Why Palantir’s Ontology Fuels Its Valuation: The Skeleton and Memory Behind AI

On March 19, 2026, the DataFun community hosted a deep‑dive live dialogue titled “Ontology and AI.” The session featured three senior practitioners—Lü Hang‑fei, a systems‑intelligence architect, Hu Shen‑min, head of intelligent R&D at Shanghai Bank, and Xi Zong‑zheng, senior R&D engineer at Alibaba Cloud—who examined why both banking risk‑control and cloud operations teams converge on Palantir’s ontology.

The discussion began by identifying three “gaps” that arise when no ontology exists. The first is the **data gap**: raw logs, metrics and traces are noisy, fragmented, and over 99 % of alerts are irrelevant, forcing engineers to manually stitch together information from disparate consoles (Prometheus, SLS, Trace APM). The second is the **model gap**: AI models act as black boxes, often hallucinating or mis‑attributing causality because they lack prior knowledge of system topology. The third is the **engineering gap**: enterprises ingest petabytes to exabytes of data daily, creating massive challenges in cleaning, storing and computing on that scale.

Hu Shen‑min added a data‑governance perspective, noting that without an ontology, vector‑based retrieval (RAG) returns facts without context, making it impossible to reconstruct a coherent scenario.

To illustrate how an ontology resolves these gaps, Xi Zong‑zheng introduced the **U‑model**. In the U‑model, a *Set* represents a node (e.g., an ECS instance) and a *Link* represents an edge (e.g., “runs on”). A server entity is defined with attributes such as IP, CPU, memory, OS, and status; a micro‑service running on that server is another entity linked by a “runs‑on” relationship. Observability data—metrics and logs—are attached to the corresponding entities via a “data” relationship, forming a minimal observable unit that can be traversed by an AI agent for automated fault isolation.

Hu Shen‑min distinguished **business ontology** (defining core business objects, processes, behaviors, rules, and data mappings) from **technical ontology** (aligning with APIs, databases, and digital‑twin representations). He emphasized that a well‑structured process model—derived from IBM’s five‑level modeling (domain, value chain, activity, task, step)—is essential for AI explainability and regulatory compliance in finance.

The panel then compared the roles of large models and ontologies. Xi described large models as the “brain” and ontologies as the “skeleton plus memory.” The skeleton defines the IT world’s structure (entities and relationships), while the memory stores expert knowledge in layered runbooks. Toolkits and skills expose actionable capabilities (query metrics, fetch logs, execute commands) that the AI agent can invoke dynamically—a capability Xi called “reflection.”

Graph‑based Retrieval‑Augmented Generation (Graph RAG) was presented as a “blood‑meat supplement”: the knowledge graph supplies precise entity relationships, while the large model provides contextual reasoning. To avoid overwhelming the model, the ontology is loaded incrementally using a directory‑like catalog that lets the model fetch only the needed sections.

When asked for advice to teams starting ontology projects, Xi warned against attempting a full‑scale rollout (“don’t try to model everything at once”) and against over‑engineering (“don’t build ontology for its own sake”). Instead, he recommended selecting the most painful scenario, building a minimal viable model, and iterating. Hu stressed the need for deep business involvement—knowledge must be co‑created with domain experts rather than imposed by a technical team alone.

Looking ahead, the speakers agreed that ontologies will not be superseded by AGI; rather, they will become a foundational operating system for future intelligent agents, providing the structured, trustworthy knowledge that large models alone cannot supply.