What Is Palantir? Gotham, Foundry, AIP, Defense AI and the Future of the Company

 Palantir Is Building the Operating System for the Real World

Inside Gotham, Foundry and AIP — and the company betting that AI becomes most valuable when it can change what happens next.

Palantir’s ambition goes far beyond data analytics: its software is designed to connect information, AI models and human decisions across governments, industry, healthcare and defense.

Palantir is an unusual technology company because most people can go their entire lives without knowingly touching its software, while institutions around them may depend on it. It has no mass-market app, and it does not train the frontier language models that dominate AI headlines. Instead, its software turns up in places where a bad decision can be expensive, politically sensitive or lethal: military headquarters, intelligence agencies, factories, banks, hospitals, supply chains and government departments.

That makes Palantir hard to categorize. “Big data” describes where it started; “AI company” describes only part of what it has become. A better way to think about Palantir is as software that sits between an institution’s information and the decisions made from it. It pulls together fragmented data, represents the organization as a working model, connects that model to algorithms and AI, and gives people a way to act on the result.

On a factory floor, that may mean spotting that a missing component will halt a production line tomorrow. In a hospital, it may mean seeing which beds, staff and operating rooms can be rearranged to shorten waiting times. At a military headquarters, it may mean deciding which satellite image, drone feed or sensor report deserves attention now. The settings are radically different. Palantir’s pitch is that the underlying problem is the same: organizations know a great deal, but struggle to turn that knowledge into a shared, current picture.

This is both the attraction and the source of the controversy. Software that makes an institution faster and more capable does not decide whether the institution’s goal is wise, humane or legitimate. Palantir argues that granular permissions, audit trails and human approval can keep powerful systems under control. Critics answer that safeguards do not settle the more political question: which institutions are being made more powerful, and how easy is it to challenge their decisions once the software becomes part of the machinery of government or war?

Palantir in 60 seconds

Founded	2003; built first for the U.S. intelligence community after the post-9/11 counterterrorism shift.
Co-founders	Alex Karp, Peter Thiel, Stephen Cohen, Joe Lonsdale and Nathan Gettings.
Core platforms	Gotham, Foundry, Apollo and AIP (Artificial Intelligence Platform).
Central idea	Turn fragmented data into an operational model of the real world, then connect decisions and actions to that model.
Customers	1,049 as of June 30, 2026, using Palantir’s SEC definition, which can count separate government units as separate customers.
2025 revenue	$4.48 billion; 54% government and 46% commercial.
Q2 2026 revenue	$1.94 billion, up 93% year over year. Full-year 2026 guidance was raised in August to about $8.15 billion.
What it is not	A consumer chatbot company, a conventional data broker, or primarily a frontier-model laboratory.

1. The company that began with a “seeing stone”

Palantir was founded in 2003, in the aftermath of the September 11 attacks, when U.S. intelligence agencies were wrestling with a paradox: they possessed enormous amounts of information, but important relationships could still disappear between databases, jurisdictions and analytical teams. Peter Thiel and other early Palantir founders brought with them an idea familiar from PayPal, where software had been used to detect fraud patterns that automated rules alone could miss. The goal was not to replace analysts, but to give them a better way to connect clues.

The company’s name comes from J.R.R. Tolkien’s palantíri, the “seeing stones” that allowed distant events to be observed. It is an unusually apt reference. In Tolkien’s world, a palantír is useful precisely because it can reveal what is far away — but what it shows, who controls it and how the image is interpreted can change the outcome. That tension has followed Palantir from the beginning: build a real-world seeing stone, then try to keep its power bounded by permissions, provenance and auditability.

One of Palantir’s earliest backers was In-Q-Tel, the venture-capital organization created to support technologies useful to the U.S. intelligence community. For years the company operated largely outside the consumer-tech spotlight, working with government agencies on terrorism, intelligence analysis and other sensitive data problems. That origin shaped both its engineering culture and its reputation: highly technical, unusually secretive, deeply embedded in state institutions and willing to take on projects that many Silicon Valley companies preferred to avoid.

Palantir went public through a direct listing in 2020. By then, it had already spent years moving beyond intelligence and defense into commercial industries. The same problem kept appearing there in a less dramatic form: companies had accumulated databases, dashboards and software systems, yet still lacked a coherent picture of how the business actually worked from end to end.

2. What Palantir actually does

Imagine a manufacturer that already owns all the software it is supposed to need. Orders live in an ERP system. Machine sensors stream into another database. Maintenance histories sit in service tickets. Supplier delays arrive by email or spreadsheet. Engineers have their own tools, finance has its forecasts, and the warehouse has inventory software. Every system may work on its own while the company as a whole remains surprisingly hard to see.

A traditional analytics project might copy those datasets into a warehouse and build dashboards. Palantir tries to go further. It connects the sources, tracks lineage and permissions, then maps the data into something closer to the language of the business: machines, parts, suppliers, orders, technicians, plants, contracts and the relationships between them. Palantir calls this semantic layer the Ontology.

A database row is just a record. In Palantir’s Ontology, an object can carry meaning, permissions, history, relationships and available actions. A “machine,” for example, can be linked to its maintenance record, the product it is making, the parts it consumes, the technicians certified to repair it, the orders that depend on it and the cost of downtime. With that context in place, software can do more than display information: it can simulate scenarios, trigger workflows, run optimizers or let an AI agent propose an action using the same rules the organization already follows.

This is why Palantir increasingly describes its products as operating systems rather than analytics tools. The aim is not merely to answer “what happened?” but to connect “what is happening?” with “what should we do next?” — and, where rules allow it, turn the approved decision into action.

A conceptual view of Palantir’s Ontology approach: instead of treating information as isolated databases, the platform connects real-world objects — suppliers, machines, orders, people and locations — into a shared operational model.

3. The four-platform stack: Gotham, Foundry, Apollo and AIP

The product names can make Palantir sound more arcane than it is. In practice, its current stack is easiest to understand as four layers that increasingly overlap.

Gotham: the government and defense operating system

Gotham grew out of Palantir’s intelligence work. It integrates data from different sources, helps analysts discover relationships and supports operational planning. In modern military deployments, that can mean combining intelligence reports, sensor data, video, geospatial information and unit positions into a common picture that commanders and operators can use. The 2025 annual report describes Gotham as a system designed to help users “see, understand, and act” in the modern battlespace, including near-real-time data from multiple domains and sensors.

In practice, Gotham is much more than a map covered in icons. Its value lies in preserving the chain behind a decision: where a piece of information came from, who is allowed to see it, how confident analysts are in it, what assets are available and what changes when the underlying situation changes.

Foundry: the commercial operating layer

Foundry takes a related architecture into companies and civilian institutions. It can connect operational databases, spreadsheets, cloud data warehouses and software applications, then build shared data products and workflows on top of them. Palantir says all of its commercial customers now use Foundry, while many government customers use it as well.

The usual pitch is not “rip everything out and start again.” Palantir is often most useful when a customer has years of accumulated systems that cannot realistically be replaced at once. Foundry sits across them, connects what was previously separate and gradually moves important decisions into a common operational environment.

Apollo: the software deployment layer

Apollo is less visible but strategically important. Palantir built it to deploy and update software across very different environments: public clouds, private infrastructure, classified networks, ships, aircraft, disconnected edge devices and other places where ordinary cloud software assumptions break down. In a world where AI systems increasingly have to run near factories, military units or sovereign data centers, that deployment capability becomes part of the product rather than back-office plumbing.

AIP: the layer that connects AI to real operations

AIP, launched in 2023, changed how many customers and investors understood Palantir. The company did not need to build the world’s best general-purpose language model. Instead, AIP could connect third-party models from OpenAI, Anthropic, Google and NVIDIA — along with open-weight or customer-hosted models — to an organization’s own data, permissions and workflows.

That puts Palantir in a different business from the frontier-model laboratories. A model company is trying to build more capable intelligence. Palantir is trying to govern where that intelligence may look, which tools it may use, what actions it may take and how its output becomes part of a real workflow. In 2026, the platform expanded support for pro-code agents that can read and write Ontology data, call tools and run inside business applications with scoped permissions.

For the broader shift from models that answer questions to systems that can take action, see Next Horizon’s Artificial Intelligence Explained: From Neural Networks to AI Agents. For Palantir, the important point is simple: an agent becomes far more useful — and far more consequential — once it can interact with real organizational systems.

4. Palantir’s biggest bet: the model is not the operating system

The AI boom encouraged a simple story: whoever builds the smartest model will own the future. Palantir is betting that the model is only one part of the stack. One laboratory may lead on reasoning for a few months while another is cheaper, faster, more private or better suited to a particular language, jurisdiction or security classification. Large organizations do not want to rebuild their operations every time that ranking changes.

Palantir therefore wants the surrounding system to outlive any one model. In October 2026, for example, AIP was adding newly available Anthropic models to UK-georestricted environments; earlier in the year it added new NVIDIA and OpenAI models. The point is not loyalty to a particular AI provider. It is the ability to replace the intelligence engine without rebuilding the data model, permissions, audit system and workflows around it.

Karp increasingly describes this as “AI sovereignty” — the idea that organizations and states should retain control over their data, prompts, models, business logic and operational knowledge rather than hand the entire stack to a single AI laboratory. In March 2026, Palantir and NVIDIA announced a reference architecture for a sovereign AI operating system that can run a full Palantir stack on dedicated NVIDIA infrastructure.

The appeal is easy to see. A ministry may not want classified workflows sent to a public model endpoint. A pharmaceutical company may use one model for research and another for regulated documentation. A manufacturer may want an agent to act on plant data without exposing proprietary processes to an outside provider. AIP is designed to make those choices part of normal operations rather than special engineering projects.

But the sovereignty argument contains a tension that will follow Palantir as it grows: if the software that protects an institution from dependence on model providers becomes itself extremely difficult to replace, has sovereignty actually increased — or has dependency merely moved to a different layer?

5. From dashboards to decisions: what commercial customers are using it for

Palantir’s commercial growth shows that this “operating system” idea is no longer confined to bespoke government projects. In the first half of 2026, commercial customers generated 48% of company revenue, and U.S. commercial revenue grew 142% year over year. Palantir reported 1,049 customers at the end of June, up from 849 a year earlier, although its SEC definition can count separate units of large government institutions as separate customers.

The product is easiest to understand through mundane, specific jobs. Palantir says Wendy’s supply-chain cooperative uses its systems for real-time inventory and order tracking across thousands of restaurants and for reallocating resources when shortages appear. Panasonic Energy has used an AIP-powered assistant to search large volumes of machine-maintenance history and help technicians diagnose equipment problems. Heineken USA has used the platform to combine brewery, warehouse, customer and vessel data so supply disruptions can be detected earlier. Citi Wealth has described compressing parts of account-opening and data-integration workflows that previously took days.

None of these is a glamorous AI demonstration, and that is partly why the examples matter. The economic value of generative AI may lie less in producing more text than in reducing the gap between what an organization knows and what it can actually do. Palantir wants to sit inside that gap.

This is also a different kind of automation. Rather than writing one brittle script for one task, a company can give an agent access to a governed model of the business, a limited set of tools and explicit permissions. That makes the surrounding controls as important as the model itself. Next Horizon’s investigation of AI agents and control failures looks at why permissions, monitoring and environment become critical once software can take actions instead of merely producing answers.

Palantir’s commercial strategy increasingly brings AI into physical operations, where technicians can combine equipment data, maintenance records and supply-chain information without removing humans from the decision process.

6. The battlefield is where Palantir’s thesis becomes most visible

Defense is not an accidental side business for Palantir. It is part of the company’s identity and of Alex Karp’s argument about what technology companies should be for. In his 2025 book The Technological Republic, written with Palantir executive Nicholas Zamiska, Karp argues that Silicon Valley spent too much of the previous era optimizing consumer engagement while backing away from difficult public missions. His alternative is explicit: advanced technology should strengthen the institutions and industrial capacity of Western democracies, including their militaries.

That worldview is now visible in a series of large programs. The U.S. Army’s Maven Smart System uses Palantir software for AI-enabled intelligence fusion, targeting, situational awareness and planning. In May 2025 the U.S. Department of Defense announced a $795 million contract modification for Maven software licenses extending through 2029. In July 2025, the Army consolidated multiple arrangements into an enterprise agreement that allows up to $10 billion of Palantir-related purchases over ten years. The ceiling is not guaranteed spending, but it shows how central the software has become to Army procurement.

NATO followed the same direction. The Alliance acquired a NATO version of the Maven Smart System in March 2025. By June 2026, NATO said it had reached full technical operational capability across Allied Command Operations. NATO describes the system as a common data-enabled warfighting platform for intelligence fusion, targeting, battlespace awareness, planning and faster decision-making.

Palantir is also moving into programs where software is no longer merely attached to hardware but helps define the system itself. TITAN, the U.S. Army’s next-generation intelligence ground station, is described by Palantir as the Army’s first “AI-defined vehicle.” Palantir is the prime contractor, coordinating partners including Northrop Grumman, Anduril and L3Harris. The vehicle is meant to take data from deep-sensing systems and turn it into information useful for long-range precision fires.

Then there is ShipOS. Announced in December 2025, the U.S. Navy program authorizes up to $448 million to deploy Foundry and AIP across the maritime industrial base. The target is not a weapon in the usual sense. It is the production system behind ships and submarines: schedules, suppliers, bottlenecks, materials and industrial capacity. Here Palantir’s defense and manufacturing strategies become the same story. The company is betting that part of the West’s military problem is really a coordination problem — and therefore, at least partly, a software problem.

A conceptual illustration of software-defined warfare: satellite imagery, drones, sensors and intelligence are fused into a common operational picture, while human operators remain responsible for reviewing and authorizing decisions.

7. Ukraine: a wartime test of Palantir’s model

Ukraine became one of the clearest demonstrations of modern Palantir because it put the company’s central claim under wartime pressure: better software can change the effectiveness of existing physical systems without changing the hardware itself.

On June 2, 2022, Alex Karp met President Volodymyr Zelenskyy in Kyiv. The Ukrainian presidency said Karp was the first chief executive of a major Western corporation to visit the capital after the full-scale invasion began. The meeting covered defense, security and digitization, and Palantir said it was prepared to open an office in Ukraine and develop technology with Ukrainian specialists.

By early 2023, Karp was publicly describing Palantir’s software as deeply involved in Ukrainian targeting. Reuters reported his claim that the company was “responsible for most of the targeting in Ukraine,” including support for identifying tanks and artillery. The figure is Karp’s own description, not an independently audited measure, but it gives a sense of the role Palantir says its software played.

The work was not limited to battlefield targeting. Ukraine also adopted Palantir technology to help investigators combine evidence relevant to alleged Russian war crimes, including intelligence, satellite imagery and photographs. The broader lesson was not that one algorithm could decide a war. It was that commercial satellites, drones, intelligence, software and human judgment could be connected into a much faster operational loop.

That experience now echoes through NATO procurement and the UK’s defense strategy. In 2025, the British government announced a strategic partnership with Palantir involving up to £1.5 billion of planned company investment in the UK, with London becoming Palantir’s European defense headquarters. The government explicitly said the partnership would develop AI-powered capabilities that had already been tested in Ukraine.

8. Does Palantir “own everyone’s data”? Not exactly — but that is not the whole privacy question

A common caricature of Palantir is that it operates a giant private database quietly accumulating personal information from around the world. That is usually the wrong mental model. Palantir says it is not a data broker and that customers remain the legal owners or controllers of the data placed in its systems. In many deployments, Palantir acts as a processor rather than deciding why the data is being used.

Its products include granular access controls, data lineage, audit logs and rules that can limit who sees a piece of information and for what purpose. These are substantive parts of the technology, not just policy language. Palantir also maintains a Privacy and Civil Liberties engineering function and publishes principles for responsible AI deployment.

But privacy is not only a question of whether Palantir itself owns the data. A government agency may already have lawful access to many databases that are individually difficult to combine. Software that links them, resolves identities, surfaces relationships and makes the results operational can dramatically increase institutional power without creating a new dataset from scratch. That is why the ethical debate follows Palantir even when the customer — not Palantir — legally controls the information.

The same architecture can therefore look very different depending on where you stand. To an engineer, it can replace uncontrolled spreadsheets with permissions and auditable access. To a civil-liberties advocate, it can make surveillance or enforcement far more efficient. Those descriptions are not mutually exclusive.

9. Why Palantir remains so controversial

The controversies around Palantir are not one controversy. Reducing them all to “privacy” obscures the harder questions: what the software makes possible, which customers the company is willing to serve, and what happens when a public institution becomes dependent on one vendor for a critical operational layer.

ICE and immigration enforcement

Palantir has worked with U.S. immigration authorities for years. In 2025 it received a roughly $30 million contract to build ImmigrationOS, a system intended to support immigration-enforcement workflows. Civil-liberties groups such as the ACLU argue that Palantir’s technology makes aggressive immigration enforcement more scalable and raises serious constitutional and human-rights concerns. In February 2026, amid renewed scrutiny, Karp defended Palantir’s surveillance technology by emphasizing technical safeguards designed to limit unauthorized access and create oversight. The disagreement is therefore not simply about whether safeguards exist. It is about whether a company should provide the capability for the mission at all.

The NHS and the problem of public trust

In England, a Palantir-led consortium won the NHS Federated Data Platform contract in 2023, worth up to £330 million over seven years. NHS England says the platform is NHS-controlled and that Palantir operates as a processor, with the goal of connecting operational information so hospitals can coordinate care, capacity and resources more effectively.

By late 2026, however, the contract had become one of Palantir’s most visible political battles. More than 44,000 people filed legal objections to the platform’s handling of their data. A Financial Times analysis reported that more than 20 NHS trusts had stopped using specific Palantir waiting-list tools, complicating public claims about their impact. The UK government was also considering domestic alternatives as part of a wider debate over technological sovereignty and dependence on U.S. suppliers. None of this, by itself, proves that the platform is insecure. It does show how difficult it is to separate technical capability from public legitimacy when health data is involved.

Israel and military use

In January 2024, Palantir announced a strategic partnership with Israel’s Ministry of Defense to provide technology for “war-related missions.” Human-rights organizations, including Amnesty International, have since criticized the company over the relationship and argued that technology suppliers should be accountable for risks connected to military operations in Gaza. Palantir has publicly defended its work with U.S. and allied militaries as consistent with its mission of supporting Western security. For readers, the important distinction is between verified facts about the partnership and broader claims about exactly how particular military decisions were made, many of which are difficult to independently assess.

Vendor lock-in: sovereignty versus dependence

One of the most important criticisms is also less dramatic than the debates over surveillance or war. Palantir sells organizations a way to escape fragmented systems and regain control over their data and AI stack. But a successful deployment can make the Palantir layer extraordinarily central. In February 2026, the UK Ministry of Defence said a direct award to Palantir was justified partly because only that supplier could meet the technical need and because changing suppliers would create disproportionate operational and maintenance difficulties. Critics read that as evidence of lock-in. Supporters read it as evidence that the system had become mission-critical.

That tension may become one of the defining problems of enterprise AI. The more useful an “operating system for the organization” becomes, the harder it is to replace. Procurement, interoperability and a credible exit plan may matter almost as much as model accuracy.

10. Why Palantir’s growth accelerated so sharply

For years, Palantir was known for long, expensive deployments that depended on teams of engineers working closely with customers. The company still relies on “forward-deployed engineers,” but its commercial playbook has changed. AIP bootcamps try to compress the first stage: customers bring real data and a real operational problem, and Palantir aims to build something useful in days instead of beginning with months of consulting and integration work.

The timing was fortunate. Companies rushed to experiment with generative AI in 2023 and 2024, then discovered that a convincing chatbot was far easier to build than a production system that could safely use proprietary data, respect permissions, interact with existing applications and show measurable economic value. Palantir had spent two decades working on precisely those less glamorous problems.

The financial shift has been dramatic. Revenue rose from $2.87 billion in 2024 to $4.48 billion in 2025. In the second quarter of 2026, revenue reached $1.94 billion, up 93% from a year earlier. U.S. commercial revenue rose 149% year over year in that quarter, while U.S. government revenue rose 90%. In August, the company lifted its full-year 2026 revenue forecast to roughly $8.15 billion.

Those numbers matter here less as an investment story than as evidence that operational AI is moving from pilots into budgets. Palantir’s next test is different: whether growth at this speed can coexist with interoperability, public trust and an architecture that does not become a closed world of its own.

11. What Palantir thinks comes next

Palantir does not publish a tidy ten-year roadmap. But its product launches, contracts and Karp’s writing point in the same direction: the company expects institutions themselves to become increasingly software-defined.

AI agents that can actually operate

The most immediate shift is from assistants to agents. In 2026, Palantir expanded tools for building agents that can query data, call software tools, write information back into the Ontology, repair workflows and run inside operational applications. The long-term goal is straightforward to describe and difficult to implement safely: a person gives the system a task — investigate a supply disruption, reconcile a production plan, prepare a maintenance response — and the agent works across governed systems instead of merely returning prose.

Sovereign AI stacks

Another theme is infrastructure independence. Palantir is positioning itself as a layer that can run across public clouds, private clouds, classified networks and dedicated sovereign data centers. Its NVIDIA partnership makes the ambition explicit: an organization should be able to own or tightly control the compute, models, data and operational logic while still using frontier AI capabilities.

Reindustrialization as a software problem

Manufacturing is another part of the same thesis. Warp Speed, Palantir’s manufacturing operating system, and ShipOS both assume that industrial capacity can be improved by giving factories and supply chains a live, shared model of themselves. Palantir increasingly talks about American reindustrialization in language that resembles its older intelligence work: fragmented information, slow decisions, poor coordination and systems that adapt too slowly when conditions change.

Defense networks where software connects sensors, decisions and production

The same logic is visible in defense. Maven links information and command workflows. TITAN brings AI into a physical ground station. ShipOS targets the industrial base. NATO is standardizing on a common AI-enabled warfighting layer. Put together, these projects point toward a military in which advantage comes not only from the quality of an individual weapon, but from how quickly sensors, intelligence, logistics, commanders, factories and weapons can be connected into one decision network.

A model-agnostic layer for the AI economy

Underneath all of this is another bet: Palantir expects the model market to remain plural. If OpenAI, Anthropic, Google, NVIDIA and open-weight ecosystems keep leapfrogging one another, customers will need a durable layer above them. Next Horizon’s profile of Anthropic and Claude examines a company whose core asset is the model itself. Palantir is making the opposite wager: the more intelligence becomes available as a commodity, the more valuable the machinery that governs and applies it may become.

Palantir’s long-term vision is an interconnected operational layer spanning critical infrastructure, industry, healthcare and defense. The promise is faster, better-informed decisions — but the same level of integration also raises questions about control, dependence and oversight.

12. The harder question: what happens if Palantir’s vision works?

For years, Palantir’s problem was explaining what it did. In 2026, the more interesting question is what happens if its idea of the organization works at scale.

The end state is an institution with a living digital representation of itself: assets, people, rules, supply chains, missions and constraints. AI models can reason over that representation; agents can take bounded actions inside it; people approve, interrupt or redirect the process. The promise is faster coordination. The risk is that more and more consequential decisions become dependent on the quality, rules and ownership of that digital layer.

That is what makes Palantir more consequential than an ordinary enterprise-software vendor. It is not mainly trying to digitize documents or bolt a chatbot onto existing work. It is trying to digitize the layer where institutions decide what to do.

Palantir says its goal is to augment human intelligence rather than replace it. That distinction matters, but it does not resolve the ethical problem. Better decision infrastructure amplifies the institution using it. A hospital network, a democratic government, a manufacturer and an immigration agency can all use the same technical ideas toward very different ends.

The metaphor in the company’s name therefore remains useful. A seeing stone is valuable because it makes distant, fragmented reality visible; it is dangerous for the same reason. If Palantir’s vision spreads, the important questions will not be only how clearly institutions can see, but who is allowed to look, what they may do with what they see, and how the people affected by those decisions can challenge them.

That is why Palantir deserves attention beyond the usual AI hype cycle. It is one of the most ambitious efforts to turn artificial intelligence from software that talks about the world into infrastructure that can change what happens inside it.

FAQ

Is Palantir an AI company?

Yes, but it is better understood as an operational software company that integrates data, analytics, AI models and workflows. It does not primarily compete by training the largest general-purpose language model.

What is the difference between Gotham and Foundry?

Gotham grew out of government, intelligence and defense use cases; Foundry is the broader data and operations platform used heavily by commercial and civilian organizations. Their architectures increasingly overlap, especially through the Ontology and AIP.

Does Palantir own the data in its customers’ systems?

Generally, no. Palantir says customers control their data and that the company often acts as a processor. The privacy debate focuses less on ownership than on how effectively the software can combine and operationalize information that customers already possess.

Does Palantir build its own ChatGPT competitor?

Not in the conventional sense. AIP is designed to connect different third-party and customer-hosted models to organizational data, tools and permissions. Palantir’s strategic asset is the operational layer around the models.

Why is Palantir controversial?

Because its software is used in sensitive domains including military operations, intelligence, immigration enforcement and healthcare; because critics question some customer relationships; and because deep deployments can raise concerns about surveillance, public accountability and vendor lock-in.

What does Palantir mean by “AI sovereignty”?

The company uses the phrase for the ability of an organization or state to control its data, models, prompts, business logic, infrastructure and permissions instead of surrendering the entire AI stack to a single external model provider.