The most valuable question in enterprise AI is no longer which model is smartest. It is who owns the weights and where they run. On June 29, Palantir Technologies and Nvidia gave U.S. government agencies and critical-infrastructure operators an answer, unveiling an engine that plants Nvidia's open-weight Nemotron models directly inside Palantir's Sovereign AI Operating System. Investors responded with force: Palantir stock jumped roughly 9% to about \$127.31 in midday trading on July 1, with cybersecurity partner Palo Alto Networks climbing about 4% in sympathy.
The pitch is deceptively simple. Customers can deploy, customize, and post-train frontier-quality models on their own data, inside air-gapped and classified environments, and walk away owning the resulting model weights outright. Nothing leaves the perimeter. No proprietary insight migrates into a closed vendor's model.
What The Deal Actually Does
The initiative pairs Nvidia's AI platform, compute, and open Nemotron models with the four pillars of Palantir's software stack: AIP, Ontology, Foundry, and Apollo. Together they form what the companies are calling a Sovereign AI Operating System Reference Architecture. Palantir's layer handles data authorization, architecturally enforced isolation, and full auditability, while Nemotron supplies a customizable, continually learning model layer running on Nvidia accelerated hardware.
Crucially, the models run on the customer's own infrastructure. Agencies train on their own data and retain full ownership of the results, including "the weights that encode their operational knowledge," as Nvidia put it in its announcement. As those models are used in production, agencies can keep improving them in place, creating what Nvidia calls a data flywheel that optimizes performance while keeping data, models, and audit trails under customer control. Enterprise-grade deployments are supported through the Nvidia AI Enterprise software suite.
Nvidia framed the move as continuous with a long American tradition of open innovation, from ARPANET and UNIX in 1969 to Linux and Docker decades later. Open models, the company argued, are now essential infrastructure for national security precisely because they can be inspected, adapted, and deployed in sensitive settings.
The Karp Broadside
Palantir CEO Alex Karp used the announcement to escalate a running feud with frontier AI labs. In the official release, he tied the deal squarely to sovereignty concerns.
"Combining Palantir infrastructure with NVIDIA's AI and Nemotron models will allow the U.S. government to unleash the full power of LLMs while removing the underlying security risks and rational concerns around proprietary insights migrating into the weights of closed models," Karp said.
He was far less measured on CNBC the following day, calling the AI industry "effing insane" and accusing leading labs of overcharging customers, exploiting their data, and jeopardizing U.S. national security. His core argument: enterprises are paying heavily for AI tokens while handing over proprietary data and IP with little to show for it, and firms like OpenAI and Anthropic lack sufficient customer-IP protection. Palantir's customers, he contended, want control over their compute, their models, and their data.
Nvidia CEO Jensen Huang has made the strategic stakes explicit, casting open-source AI as foundational to national security and continued U.S. technology leadership, a message Nvidia reinforced in its own blog post accompanying the launch.
Why It Matters
For years the AI infrastructure race has been narrated as a contest of raw model capability, waged in hyperscaler data centers reachable only through an API. This deal advances a rival thesis: that for the highest-stakes buyers, control beats cleverness. The U.S. government, with roughly 3 million civilian employees spanning commerce, energy, healthcare, and transportation, is effectively one of the world's largest enterprises, and it cannot pipe classified data to a hosted cloud model.
The prize is enormous. McKinsey estimates the sovereign AI market could grow into a \$600 billion opportunity by 2030, driven by governments and regulated enterprises that structurally cannot use hosted services. By fusing Nvidia's silicon and open weights with Palantir's authorization-and-audit layer, the pair is positioning itself as the default on-ramp for that spend, and reframing the competitive battleground away from the closed frontier labs Karp is publicly attacking.
There is a deeper shift embedded here too. Open-weight models, once dismissed as trailing closed frontier systems, are being recast as a strategic advantage rather than a compromise. When a model's weights can be owned, inspected, and refined behind an agency's own firewall, transparency becomes a selling point instead of a liability, and the token-metered economics of the closed labs start to look like a tax buyers can route around.
What To Watch
The announcement is a reference architecture and an engine, not yet a portfolio of disclosed contracts. The signal to watch is conversion: named agency deployments, dollar figures, and evidence that air-gapped Nemotron models perform close enough to closed frontier systems to justify the trade. Also worth tracking is whether the closed labs respond with their own sovereign or on-premise offerings, and whether Palantir can extend the same playbook beyond Washington into regulated commercial sectors such as finance and healthcare. For now, a 9% one-day pop shows the market believes the sovereignty thesis has teeth. The contracts will decide whether it has revenue.
"Combining Palantir infrastructure with NVIDIA's AI and Nemotron models will allow the U.S. government to unleash the full power of LLMs while removing the underlying security risks."— Alex Karp, CEO, Palantir