PALANTIR: War, Surveillance, Artificial Intelligence

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

Chapter 4. Palantir's Four Major Platforms

Attorney Kyungjin Kim

A. Gotham: The Watcher in the Dark

(1) A Person-and-Relationship Tracking System for Counterterrorism Operations and Intelligence Agencies

One day in 2008, in an analysis room deep beneath CIA headquarters, a young officer was staring at her screen. Her mission was to track the network of al-Qaeda operatives. In the past, this would have required manually searching dozens of databases. But on the screen, the relationship web between individuals was unfolding in real time like a spider's web. When she clicked on one person, his call records, financial transactions, travel routes, and known associates were connected in a single picture. This was Gotham's first operational deployment.

Gotham was the first platform Palantir launched, in 2008. The name was taken from Gotham City in the Batman franchise. It was no coincidence. In Christopher Nolan's film The Dark Knight, Lucius Fox builds a system that surveils all of Gotham City. In the movie, Batman destroys the system after the mission is complete. The real-world Gotham was not destroyed. Seventeen years later, in 2025, Gotham has become even more powerful.

Gotham's core function is person-centric analysis. The platform integrates data from diverse sources to construct a complete picture of a specific individual or organization. Phone call records. Email metadata. Financial transaction histories. Immigration records. Social media activity. Satellite imagery. Drone footage. All of these fragments are reconstructed into a unified view through ontology technology.

Gotham's interface integrates multiple analysis tools into a single workspace. Geospatial analysis visualizes the movement patterns of specific individuals or vehicles on a map. Network analysis represents relationships between individuals as nodes and links, revealing hidden connections. Call record analysis detects abnormal activity in phone call patterns. AI and machine learning algorithms analyze all of this data in real time, automatically detecting patterns that an analyst might miss.

Gotham also features Mixed Reality capabilities. Commanders can collaborate in virtual operations rooms and visualize battlefield situations in three dimensions. Satellite integration enables acquisition of real-time reconnaissance imagery from anywhere in the world. Algorithms automatically coordinate which sensors should collect which data.

Security is the core of Gotham. The platform is equipped with multilayered permission systems including classification levels, the need-to-know principle, and time-limited access. Every query and analysis activity leaves a complete audit trail. Who saw what and when is recorded. This is a mechanism for ensuring accountability, but paradoxically, it is also a mechanism for surveilling the surveillance system itself.

Gotham's client list is not publicly disclosed. However, certain known user agencies exist: the CIA and NSA, among other intelligence agencies; the U.S. Army and Navy; Europol. The Danish police have been using Gotham since 2017 for a predictive policing project called POL-INTEL. In 2025, Denmark decided to expand Gotham's use to its military and intelligence agencies.

The most notable recent contract is the $10 billion deal with the U.S. Army concluded in mid-2025. This 10-year contract consolidated 75 individual software contracts into one. Project Maven has also been expanded. This contract, valued at $795 million in 2025, could grow to as much as $1.3 billion by 2029.

Gotham is not a mere analysis tool. As of 2025, the platform supports AI-based kill chains. It automates the process from target identification to strike asset assignment. Of course, the final decision remains with humans. The human-in-the-loop principle is still maintained. But every step before that is handled by algorithms.

This is the essence of Gotham. An eye that sees everything in the darkness. The question is what that eye is watching, and who controls it.

(2) The Truth and Fiction of the Bin Laden Kill Operation (Neptune Spear) Support Claim

In the early hours of May 2, 2011, two U.S. special forces helicopters landed in a residential compound in Abbottabad, Pakistan. Forty minutes later, Osama bin Laden was dead. The code name of the operation, carried out by Navy SEAL Team Six, was "Neptune Spear." After the operation, a question arose: was Palantir involved?

There is no official answer to this question. Neither the CIA nor Palantir has confirmed it. But there is circumstantial evidence.

The tracking of bin Laden was the result of years of intelligence gathering. Between 2002 and 2005, the CIA learned from al-Qaeda prisoners about the existence of a courier named Abu Ahmed al-Kuwaiti. It took another six years to track this courier. In 2010, CIA officers finally located him in Pakistan and followed him to the residential compound in Abbottabad.

The likelihood that Palantir's Gotham was used during this tracking process is high. The CIA had been one of Gotham's earliest clients since 2008. Person tracking, relationship network analysis, and geospatial analysis are precisely what Gotham was designed for , integrating dozens of data sources to construct a unified picture. The bin Laden operation required exactly such capabilities.

Palantir has never directly claimed this connection. However, it has implicitly used it. In the company's marketing materials and investor presentations, the phrase "success in counterterrorism operations" frequently appears. It is a way of stimulating the audience's imagination without mentioning specific operation names.

The truth lies within classified secrets. Immediately after the 2011 operation, Admiral William McRaven, commander of Special Operations Command, ordered all files related to the bin Laden operation deleted from Defense Department computers and transferred to the CIA. Physical evidence has never been released. No photographs, no video, no DNA test results have been made public. All FOIA requests have been denied.

What is certain is this: the bin Laden operation was a case where "connecting intelligence" was the key to victory. The ability to assemble fragmented clues into a single picture , this was exactly what U.S. intelligence agencies desperately needed after 9/11, and what Palantir promised to provide. The operation's success suggests that promise was, to some extent, fulfilled.

Whatever Palantir's role was in the bin Laden kill operation, the operation contributed to strengthening the mythology of the company. The image of "technology that catches terrorists" became one of the company's most powerful weapons. Whether that image is based on fact or carefully cultivated remains shrouded in mystery.

B. Foundry: The Central Nervous System of Enterprise

(1) The Structure of a Data Integration Platform for the Private Sector

In 2016, Palantir stood at a critical turning point. Gotham had achieved success with intelligence agencies and the military, but the government market alone had limits for growth. The company needed to enter the private enterprise market. But companies were reluctant to apply software used by the CIA to their own businesses. The surveillance technology image was the problem. A new brand was needed.

That is how Foundry was born. The name, meaning "a place where raw materials are cast," carries the meaning of transforming raw material (data) into valuable products (insights). The core technology was the same ontology engine as Gotham, but the packaging was completely different. Instead of surveillance, "digital transformation." Instead of tracking, "operational optimization." The language changed.

To understand Foundry's structure, think of three layers.

The first is the data connection layer. Enterprises operate dozens, sometimes hundreds, of different systems , ERP, CRM, production management, logistics, HR. These systems typically do not communicate with each other. Foundry connects to all of these systems and ingests data. Whether through real-time streaming or periodic batch uploads, it ingests regardless of data format or source.

The second is the transformation layer. Ingested data is cleaned and standardized. Throughout this process, data lineage is tracked. Where each piece of data came from and what transformations it underwent are all recorded. This ensures auditability and enables root cause tracing when errors occur.

The third is the ontology layer. This is where the magic happens. Cleaned data is transformed into business objects. For an airline, for example, "aircraft," "route," "passenger," "maintenance record," and "part" each become objects. These objects have properties. An aircraft object has properties such as model, manufacture date, total flight hours, and current location. And objects are connected by links. This aircraft operates on this route, is composed of these parts, and has these maintenance records.

This is the "digital twin" , replicating the complex operations of the real world in the digital world. However, this replica is updated in real time, and simulation and analysis are possible.

Foundry's strength is "closed-loop" operation. It does not end with data analysis. Analysis results are immediately connected to real actions. For example, once a supply chain optimization analysis is completed, the results can be automatically reflected in the ordering system. The time from insight to action is dramatically shortened.

As of 2025, Foundry is recording explosive growth in the U.S. commercial sector. U.S. commercial revenue in Q1 2025 increased 71% year-over-year. The number of clients grew 43% to 849. There were 139 contracts exceeding $1 million, 51 exceeding $5 million, and 31 exceeding $10 million. Customer retention rate reached 98%.

The secret to this growth lies in the "Bootcamp" strategy. Traditional enterprise software sales took months. Palantir compressed this to five days. Potential customers are invited to experience Foundry with their own actual data. A working prototype is built in five days. The customer sees the value firsthand and then decides to purchase. This method has dramatically shortened the sales cycle.

(2) Application Cases in Manufacturing, Finance, and Healthcare

An engineer at Airbus headquarters stared at the screen. Data from thousands of aircraft operating worldwide was streaming in real time. Engine temperatures, fuel consumption, component wear status. An algorithm analyzed the patterns and raised an alert: a specific aircraft model's landing gear component was wearing out faster than expected. A replacement schedule was automatically generated before the part could fail.

This is Skywise , an aviation data platform jointly developed by Airbus and Palantir. Launched in 2017, more than 100 airlines currently participate in this platform. Aircraft operational data, maintenance records, and parts inventory information are all integrated. With predictive maintenance now possible, airlines have been able to reduce delays caused by unplanned maintenance.

In the energy sector, BP is a representative case. BP used Foundry to build a "digital twin" of its global supply chain. The entire process from crude oil production to refining, transportation, and sales is monitored in real time. Algorithms automatically perform demand forecasting, inventory optimization, and logistics route planning. During the 2022 energy crisis, this system played a critical role in rapidly responding to volatile market conditions.

The healthcare case is controversial. In November 2023, the UK's NHS signed a contract with Palantir worth 330 million pounds over seven years. The purpose is to build a "Federated Data Platform." The goal is to connect patient data from hospitals across England to optimize bed management, surgical scheduling, and resource allocation.

The contract provoked immediate backlash. The British Medical Association (BMA) and the Doctors' Association UK raised concerns about patient data privacy. Cybersecurity experts questioned the transparency of the procurement process. In April 2024, medical professionals protested outside NHS England headquarters. Their demand was for the NHS contract to be canceled because of Palantir's contract with the Israeli Defense Forces (IDF).

As of early 2026, only about 15% of NHS trusts are actually running the platform. The adoption rate is slower than expected. Political and ethical debates, more than technical issues, are holding it back.

In manufacturing, the partnership with Samsung is drawing attention. In 2025, Samsung began a project in collaboration with Palantir to improve semiconductor yield and quality. The goal is to analyze the massive volumes of data generated in semiconductor manufacturing processes, identify causes of defects, and optimize process variables.

In Korea, collaboration with HD Hyundai is also underway. Targeting the digital transformation of shipbuilding, Palantir, HD Hyundai, and Siemens are jointly pursuing a project to automate shipyard operations. KT has taken on the role of partner to spread Foundry and AIP across Korean industry.

Foundry's success cases are impressive. But criticism also exists. Foundry's ontology is a proprietary and closed structure. Once adopted, it is difficult to leave. While data export functionality exists, the business logic embedded in the ontology cannot be taken along. This is another reason explaining the 98% customer retention rate.

C. Apollo: A Revolution in Software Deployment

(1) Seamless Deployment from Cloud to Battlefield Edge

"We bring SaaS to places SaaS has never been."

This is the phrase Palantir uses to describe Apollo. The back seat of a Humvee. Inside a submarine. A satellite. The idea of cloud software operating in such environments was hard to imagine until recently. Apollo made it possible.

To understand why Apollo exists, one must look back at Palantir's history. When Gotham was first deployed in 2008, cloud computing was in its early stages. Most enterprise software was installed on the customer's own servers. Government agencies and the military, in particular, could not use external clouds for security reasons. Software upgrades were rare and painful.

There was a problem. Palantir's software consists of hundreds of individual services. Each service is managed by an independent development team. Deploying and updating these services simultaneously across dozens of different environments , cloud, on-premises, classified networks, and fully isolated air-gapped environments , was a nightmarish task.

Apollo was created to solve this problem. The core idea is a "hub-and-spoke" architecture. There is a central Apollo hub, and spokes (deployment platforms) are installed in each deployment environment. The hub manages all versions of all products. The spokes "pull" updates suited to their own environment.

The "pull" method is important. Traditional software deployment uses a "push" method: developers push out updates. This assumes the developer knows the state and constraints of every environment. As environments multiply, management becomes impossible.

In Apollo's pull method, each environment defines its own constraints. This environment only follows "Release Channel Stable." That environment also accepts "canary" versions. Apollo automatically orchestrates updates while respecting these constraints.

The numbers prove it. Apollo processes more than 41,000 automated updates per week. Lead time for applying changes is under four minutes. Blue-green deployment and automatic rollback capabilities ensure stability. If a problem occurs, the system can immediately revert to the previous version.

Operation in edge environments is Apollo's true differentiator. Environments where internet connectivity is intermittent or nonexistent. Environments where bandwidth is extremely limited. Apollo is designed to function even in such settings. It operates with minimal resources and synchronizes with the central hub when connectivity is restored.

Consider why this matters in military operations. A commander in the field may need to conduct operations for days without connection to headquarters. During that period, the Palantir software installed on his equipment must continue to function. If new threat intelligence becomes available, it must be updated the moment connectivity is restored. Apollo makes this possible.

(2) Lock-in Effects and the SaaS-Based Revenue Model

Apollo is not only a technological achievement but also a business model.

In the traditional enterprise software model, customers purchased and installed software. Upgrades were optional. Many companies used the same version for years. This was unfavorable for software companies. If the customer did not see value in the new version, there was no additional revenue.

The SaaS (Software as a Service) model solved this problem. Customers subscribe instead of purchasing. Software runs in the cloud, and updates are applied automatically. Customers always use the latest version. The company earns continuous subscription revenue.

The problem was that Palantir's main customers , intelligence agencies, military, and heavily regulated companies , could not use the traditional SaaS model. Data could not leave for an external cloud. But reverting to on-premises installation meant losing SaaS's advantages.

Apollo created a new model called "Private SaaS." Software runs in the customer's environment, but updates and management are centrally automated like SaaS. The customer maintains data sovereignty while continuously receiving the latest features. Palantir maintains subscription-based revenue.

The lock-in effect arises here. Palantir software deployed through Apollo becomes deeply integrated into the customer's operational processes. Data pipelines, analytics workflows, and decision-making processes are all built on top of the platform. Over time, this dependency deepens.

What would it take to replace Palantir software with something else? All data integrations would need to be rebuilt. The ontology would need to be redesigned. Users would need to be retrained. The risk of operational disruption would need to be accepted. For most organizations, this is a practically impossible choice.

Average contract duration of 3 to 4 years. A 98% customer retention rate. A 124% net dollar retention rate. These numbers reflect the lock-in effect. A net dollar retention rate exceeding 100% means existing customers are spending more money each year.

Apollo is also available on the AWS Marketplace and Microsoft Azure Government Cloud. It has obtained IL5 and IL6 certifications, enabling operation on classified networks. It also holds FedRAMP certification. These certifications represent barriers to entry that competitors cannot easily overcome.

Apollo is Palantir's unsung hero. It does not receive as much attention as Gotham or Foundry, but it is Apollo that enables these two platforms to operate simultaneously across thousands of environments worldwide. By solving the seemingly mundane problem of software deployment, Palantir has been able to plant its software in places no other company has reached.

D. AIP (Artificial Intelligence Platform): Deploying Generative AI in the Field

(1) Controlling LLM Hallucination and Applying It to Real Work

In April 2023, Palantir announced a new platform: AIP, the Artificial Intelligence Platform. It was a tool for integrating large language models (LLMs) like ChatGPT into enterprise environments. In the announcement video, CEO Alex Karp said: "We are landing LLMs in the real world."

The word "landing" was apt. LLMs have remarkable capabilities, but they also have a serious problem: hallucination , the phenomenon where a model confidently states information that is not true as though it were fact. If an LLM occasionally says something wrong in a casual conversation, you can laugh it off. If an LLM hallucinates during a military operation, medical diagnosis, or financial decision, it becomes a disaster.

Palantir's solution was the ontology. Once again, the ontology.

In AIP, the LLM does not generate answers freely. It is "grounded" in the ontology. When the LLM receives a question, it first searches for relevant data in the ontology. Answers are generated based on verified enterprise data. Sources are cited. Users can verify which data the AI's answer came from.

Here is an example. A logistics manager asks: "What is the current inventory status of the Chicago warehouse?" A typical chatbot would answer based on training data , probably outdated information. An AIP agent retrieves data from the real-time inventory system connected to the ontology and responds. The answer includes the last data update time.

AIP supports multiple LLMs: OpenAI's GPT series, Anthropic's Claude, Google's models, Meta's Llama, and NVIDIA's Nemotron. Customers can select the appropriate model for their use case. For security-sensitive cases, self-hosted models running on-premises can be used.

In August 2024, Palantir announced a partnership with Microsoft. Palantir's entire product suite will be deployed on the Microsoft Azure Government Cloud. This means GPT-4 can be used on classified networks for the first time. Military and intelligence agency analysts can now harness the power of large language models.

Trust is core to AIP's design. The platform provides detailed audit trails for all AI decisions. Who asked what question, what data the AI referenced, and what logic it used to generate the answer , all recorded. There is also a "proposal-based" pattern. The AI does not execute actions directly but only makes proposals. Execution requires human approval.

However, Palantir itself acknowledges the limitations of LLMs. Its documentation warns: "Synthetic reasoning carries the risk of hallucination." Ontology grounding is effective for data retrieval, but errors can still occur in the process where the model interprets and reasons with data. Users must understand this limitation.

(2) Agent-Based Task Automation and Evolution Toward an "AI Operating System"

2023 and 2024 were the years of the chatbot. 2025 is the year of the agent.

Let me explain the difference. A chatbot answers questions. An agent acts. Ask a chatbot "What's the weather tomorrow?" and it tells you the weather. Tell an agent "If it rains tomorrow, change the outdoor event schedule" , it checks the weather, and if rain is forecast, it actually changes the schedule.

AIP Agent Studio makes this possible. Users can build agents. An agent combines an LLM, the ontology, documents, and custom tools. Agents are context-aware. They dynamically execute workflows.

Let us look at real cases. A logistics company can deploy an agent that automatically reroutes deliveries when storms are forecast. A hospital can use an agent that manages nurse schedules in real time. An energy company can run an agent that automatically adjusts the power grid during demand surges.

In the defense sector, the Maven Smart System has evolved into a fully integrated AI target acquisition and situational awareness tool. It is now used as a standard by multiple NATO member nations. The TITAN (Tactical Intelligence Targeting Access Node) ground station serves as the "brain" of the U.S. Army's AI-defined battlefield.

Let us look at AIP's other components. AIP Logic is a no-code LLM function builder. Even business users who are not developers can create LLM-based functions. AIP Assist is a chatbot available across the entire platform. It can search not only Palantir documentation but also custom documents uploaded by customers. Pipeline Builder can include LLM nodes to automate data classification, sentiment analysis, summarization, and more.

The numbers speak to AIP's impact. In Q1 2025, Palantir signed 139 contracts worth $1 million or more. 51 contracts over $5 million. 31 contracts over $10 million. A significant number of these contracts are related to AIP.

Palantir positions AIP as an "AI operating system." Just as an operating system is the middle layer between hardware and applications, AIP aims to become the middle layer between raw data and AI applications. Any LLM, any data source, any application connects through AIP.

CTO Shyam Sankar led the transition to agentic AI in 2025. According to his vision, future enterprise operations will be a collaboration between humans and AI agents. Agents handle repetitive and data-intensive tasks. Humans focus on decisions requiring judgment and creativity.

Of course, risks exist. Allowing AI agents to act autonomously means relinquishing some control. If an agent makes a wrong decision, who is responsible? How do you ensure an agent's actions are ethical? These questions have not yet been fully answered.

Palantir's answer is, again, transparency and auditability. All agent actions are recorded. Actions requiring approval are distinguished from those that can be performed autonomously. However, while technology advances rapidly, governance and regulation are struggling to keep pace.

AIP is Palantir's present and future. If Gotham and Foundry laid the company's foundation, AIP is the growth engine. Since its 2023 launch, the company's stock has soared. It was added to the S&P 500 at the end of 2024. In November 2025, the stock price reached an all-time high of $207. Market capitalization surpassed $400 billion.

This is what it means to deploy generative AI in the field. Beyond chatbots, AI that acts in the real world. Controlling hallucination, grounding in data, making it auditable. Palantir is presenting one answer to this difficult problem. Whether that answer is the only one, or the best one, is still too early to judge.