Palantir CEO Alex Karp argues in a new white paper that AI labs extract customer decision data through token billing — the per query units they charge for — then use that spend to fund competing products in the same markets, naming the procurement
Every dollar an enterprise spends on AI inference now doubles as an R&D budget for the lab that wants to compete in the same category. That is the mechanism Palantir CEO Alex Karp named this week in a formal white paper and a nearly 20-minute CNBC appearance covered by Hindustan Times, and it is the procurement problem enterprise technology buyers have been quietly sitting on since the frontier labs moved from API-only access to category-specific products.
Tokens, the units AI vendors bill per query or per document processed, look like a usage fee. In practice they are a continuous data handover: every prompt, every retrieved document, every tool call carries a customer's operational context, decision patterns, and institutional know-how into the lab's models and product roadmap. Palantir's white paper, published this month and summarized in a company post on X, calls the result an institutional sovereignty risk: an enterprise's compounding advantage bleeding into the lab that prices its tokens.
On the technical side, fine-tuning, retrieval-augmented generation, and tool-use traces all send labeled customer behavior into the lab's training and evaluation pipelines, often under permissive data-use clauses the customer signed at procurement. On the commercial side, frontier labs have moved in the past year from selling raw model access to shipping branded products in customer categories. Former White House AI policy lead David Sacks, amplifying Karp's argument on social media, named the pattern in a single sentence: Anthropic has launched Claude Science, Claude Security, Claude Legal, and Claude Code, each expanding into categories previously served by companies building on top of the lab's models.
The white paper's prescription is to anchor value to assets the enterprise uniquely owns, including decisions, workflows, permissions, context, and institutional memory, and to stop treating model leadership as a moat. Model leadership, in Palantir's framing, may flip every few quarters. Institutional memory compounds over years. The 15 enumerated steps in the paper run from data-egress controls and on-prem deployment to forbidding model-training use of customer inputs by contract. A summary by NextBigWhat groups the 15 steps under five strategic categories, including protecting operational data, controlling the deployment surface, and retaining decision rights against the gravitational pull of the lab's roadmap.
A buyer who treats AI spend as a SaaS line item is reading the wrong contract. The clauses that matter are the data-use license (does the lab get to train on customer inputs?), the model-update cadence (does every fine-tune transfer learnings back to the shared base model?), the output ownership clause (who owns the model's outputs in the customer's domain?), and the substitution clause (can the lab ship a competing product in this customer's category during the contract term?). Each of these is negotiable today. Each will be harder to renegotiate once the spend base is locked in.
Karp's framing should be read as both a structural critique and a sales pitch. Palantir sells an ontology and decision-stack platform whose value proposition rises as enterprise buyers grow skeptical of the frontier labs. The argument is more credible for being true: the data-use and substitution risks are real, and the labs' product roadmaps confirm them. The reader should treat the mechanism as real and the solution as one option among several, including on-prem model deployments, hosted open-weight models, and in-house fine-tunes with explicit training carve-outs.
The next trigger to watch is whether OpenAI, Anthropic, or any other frontier lab publicly responds to the white paper. None had responded as of the paper's release; the labs' commercial incentive is to keep data-use clauses permissive and substitution clauses unenumerated. The renewal cycle for enterprise AI contracts this fall will be the first stress test of how seriously buyers read the clauses they signed twelve months ago.