Palantir CEO Alex Karp's 'Institutional Sovereignty' — his framework for enterprise and government autonomy from AI vendor dependency — names what OpenAI and Anthropic won't: customers pay per token while labs harvest the data.
Alex Karp's white paper reads like Palantir marketing. The mechanism inside it isn't.
The chief executive of data-analytics firm Palantir Technologies (PLTR) released a 15-step manifesto this week, titled "Institutional Sovereignty in the Age of AI," laying out a defense plan for enterprises and governments against the frontier model labs whose models they depend on. The paper arrived a week after Karp's heated TV appearance, where he declared that "something has gone completely wrong" in the relationship between AI labs and their enterprise customers. Karp told the Wall Street Journal this week that AI labs hype their capabilities and overcharge for tokens, the metered units of AI model use that labs charge per. He separately called the AI industry "effing insane" in a July Forbes interview and criticized closed-model token pricing in Business Insider's coverage of a CNBC appearance.
Karp is a direct competitor with OpenAI and Anthropic for enterprise AI spend, and the 15-step paper is also a product pitch: Palantir's ontology-and-platform stack is the implicit answer to most of its questions. That tension is part of why the wire coverage frames this as an angry CEO venting. The more useful read is that Karp has named a structural mechanism that any enterprise AI buyer should pressure-test, independent of who is selling the answer.
The mechanism has three parts.
First, data-and-insight extraction. Frontier labs fine-tune their models on customer workflows. The customer pays for tokens; the lab keeps the upgraded model. In Karp's telling to the WSJ, customers describe paying "for tokens that create no value" while the lab captures the "magic sauce" behind what makes the company successful. A Hacker News discussion of the piece frames the dynamic more bluntly: "if you're an AI product then the model product synthesis is your product, and you don't get any of that from a proprietary closed model."
Second, subsidized first-party pricing against API customers. Once a frontier lab ships its own competing product, the lab can price that product at a steep loss and recover margin elsewhere in the stack. The HN thread pointed to Anthropic's Claude Code: community commenters alleged that Anthropic could subsidize Claude Code at API-cost-plus while charging other AI coding tools full API rates for the same underlying model, putting those tools at a structural disadvantage. The specifics of that pricing dynamic are community testimony rather than confirmed business reporting, and the framing is sharper than the available evidence supports. The dynamic itself is the kind of conflict that exists in any platform market where the platform owner ships a first-party competitor to its API customers.
Third, ontology and ownership capture. The lab's model becomes the authoritative map of the customer's domain, and whoever defines the ontology controls how decisions are described, scored, and audited. Karp's paper argues that this layer is the asset and should sit with the customer or the government, not with the model vendor. "Institutional Sovereignty" in Karp's usage is shorthand for enterprise and government autonomy from AI-platform dependency.
Three concrete tests follow from this framing, and they apply whether the vendor on the other side of the table is OpenAI, Anthropic, or Palantir.
Test the data clause. Ask which model improvement loops include customer inputs by default, and whether the contract requires opt-in consent, separate compensation, or a hold-back. If the answer is "we use everything to improve the model unless you pay extra," the asymmetry Karp names is already in the contract.
Test the product roadmap. Ask whether the lab ships a first-party product that competes with anything in the customer's stack, and at what price floor. A lab that sells API access and a competing end-user product at a loss is a channel partner for some workloads and a competitor for others. Procurement needs to know which.
Test the ontology. Ask where the canonical representation of the customer's domain lives, who can edit it, and who can audit the model's reasoning against it. If the lab owns the ontology and the customer does not, the customer is renting its own decision logic.
These three questions sit inside a wider tension that Karp, on Palantir's account, is trying to make politically legible. The WSJ noted that the debate is also playing out in bipartisan AI oversight and in U.S.–China competition for tactical edges in compute and deployment. Both are real stakes, and both are downstream of the same enterprise-economics question: who captures the value when a closed model becomes the default interface to a company's data?
The HN discussion made a separate prediction worth tracking. Token costs on the Pareto frontier are reportedly flattening, and enterprise token spend is described as "untenable" by community commenters. If that holds, the buyer playbook Karp is selling shifts: many organizations will default to distilled models, either vendor-distilled like a smaller Sonnet variant or China-sourced. The sovereignty question then moves from "which frontier lab" to "which distilled model do we trust on our data." That is a different negotiation, with different vendors and different risks.
The white paper is Karp-shaped and serves Palantir's positioning. The mechanism inside it does not require buying Palantir to be useful. It requires buyers to ask three questions before the next AI-vendor renewal. Karp has put those questions on the public record this month. Whether OpenAI, Anthropic, or any of the other labs answers them in writing is the next data point worth watching.