PALANTIR: War, Surveillance, Artificial Intelligence

Part 2: Core Technology, Ontology and the Decision-Making Revolution

Chapter 3. Ontology: The Magic of Turning Data into Knowledge

Attorney Kyungjin Kim

A. The Heart of Palantir's Technology: Defining Ontology

(1) Implementing a "Digital Twin" Beyond Simple Data Integration

Late one night in 2008, an emergency meeting was held at the headquarters of a major bank in London's financial district. Thick reports were piled on the table. The CEO asked: how much is our bank exposed to Greek sovereign debt? It was a straightforward question. The chief risk officer opened his laptop and called another department. An hour passed. Wiping the sweat from his brow, he answered: I'm not sure exactly, but it seems to be around $1 billion. The CEO was despairing. One of the world's premier financial institutions didn't even know where its own money was.

The problem was not a lack of data. The bank had complete data. It was just written in different languages. On the lending department's books, Greece was labeled "GR." On the bond trading desk, it was recorded as "Hellas." The derivatives department used a numerical country code. The computers did not know that these three things pointed to the same country. This was the fundamental problem Palantir sought to solve.

Palantir's engineers found inspiration in an unexpected place to solve this problem: philosophy books. The concept they seized upon was Ontology. It is a philosophical term meaning the study of being , the discipline that explores what exists in the world and how those things relate to one another. Palantir took this concept, first established by the ancient Greek philosopher Aristotle, and made it the core principle of its software.

Palantir's official documentation defines ontology as "creating a complete picture of an organization's world." It maps datasets and models into concepts like object types, properties, links, and actions, reconstructing real-world things and events within the platform. It serves as a translator that appears as business language to human eyes and as a computable structure to machine eyes. That is why ontology does not end as a mere catalog or schema design tool. It becomes a resilient foundation for end-user workflows.

This is where the concept of the digital twin enters. Many companies misunderstand the digital twin as merely a flashy 3D display , a pretty rendering of a factory in virtual space. The digital twin Palantir speaks of is something entirely different. It is not decoration. It is a working replica that contains both the semantic system and the behavioral system of an organization. According to Palantir documentation, the ontology serves as the organization's digital twin in many environments, encompassing not only semantic elements like objects, properties, and links, but also dynamic elements like actions, functions, and dynamic security. There is a reason it bears the name "twin." When reality changes, the screen must change; when the screen changes, the next action must change.

In the spring of 2025, the American fast-food chain Wendy's ran short of syrup. It was a problem affecting 6,450 stores. In the past, it would have taken 15 people working all day to resolve. Pete Surkein, CEO of Wendy's Quality Supply Chain Cooperative, said at a Palantir customer conference: "We solved the problem in five minutes." How was that possible? Because Wendy's digital twin had connected 3,500 trucks, 34 distribution centers, and 6,450 stores on a single screen. The moment the signal that syrup was running short came in, the system instantly calculated which truck from the nearest distribution center should deliver via which route.

The true power of the digital twin lies in its writability. In advanced deployment environments, Palantir's digital twin functions not as a mere mirror but as a control panel. Changes made in Foundry are backpropagated to downstream systems. It can update ERP status, trigger work orders, or reconfigure supply plans. This is the decisive difference that distinguishes a static digital shadow from a true digital twin. Organizations leap from the level of "seeing what is happening" to "continuously re-optimizing what is happening."

In Q1 2025, Palantir's U.S. commercial segment revenue increased 71% year-over-year. In Q2 of the same year, quarterly revenue surpassed $1 billion for the first time. Market capitalization exceeded $430 billion. The Economist noted that Palantir may be the most overvalued company in history, as it was valued at more than 600 times its 2024 earnings. These numbers tell us something: the market is placing astronomical value on the concept of the digital twin. It is betting money not on technology that gathers data, but on a method that defines reality.

(2) Reconstructing the World Through Objects, Properties, and Links

Alex Karp often asked his technologists: do you see the world like an Excel spreadsheet, or like a story? Most data engineers see the world in rows and columns. For them, the world is a giant spreadsheet. But Palantir reconstructs the world in sentences. They use three elements: objects, properties, and links. It is like the subjects, adjectives, and verbs taught in elementary school language class.

First, let us look at objects. Objects are the nouns that exist in the world. In Palantir's documentation, an object type serves as the blueprint for that category of entity. Take military operations as an example. A tank is an object. A platoon leader is an object. An enemy supply truck is an object. In legacy systems, these were scattered across different database tables. The personnel department's database stored the platoon leader's serial number; the logistics department's database stored the tank's specifications. Palantir places all of these on a single map. In private enterprises, sensors, assets, work orders, and employees are each defined as independent object types. In healthcare settings, patients, visits, diagnoses, and medications become objects.

Next come properties. Properties are the adjectives that describe objects. According to Palantir documentation, a property is a schema definition representing a characteristic of some real-world entity or event. The object "tank" has a property called "fuel level." It also has a property called "current location." A property called "last maintenance date" can be added as well. Palantir supports various primitive data types: strings, integers, floating-point numbers, dates, timestamps. Advanced types like geographic coordinates, geometries, media references, structs, and time series are also included. Up to this point, it does not look very different from a conventional database.

The magic happens with the third element: links. Links are verbs. Palantir's system connects objects to each other with lines and names those lines. The platoon leader rides the tank , that is a link. The tank targets the enemy truck , that is also a link. The enemy truck transports ammunition. A link type is the schema definition of the connection between two object types; a link is a single instance of the relationship between two objects. When these simple sentences accumulate, a grand narrative is constructed. Previously, analysts had to imagine these relationships in their heads. They had to find the platoon leader in the personnel file, find the tank in the logistics file, and then cross-reference enemy positions on a map. Palantir's ontology engine predefines these relationships.

Now the system can reason. This is the revolutionary point. Here is an example. The system discovers an object called "enemy truck." That truck has stopped near an object called "ammunition depot." It is then moving toward the front line. The system combines these relationships to infer a new fact: this truck is likely en route to resupply ammunition. On the commander's screen, what appears is not a mere dot but an operational recommendation: "Supply route interdiction needed." This is the foundation of what Palantir calls software-defined warfare.

The German pharmaceutical company Boehringer Ingelheim applied this principle to drug development. They built an ontology connecting cross-team data and relating terms like targets, genes, and diseases. They created an enterprise knowledge graph on top of their data lake. As a result, approximately 90% of R&D data could be served through a unified one-stop space via a semantic layer. When a scientist searches by gene or disease, relevant data from across silos can be instantly reviewed. There is no longer any need to manually combine data from different sources.

Among Palantir's solution architects, there is a common saying: "The ontology is essentially our company's nervous system." There is another one too: "Once you've mapped something into the ontology, you never again need to argue about what a customer is or what an asset is."

One data engineer put it this way: "I spent two years building data models in Databricks, and I threw it all away in two weeks with the ontology." Data is no longer dead numbers. It becomes sentences made of nouns and verbs, and sentences accumulate into context. Translating the complex real world into a language that computers can understand , that is the heart of Palantir's technology.

B. Destroying Data Silos

(1) The Principle of Unifying Fragmented Data into a "Single Source of Truth"

In the early 2000s, on the battlefields of Afghanistan, U.S. forces were fighting an invisible enemy. But their greater adversary was internal: the login screen. For an intelligence officer to determine a terrorist's location, they had to access more than ten different programs. There was a separate program to view drone footage. There was a separate system for accessing wiretap records. Reports from human intelligence sources, called HUMINT, were filed in paper binders. By the time an analyst tried to synthesize all this information, the terrorist had already fled.

This is the tragedy of data silos. A silo originally refers to the cylindrical storage structures used to store grain. The wheat in one silo does not mix with the barley in the silo next to it. The same was true of organizational data. Each department and agency used different systems, and data formats varied. According to one analysis of the pharmaceutical industry, about 48% of drug development executives reported that data silos undermined cross-departmental efficiency. Teams spent valuable time reconciling which numbers were correct. Resources were wasted not on extracting insights but on verifying data integrity.

There is a scene that commonly unfolds in meeting rooms. The sales team's numbers differ from the finance team's numbers. The plant team says "our system is fine." The logistics team says "the field is a nightmare." Everyone is telling the truth , just different truths. A silo is not malice but structure. When systems are separated, languages separate too. And separated languages slow an organization down.

Palantir did not demolish these silos. Instead, it built bridges over them. Many companies, when attempting data integration, try to tear out all existing systems and build one massive new database from scratch. This is called a data warehouse. But this approach takes years and costs astronomical sums. Above all, it faces fierce resistance from the field. Each department does not want to give up the tools they are familiar with. Palantir's approach was different. It chose virtual integration.

The principle works like this: existing databases are left in place. HR continues using Oracle. Finance continues using SAP. Instead, Palantir's Foundry platform connects to all of these systems. It does not physically move the data. It brings over the meaning of the data. This is called mapping. It tells the system that Customer ID in System A and Client Number in System B actually refer to the same person. Users can then access all data through a single window called Foundry.

This is called a Single Source of Truth, or SSOT for short. It is a data management concept in which important data is stored and updated in a central repository. By routing all information through one authoritative source, it ensures everyone works from the same playbook. There is no longer a need to fight over data sources in the meeting room. The revenue figure brought by the sales director will not differ from the one brought by the finance director. On Foundry, everyone sees the same numbers. When someone modifies data, the change is reflected in real time for all users.

The biotechnology company Biogen applied this principle to its manufacturing floor. They built a single source of truth for manufacturing data, treating in-process data as a single source. The results were remarkable. Instead of waiting for a separate end-of-line test silo, continuous verification became possible. Biogen's head of global analytics said: "The combination of time-series data with manufacturing context enabled advanced applications. Machine learning models could now rapidly assess batch quality."

The economic impact is dramatic. According to research by Veeva Systems, in a siloed environment, a simple protocol amendment required 25 or more manual steps and multiple documents. Using an integrated vault turned it into a single-source, single-step operation. Update times were cut from weeks to minutes. Drug development teams could spend less time searching for data and more time on actual innovation.

It becomes a single-source, single-step operation. Update times were cut from weeks to minutes. This demonstrates how important an integrated data flow across the entire drug lifecycle is for eliminating redundancy and maintaining a single version of the truth for each drug. Palantir extended this integration to all of an organization's data assets. It created an environment where data, logic, and decision-making coexist. Competing versions of data were eliminated, and it was ensured that everyone uses the same correct numbers.

(2) The Design of Metadata, Lineage, and Permissions Systems

Centralizing data is convenient. But it is dangerous. Being able to see all information also means that anyone could steal or manipulate it. Palantir grew up alongside intelligence agencies from its birth. For them, security was not a feature but a matter of survival. That is why they treated data about data , metadata , as more important than the data itself.

Every piece of data in the Palantir system has a tag attached. When was this data created? Who made it? What was the original source? Palantir explains that the ontology can include rich metadata about every field and fine-grained governance over changes. It is a method of recording who defined what, when it was changed, and why it was changed. This record is not an ethical ornament. When an incident occurs in the field, data stands in court too.

Palantir's 2024 Privacy and Governance white paper states: understanding data lineage allows administrators to visualize data flows across the platform. Reading from left to right, one can see how data in the platform flows from ingestion through data transformation to platform applications. A comprehensive view of data lifecycle and interactions provides a clear picture at scale.

Lineage , this concept is like a data resume. It is a mechanism for tracking the path data takes through a system. It visualizes the entire journey from data's origin to its final use. Palantir's data lineage tools provide a holistic view of how data flows through the Foundry platform.

Here is an example. A general is viewing a map on a tablet that shows the predicted movement route of enemy forces. He is curious about how this map was created. With a single click, he can trace the ancestry of this information. He can learn that this route was predicted by AI, and that the AI analyzed it based on drone photos taken 30 minutes ago and wiretapped radio transmissions intercepted an hour ago.

Palantir's 2024 10-K report notes that Foundry identifies the reason data projects fail as the difficulty in understanding and reproducing the steps and methods of building pipelines. To address this, it enables users to track pipelines and follow traces to understand what the rows and columns of a table mean and why they are there. The moment you ask "why is it there," data becomes not numbers but evidence.

This lineage changes politics as well. When someone says "that figure is wrong," the other party can trace the lineage back to find where it went wrong. Arguments become tracing exercises. In the pharmaceutical field, this means a researcher can trace analytical results back to the original instrument files and laboratory notebook entries. Audit trails are maintained at each step. With a proper audit trail in a central clinical data repository, what was submitted to the FDA can be traced back to source data, and post-submission updates are also tracked.

The last element is the permissions system. When silos are broken down, the first fear people think of is: "So will everything be visible?" Palantir says it places fine-grained and flexible security controls within the ontology itself. Permissions can be applied not only to design elements like object types and link types, but also to actual object and link data.

Palantir's data protection and governance documentation lists specific tools. Checkpoint is a Foundry application that enables justification requests before certain sensitive data operations. Sensitive Data Scanner allows administrators to create organization-specific sensitive data definitions and policies for how to handle information when it is identified. Cipher Service obfuscates data through cryptographic operations such as encryption, decryption, and hashing.

This permissions system is not simply a matter of whether you can or cannot open a file. Permissions can be set down to the smallest unit of data. Suppose a CIA analyst and an FBI investigator are looking at the same terrorist file. On the CIA officer's screen, the terrorist's name and overseas hideout information are visible. But the informant's real name is redacted. On the FBI agent's screen, the terrorist's criminal record within the United States is visible, but overseas operational information is not. This is called purpose-based access control. It is not that all data is available just because a user has high rank. Only the data necessary for the specific mission being performed can be viewed.

Palantir tore down the walls of data while simultaneously building the safest vault. It resolved the contradiction between openness and security through technology. Palantir's AIP is built as a traceable and auditable system. It captures a complete audit trail to ensure trust and accountability in responsible human-machine collaboration. It can fully track which data an AI model used, where that data came from, and who had access rights.

C. Human-Machine Teaming

(1) Not Black-Box AI, but a Tool that Amplifies Human Intuition

Many AI companies in Silicon Valley try to exclude humans. Their goal is full automation , because humans make mistakes, get tired, and are emotional. But Palantir's philosophy is the opposite. Alex Karp asks: would you give an algorithm the authority to pull the trigger on the battlefield? Ethics aside, it is strategically foolish as well. AI is skilled at reading patterns but poor at reading context.

Palantir Technologies was founded in 2003 with a mission born from the national security challenges after 9/11: analyzing vast, heterogeneous intelligence datasets to prevent terrorist attacks without undermining civil liberties. Co-founder Peter Thiel envisioned a mission-driven company that would apply software principles similar to PayPal's fraud recognition system to this complex problem. From inception, the company's philosophy was not about replacing human analysts with omniscient AI. It was about intelligence augmentation , providing human decision-makers with superior tools to navigate complexity.

Palantir rejects black-box AI. A black box is a system where input and output exist, but the process in between is unknowable. Most cutting-edge AI systems based on deep learning are like this. Even the developers do not know why AlphaGo placed its stone in a particular position. With the game of Go, that is fine. But with launching a missile or denying a loan, the story is different. It must be possible to explain why a given decision was made.

Palantir's approach employs two core techniques. The first is chain of thought , Chain-of-Thought prompting. AIP logic functions are configured by default to use this prompting. It causes large language models to respond iteratively according to a structured plan. The prompt instructs the LLM to first outline a plan, then execute the plan step by step, and finally provide a final answer. AIP's LLM Debugger gives users insight into the inner workings of AIP logic functions, allowing them to see intermediate steps and tool calls.

The second is a tool delegation strategy. Instead of processing the entire task within the LLM's black box, AIP delegates specific logical tasks to more interpretable tools. This approach allows reliance on trustworthy and more interpretable logic. If the AI system begins exhibiting unexpected behavior due to tool usage, the specific logic and steps of that tool can be inspected to identify where the logic fails.

An experiment was conducted in the medical field. When explainable AI was used instead of black-box AI , specifically, when heatmap explanations for chest X-ray images were provided , radiologists' work performance improved significantly. Explainable AI went beyond merely providing the AI's prediction score. It offered visual explanations of why a particular judgment was made, enabling medical professionals to critically evaluate the AI's recommendations and combine them with their own expertise.

Palantir's philosophy views AI not as a tool that replaces humans but as one that amplifies human intuition and expertise. It is like the Iron Man suit. When Tony Stark puts on the suit, the suit does not fight on its own. The suit maximizes Tony Stark's capabilities. In 2025, this concept was actually validated in the U.S. Air Force's ShOC-N Capstone experiment. Palantir's Maven Smart System and Maverick AI were tested for integration into the dynamic targeting process to improve speed, scale, and accuracy. The joint integrated experiment improved combatant decision advantage and overall situational awareness.

In the 2024 Forrester Wave AI/ML Platform Q3 report, Palantir was named a Leader. It received the highest ranking in current offering. The report stated: Palantir has one of the strongest offerings in AI/ML and has a vision and roadmap for creating a platform that integrates humans and machines into a co-decision-making model. Palantir's Chief Architect Akshay Krishnaswamy said: Palantir AIP powers the most demanding use cases in both the public and private sectors, and is uniquely designed to connect AI directly to frontline operations. We believe our investment in multimodal guardrails for human-AI collaboration, decision-centric ontology, and the full range of capabilities enterprises need to move from AI prototypes to production has been validated.

(2) Transparent Algorithms for Decision Advantage

Whether in war or business, the essence is the same: a continuous stream of decisions. The side that decides faster and more accurately wins. This is called Decision Advantage. In the past, the side with more information won. Now it is different. Information overflows. The battle now is over who can find meaningful signals in that flood more quickly.

Take the Ukrainian battlefield as an example. A reconnaissance drone spots an enemy tank. In the past, it would have taken dozens of minutes for this information to reach the artillery battalion. Coordinates had to be called out, marked on a map, and radioed in. In the Palantir system, the moment the drone sends its video, AI identifies the tank. It recommends the most suitable friendly artillery within firing range. The time this takes is mere seconds.

But the final button is pressed by a human. The system asks on screen: do you wish to strike? The commander reviews the weapon recommended by the AI and the estimated collateral damage, then approves. The algorithm is transparent. It shows with data why it recommended this artillery unit and why it marked these coordinates. The commander becomes not a slave of the machine but the conductor of the machine.

Palantir's 2024 10-K report describes AIP as providing an interface that enables organizations to operationally use AI and LLMs. It mentions safe handoffs between AI agents and human operators, extensive security and audit controls, and integrated human review checkpoints across workflows. Humans are not designed to be spectators applauding from behind. They are designed as brakes that can stop things at any point along the way.

Palantir's writings on AI governance point in the same direction. They tie critical decisions to human-in-the-loop gates, restrict data and applications visible to each role through access controls, and emphasize data minimization and purpose-based controls. This may look like an ethics statement, but in practical terms, it is a product requirement called "auditable automation." Automate, but leave clear accountability. If accountability is not preserved, automation is banned within the organization.

The core of transparent algorithms is the ability to track and audit AI's decision-making process. Palantir's AIP operates on the principle of building AI systems that are interpretable, understandable, and transparent. AI systems must not become black boxes. Users must understand how the system works in order to build trust. Palantir's approach ensures that AI cannot independently execute actions such as military targeting operations or major financial transactions. Nothing is executed without explicit human approval.

Transparency is not merely an ethical requirement but a practical necessity. Human-AI collaboration research shows that information asymmetry and capability asymmetry are two key sources of complementary team performance. The best collaboration outcomes occur when humans and AI each possess different information and capabilities. To effectively use this, each agent must know what information the other uses and what strategy it adopts.

The Human-Machine Collaboration Wargame for Decision Advantage, conducted in July 2025, was an important milestone in validating this concept. This series of experiments represents critical advances in AI integration and joint integration in battle management. By fusing vast sensor data with machine learning and human oversight, it improved situational awareness and accelerated military decision-making.

Palantir's ontology is the infrastructure that provides transparency. Data is ingested into Foundry or Gotham, contextualized by the ontology, and activated by AIP, enabling humans and AI to run simulations and propose actions. When decisions are made, those actions and outcomes are recorded back into the ontology. This feedback loop enriches the digital twin over time, making future AI-based recommendations more accurate and enabling organizational leaders to learn from the effects of their decisions.

The long-term value of transparent algorithms is the facilitation of organizational learning. Palantir's closed-loop system creates a compounding advantage where the more an organization uses the platform, the smarter and more indispensable it becomes. This is not simply about collecting more data. It is about systematically capturing the context of decisions, their outcomes, and lessons learned, and reflecting these in future decision-making. Decision advantage does not come from a one-time technological edge but from a culture of continuous learning and improvement. Palantir provides the technological infrastructure that makes such a culture possible.

In 2025, Palantir's annual revenue was approximately $4.4 billion , a 53% increase year-over-year. Operating margin reached 51%. It had no debt and held more than $4.5 billion in cash. As of January 2026, the stock price was trading between $185 and $195. These numbers tell us something: the market is placing enormous value on the concept of decision advantage. Technology computes; humans decide. And the speed of that decision determines victory or defeat. This is what Palantir calls the proper relationship between technology and humans.