Palantir Technologies Inc. (PLTR) Deep-Dive: The Operating System for Deploying Large Models — Can It Carry an AI-Infrastructure Valuation Far Above SaaS?

Thesis verdict: holds, with conditions — AIP and the Ontology genuinely solve the physical bottleneck of putting large models to work, driving 149% U.S. commercial growth and a 62% margin; but an extreme, long-duration valuation is highly fragile at 5.00% interest rates, and receivables concentration plus termination-for-convenience clauses add risk.

Not selling compute or models, but the "data operating system" between the enterprise and AI

Palantir is a software company that provides a data operating system to governments and large enterprises. It sells neither the underlying compute chips nor a general-purpose large model. Instead it delivers platform-level software — Gotham, Foundry, Apollo and the Artificial Intelligence Platform (AIP) — and customers pay by subscription or by actual usage. The company's core value lies in solving a real pain point: when a large institution tries to adopt generative AI, it discovers that the underlying model can neither understand nor reach the highly confidential, scattered business data and logic inside the organization. Palantir's software is the middle layer that safely stitches together private enterprise data, business-process rules and external large models.

Looking at the business base, Palantir's revenue comes from a commercial segment and a government segment of comparable size, but the incremental contribution is tilting decisively:

In the first half of fiscal 2026 (H1 2026), the commercial segment contributed $1.72B of revenue, 48.2% of the total; of the $1.68B total year-over-year increase, commercial delivered $872M, or 51.9%. The core driver of the commercial surge is U.S. commercial customers. On the news side, the company struck a sovereign-AI infrastructure partnership with Nebius and Nvidia (2026-09-09); management raised full-year U.S. commercial revenue guidance to above $3.424B (implying growth above 134%); and on quantified metrics, U.S. commercial remaining deal value (RDV) jumped 124% year over year.

The government segment contributed $1.85B of revenue in H1 2026, 51.8% of the total, with an $808M year-over-year increase (48.1% of the increment). Government business rests on deep budget support; the company recently won the U.S. Army TITAN order (2026-09-02), and overall government growth re-accelerated to 90%.

On contractual obligations: as a pure software and subscription business, the key forward indicator is RPO (remaining performance obligations, i.e. contracted orders not yet recognized as revenue). Total RPO currently stands at $4.9B, 43% of which will be recognized as revenue within the next 12 months. One thing must be made clear, though: according to Note 3 of the 10-Q, the vast majority of customer contracts include a unilateral "termination for convenience" clause (the customer can end the contract without cause after paying an agreed compensation). That makes Palantir's RPO less rigid than traditional SaaS; if the external economy deteriorates, customers hold the legal right to cut contracts at any time.

In the coordinates of the software value stack (the layered architecture of software technology), Palantir's position is unusual:

• It has fully shed layers one and two — raw materials and compute. The company builds no fabs, buys no chips to stack compute centers, and trains no foundation large language model (LLM), sidestepping the heavy-asset depreciation burden of the infrastructure layer.

• It firmly occupies layer four (orchestration, workflow and execution), acting as the steward that schedules foundation models and enterprise data, weaving itself tightly into customers' daily operations.

• It holds absolute control of layer five (security, governance and compliance). AIP's core selling point is building "AI guardrails" for the enterprise, ensuring that when a large model touches core assets it strictly obeys fine-grained role- or attribute-based access controls.

Right now the underlying large models are sinking into brutal price competition and commoditization. Yet the more homogeneous and cheap the base models become, the more a large institution needs a neutral software orchestration layer that can swap between them seamlessly. Palantir uses that to anchor value in the upper layer. The embedding runs deep: its systems directly control Airbus's aircraft-assembly scheduling, BP's real-time well drilling and the military's live operational dispatch. If the software stopped, customers' production and tactical operations would grind to a halt.