The concrete deal behind the rhetoric
On June 29, 2026, Palantir and Nvidia announced a Sovereign AI Operating System Reference Architecture: a full-stack design that runs customized NVIDIA Nemotron open models inside air-gapped environments on Nvidia's Blackwell Ultra GPUs, layered on Palantir's AIP, Ontology, Foundry, and Apollo products. Per Nvidia's own announcement, agencies and infrastructure operators using it can train models on their own data and retain full ownership of the resulting weights, rather than calling a hosted API. This is the product Karp is implicitly selling against when he criticizes token-based pricing, so his public criticism of OpenAI and Anthropic doubles as marketing for a real, already-shipping alternative.
The criticism
Speaking on CNBC's Squawk Box on July 1, Karp said "something has gone completely wrong" with how OpenAI and Anthropic sell frontier models, and argued that many enterprises are unhappy with what he called "tokenmaxxing", per Karp's account, the practice of maximizing token consumption regardless of returned value. He said the prevailing enterprise fear is that they will "chillax and waste my time with tokens, I'm going to get no value and they're going to get my IP." That last claim, that using OpenAI's or Anthropic's hosted models puts customer IP at risk, is Karp's own allegation and has not been independently corroborated by the labs or documented in the available reporting; it should be read as one competitor's characterization, not a verified fact. Karp separately said Palantir now sells a model-agnostic product that lets customers switch between underlying models, and argued that government and defense customers ("warfighters," in his phrasing) have particular trust concerns about relying on any single frontier lab's infrastructure.
Context
The remarks land amid a broader shift in enterprise sentiment on AI costs. CNBC reported days earlier that some organizations are moving away from indiscriminate "tokenmaxxing" toward efficiency-focused deployment, and other reporting has described both OpenAI and Anthropic exploring pricing changes as competition intensifies. Karp's framing draws an explicit line between building a frontier model and safely operationalizing AI inside regulated enterprise or government environments, an argument that benefits Palantir's own defense- and government-heavy customer base.
What to watch
- •Whether OpenAI or Anthropic publicly respond to the IP-risk claim, or adjust enterprise pricing/data-handling terms
- •Adoption of the Palantir-Nvidia Sovereign AI Operating System by named government agencies or infrastructure operators
- •Whether other enterprise vendors follow Palantir in pitching open-weight, self-hosted alternatives as a hedge against per-token pricing
Key Points
- 1Karp's "tokenmaxxing" and IP-risk criticism of OpenAI/Anthropic coincides with Palantir's own competing Nvidia Nemotron-based sovereign AI product launched June 29.
- 2The claim that hosted-API token pricing puts customer IP at risk is Karp's unverified allegation, not a documented practice by OpenAI or Anthropic.
- 3The Palantir-Nvidia Sovereign AI Operating System runs open Nemotron models on air-gapped Blackwell Ultra hardware, giving customers ownership of model weights.
Scoring Rationale
Notable: pairs a widely covered executive rivalry narrative with a concrete, verifiable product announcement (Palantir-Nvidia Sovereign AI Operating System on Nemotron open models, confirmed via Nvidia's own post) that materially affects enterprise build-vs-buy decisions for regulated AI deployments. Scored just above 'solid' rather than 'major' because the core news is a competitive sales pitch and executive commentary, not a frontier model or infrastructure breakthrough itself.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
