Palantir and Nvidia gave the US government something no rival power can easily match: frontier-grade AI that runs entirely inside sealed, air-gapped systems, with Washington owning the chips, the models and the weights end to end. Two days later, Palantir posted a nine-point manifesto that turned the engineering into ideology, warning that any institution depending on someone else’s AI is quietly handing over its future. Read the announcement through the manifesto and the message to the rest of the world sharpens into something colder. The people who just built America a self-sufficient AI machine are also telling Berlin, Delhi and Riyadh exactly what they lose by running on somebody else’s, and right now, that somebody else is the United States.
What Nvidia and Palantir have actually shipped
The engine pairs Nvidia’s open Nemotron models with Palantir’s Sovereign AI Operating System, built on AIP, Ontology, Foundry and Apollo. Agencies run customized Nemotron models on their own hardware, train them on their own data, and keep the resulting weights, the layer that encodes how an organization actually works. It runs on Nvidia’s Blackwell Ultra GPUs. Nvidia wraps the whole thing in a US-heritage story, drawing a line from DARPA’s 1969 university network through UNIX, Linux and Docker to open models today. The customer here is enormous: the federal government employs around 3 million civilians across commerce, energy, healthcare and transportation.
Why one word is doing all the work
Palantir’s July 1 post is where the sales pitch turns into worldview. Surrender your AI sovereignty, it argues, and you hand your institution’s future to people who will use it against you. It mocks “tokenmaxxing,” the chase for high token usage, as an addictive sense of false progress, an unsubtle dig at closed providers who charge by the token. Its cleanest line: controlling your weights is controlling your fate. Weights, in Palantir’s telling, are distilled institutional knowledge, and letting someone else hold them lets them migrate your advantage to their books. It’s a compelling argument. It’s also one that works in reverse the moment you’re not American.
The dependency nobody wants to name
That reversal is the whole story abroad. Every reason Palantir gives Washington to build its own stack is a reason for Berlin, Delhi or Riyadh to worry about buying an American one. The architecture is genuinely sovereign for its owner, and that owner runs on US chips, US models and a US company’s operating system. Nations that plug in inherit a dependency they can’t quietly reverse later. Palantir even warns that “techno-politicization” erodes agency “especially on the battlefield in the West,” language that assumes a hierarchy of who’s inside the tent. The fear isn’t a literal switch someone flips. It’s structural: the knowledge, the tuning, the weights all accumulate wherever the full stack lives, and this stack lives in America.Nvidia’s Justin Boitano says the partnership lets industries and nations turn data into intelligence with speed, efficiency and trust. Palantir insists data authorization, enforced isolation and auditability stay with the customer. Both claims hold up. Neither settles the larger issue. Palantir’s own tagline is that the future of AI is on-prem, and the on-prem it’s selling the world was designed in the US, built on US silicon, wrapped in US software. For everyone outside that arrangement, adopting it means accepting a quiet truth: sovereignty, sold this hard, tends to belong to whoever is doing the selling.
