Rent intelligence by the token and your alpha transfers to the model landlord. Put a neutral control plane between your data and the model, and the models become the swappable part. Toggle the two paths and watch where the alpha goes.
Green dots are your alpha: proprietary data, workflows, the decisions that make your business worth more than its inputs. Pick a path, then try swapping the model.
Palantir CEO Alex Karp, on CNBC this July: "What the technical customers want is control over their compute, their models, their data stack, and their alpha." Four nouns. Two public companies sell the layer that satisfies them without giving up frontier models.
Compute → owned GPUs, on-prem or contracted, not a lab queue. Models → open weights you can tune inside your own perimeter, or closed models held at arm's length. Data stack → stays inside the perimeter; the model sees only what the control plane brokers. Alpha → the workflow knowledge that stays on your balance sheet instead of training the landlord's next product. The first two are an iron purchase. The last two are a control-plane purchase.
Palantir's Ontology maps the business itself — the objects, relationships, and decisions that make up how a company actually operates — and sits between the raw data and whatever model is doing the reasoning. AIP runs the agents on top. Karp's pitch is that battlefield LLM deployments already run on top of this layer: the model can be anyone's, the operational graph is Palantir's. That is a control-plane business wearing a defense contractor's jacket.
ServiceNow spent $11.6B assembling the same layer for the enterprise: AI Control Tower governs any agent on any model — OpenAI, Anthropic, Google, custom — against the CMDB, the system of record for what the enterprise owns and how it connects. Veza brings identity (who can access what), Armis brings visibility (what is actually on the network), Moveworks brings orchestration. The platform is model-agnostic on purpose: the value is not in the model, it is in the layer every model depends on.
Why would an enterprise fear its own model vendor? Because the lab can watch where value pools on top of its model, then ship that vertical itself. Per The Information's reporting, Figma found out how fast.
Karp is talking his book — Palantir sells the layer he says enterprises are missing. But the market has already priced the book, and it has priced the same layer twice, at two very different levels.
Forward P/E, as of early July 2026 — multiples drift; check the tape. PLTR carries the richest multiple in large-cap software: that is what the market pays when it believes the control plane captures the value. NOW sells the same layer with the stock down more than half over the past year. One name priced for perfection, one priced for doubt. Neither is a recommendation; the spread is the point.