Sovereign AI Gives Back Your Data, Not the Machinery: Who Captures More of Palantir and Nvidia's Economics

On September 10, the two companies that have become shorthand for the AI boom announced they'd sell a packaged "sovereign AI" stack to governments and big corporations — the tools to run AI on your own premises, on your own data, rather than handing your information to a public cloud. The first customer is Nvidia itself. It plans to use the software to track the roughly 1.3 million parts that go into a single one of its newest server racks, following components across suppliers to see which shortages are holding up builds.

That last detail matters. If Nvidia — which runs what may be the most valuable, intricate supply chain on the planet, and which makes the chips everyone else is waiting on — decided its own crown-jewel logistics deserved this tool, the collaboration is a real deployment, not a press release.

What "sovereign" actually means

The pitch is easy to grasp: nations, utilities, and defense agencies don't want to send their most sensitive data through an American hyperscaler's cloud. "Sovereign AI" promises they don't have to. Palantir supplies the software layer — its Foundry data platform, its AIP operating system, and its "ontology," the model that ties a company's scattered data into a coherent picture of how its business actually works. Nvidia supplies the reasoning models (the open Nemotron family), the cuOpt optimization engine, and the hardware underneath. The whole thing is packaged as Palantir's "Sovereign AI OS," running on Dell, Cisco, Rackspace, or Nebius infrastructure, on-premises or in a colocation center.

Read the fine print of what the customer actually controls and the name overpromises. The buyer keeps its data, but it still runs on Nvidia's accelerators, reasons through Nvidia's models, and organizes its world through Palantir's ontology. As one observer put it, the deal sells sovereignty over the contents and dependency on the machinery. The customer stays dependent on two U.S. suppliers — and in an era of export controls and economic nationalism, that dependency is itself the product's whole commercial appeal.

That softens the apparent answer to "who captures more of the economics." Neither company is giving its moat away. But their moats are different in kind, and so is how heavily each stock leans on this.

Two different kinds of moat

Nvidia's edge is physical scarcity. Its accelerators are the most concentrated, hardest-to-substitute node in AI capital spending today, and "sovereign" actually helps it: instead of renting GPU time from the cloud, each customer buys its own fleet in the country that wants the capability. Every sovereign deal moves the same chips, just to a different destination.

Palantir's edge is lock-in by inertia. An ontology is embedded deeply into how a customer represents its business, and it is painful to rip out once workflows are built on it. Nvidia choosing Palantir internally matters precisely because that same stickiness will be hard for a customer to reverse later. Palantir gets a recurring, near-pure-margin subscription on top of a hardware purchase it doesn't have to fund.

Now the economics. A chip sale is far bigger in dollars than any software license; on a per-deal basis Nvidia pockets the large majority. But proportionally, the two are levered oppositely.

Nvidia is a ~$300 billion revenue business that grew about 83% over the trailing year, at 74% gross margins, 64% operating margins, and over 86% return on invested capital. One sovereign contract, even a big one, is a rounding error on that base. For Nvidia, sovereign AI is one channel of many; the deals mostly determine where its chips land, not how many ship.

For Palantir, sovereign AI is closer to the whole game. Trailing revenue is roughly $6.2 billion — about 1/50th of Nvidia's — growing around 79%. Its gross margin is near 85%, operating margin near 43%, and free-cash-flow margin above 50%, with almost no capital spending: software that embeds into a customer is close to pure margin once built. Management just raised full-year 2026 guidance to about $8.15 billion, up 82%, and U.S. commercial revenue — the segment most exposed to this enterprise-AI-sovereignty story — jumped 149% year over year in the second quarter to $764 million, with the full year guided to more than $3.4 billion, up at least 134%. When Palantir sells a sovereign stack, the growth is highly incremental to a small base.

The market has already picked a winner

Here is where the structure stops doing the work and price takes over. The market has already decided that Palantir captures more economics — it paid for that belief. Palantir trades around 65 times sales and 133 times trailing earnings. Nvidia trades around 17 times sales and 27 times earnings. For every dollar of revenue, investors are paying roughly four times more for Palantir than for Nvidia, on the theory that software lock-in is stickier and higher-margin than hardware that faces growing competition from custom chips and hyperscaler in-house designs.

That premium is defensible on margins. It is also a bet that the ontology lock-in outlasts the hardware advantage, and it removes the easy version of the trade: the structural relationship between the two companies is real, but Palantir's stock no longer offers an unappreciated version of it.

The irony in the pitch

Now the falsifiable part, and it is a genuinely awkward one for Palantir specifically. Sovereign AI is aimed hardest at governments — and the governments most animated by the word "sovereign" are exactly the ones that do not want to depend on an American vendor.

Europe is ground zero for the pitch and for the pushback. Germany ruled Palantir out of military decision support in April, is reviewing roughly 30 mostly German alternatives, and wants a prototype by 2027. France's intelligence service is replacing Palantir with a homegrown rival. In July, France and Germany signed a joint declaration to build a European alternative to Palantir's software, modeled on France's "Arcadia," and both have already dropped Palantir for the French firm ChapsVision. The EU is drafting "Made in EU" procurement rules.

Nvidia's chips are far less politically branded — a European sovereign AI data center still depends on Nvidia silicon. But Palantir is the named target of Europe's de-Americanization effort. The trend that should lift Palantir most, governments buying national AI stacks, is the same trend that produces buyers determined to avoid Palantir in the market where the pitch matters most. That is not a killer for the thesis — the U.S. and allied-government market is large, and Palantir reports live deployments at eight sovereign entities including Israel, Switzerland, the UN World Food Programme, and NATO-aligned groups. It is a real boundary on how clean the export of this relationship is.

So the useful takeaway is not a clean winner. Per dollar of deal, Nvidia captures more; per dollar of company, Palantir is far more levered; and the market has already priced Palantir's software lock-in as the scarcer, more valuable asset in this pair. The discipline is to remember that structure and price are different questions. Nvidia's scarcity is physical, durable, and comparatively cheap. Palantir's lock-in is real but contested — Europe is actively trying to design around it — and it is expensively priced. Both are building the same stack. Only one of them is pricing in flawless execution.