, and sensor feeds. None of them agree cleanly with each other. A hospital has EHRs, scheduling systems, staffing systems, discharge data, and supply tools. An airline has maintenance systems, operations data, flight telemetry, engineering logs, and vendor data. A police or intelligence agency has case systems, registries, tip lines, and surveillance feeds.

Palantir’s claim is that it can make those systems legible as one operational picture, then let people actually run the business on top of that picture.

The case studies are unusually specific:

• A Fortune 100 consumer goods company integrated seven ERP sources in five days and Palantir claims up to $100 million in annual savings.
• A major shipping company used Foundry to prioritize tasks across 100+ countries, with Palantir claiming tens of millions in savings.
• A railroad used Foundry to recover $20 million+ per year in warranty claims.
• Airbus says Palantir helped accelerate A350 delivery by 33% and later expanded into the broader Skywise ecosystem.
• United said the combination of Palantir Foundry and Skywise would improve maintenance and operational reliability.
• Ferrari says Palantir lets race engineers turn analyses that once took minutes into seconds.
• Rio Tinto says Foundry helps coordinate 53 driverless trains and monitor mine risk from thousands of sensors.
• Tampa General said Foundry helped drive a 30% improvement in nurse staffing ratio, a 28% reduction in PACU hold time, and a disaster-response app in under 24 hours.

The through-line is operational compression. Fewer handoffs. Fewer spreadsheet rituals. Faster decisions with more context.

Why governments buy it

Governments have the same mess as big companies, but with worse procurement, worse data sprawl, stricter security, and higher stakes.

The British government’s biggest public Palantir deployment is the NHS Federated Data Platform. It is built on Foundry, under a £330 million contract over seven years. NHS England says the system is meant to support elective recovery, care coordination, vaccination, population health planning, and supply chain management. Its FAQ says local trust instances can hold identifiable care data for direct care, while national planning views use de-identified or aggregate data. By late 2024, NHS England said 91 trusts and 31 systems had signed up.

In the U.S. military, Palantir is now deeply embedded. Reuters reports the Pentagon awarded a $480 million Maven contract in 2024, then the Army expanded it with more licenses, and later consolidated contracts into an agreement worth up to $10 billion over ten years. Maven is not a vague “AI for defense” story. Reuters describes it as battlefield software that helps analyze sensor and intelligence data and identify targets or military points of interest.

Palantir’s work with immigration enforcement is even more concrete, and more politically explosive. The ACLU describes four main systems or programs: ELITE for target identification, ICM for case management, ImmigrationOS for streamlining deportation operations, and AI-enhanced tip processing. 404 Media says ELITE can display targets on a map, show dossiers on individuals, and assign an address confidence score. Reuters says ICE awarded Palantir a $30 million contract related to systems that identify undocumented immigrants and track self-deportations.

Germany gives a different picture of the same core technology. Deutsche Welle reports that Palantir’s Gotham software is already used by police in Hesse, Bavaria, and North Rhine-Westphalia, with Baden-Württemberg planning to implement it. In Hesse it is branded HessenData. The Hesse interior ministry says it has been used since 2017 against organized crime and terrorism and credits it with helping prevent a 2018 Islamist attack.

So when people ask which governments use Palantir, the answer is not vague. Public reporting and official materials show real deployments in the U.S. military, U.S. immigration enforcement, NHS England, and multiple German state police systems, among others.

The AI layer changes the pitch, not the core product

A lot of recent coverage makes Palantir sound like an AI company that happened to have government contracts first. That gets the order backwards.

The AI story works because Palantir already spent years building the data model, permissions, and workflow layer. Its own docs say AIP agents are grounded in the Ontology, documents, and custom tools, with LLM access scoped by platform security. AIP Logic is for building LLM-powered functions that can read ontology objects, return outputs, and even stage edits for review. Palantir’s filings emphasize human review checkpoints and audit controls.

That is why Palantir can sell AI to very conservative organizations. It is not saying, “let the chatbot loose on your data.” It is saying, “we already know what your data means, who can touch it, and what actions are allowed. Now we can put an LLM inside that box.”

The uncomfortable truth

Palantir’s commercial and government businesses look very different morally, but they are technically closer than many people want to admit.

The same features that help a hospital coordinate beds and discharges can help an immigration agency coordinate detention and removal.

The same features that help a utility map wildfire risk can help a military map targets.

The same “single source of truth” story that sounds good in manufacturing also means centralizing very sensitive information about people.

That is why critics focus less on the software buzzwords and more on the consequences of putting this kind of system in the hands of states. The ACLU argues that Palantir systems help create centralized dossiers, expand the use of brokered personal data, and enable opaque targeting decisions. In Germany, civil liberties groups have challenged Palantir-style police data mining in court. In Britain, critics of the NHS rollout worry less about today’s official contract language than about what a centralized health data platform could enable later if laws or politics change.

Palantir’s own answer is that its permissioning, logging, and governance controls constrain misuse. Sometimes that is probably true. But the controls do not change the fact that the software is designed to make large institutions more legible to themselves and more capable of coordinated action. Whether that feels admirable or chilling depends a lot on who the customer is and what they are trying to do.

Final view

Palantir is not mainly selling dashboards. It is not mainly selling a database. It is not even mainly selling AI.

It sells software for taking a sprawling institution, modeling the important parts of it as a live system, and letting people operate through that model.

For companies, that means better maintenance, fewer supply chain blind spots, faster planning, and more disciplined operations.

For governments, it can mean better hospital coordination or military logistics. It can also mean target selection, mass surveillance, and deportation infrastructure.

That is what Palantir actually does. It makes complicated institutions easier to see, easier to query, and easier to steer.