Logo via Wikimedia Commons; treatment-A cover, license to verify on approval
The two companies are combining open AI models with zero-trust security platforms to give US agencies full control over their data and deployments.
Palantir and Nvidia are deepening their partnership to deliver what both companies are calling a “sovereign AI” stack, purpose-built for US government agencies and critical infrastructure operators that can’t afford to hand their data to third-party model providers. The pitch is straightforward: run powerful AI models on your own hardware, behind your own firewalls, with zero data leakage to the outside world.
The collaboration integrates Nvidia’s open Nemotron models and AI Enterprise software with Palantir’s suite of platforms, including AIP, Foundry, Ontology, Apollo, and the company’s Rubix zero-trust security layer. Together, they form what the companies describe as a Sovereign AI Operating System Reference Architecture.
Why closed models are a dealbreaker for defense #
The core value proposition here is data sovereignty. When a defense agency or critical infrastructure operator uses a closed model from providers like OpenAI or Anthropic, they’re trusting that vendor with potentially classified or operationally sensitive data. That’s a non-starter for many government use cases, particularly those involving classified environments or air-gapped networks that are physically disconnected from the internet.
Palantir CEO Alex Karp has been vocal about this gap, positioning the company’s offerings as a direct answer to the frustrations government clients face with token-based pricing models and the data exposure risks that come with commercial AI platforms. Multiple US critical infrastructure clients are already running Nemotron models through Palantir’s platforms, according to Karp.
Nvidia CEO Jensen Huang has framed the open model approach as essential for national security and maintaining US leadership in artificial intelligence. Open models give customers the ability to train and fine-tune on proprietary data without exposing model weights or operational insights to anyone outside their organization.
The technical architecture #
The reference architecture supports on-premises, edge, and sovereign cloud deployments. On the hardware side, it runs on Nvidia Blackwell Ultra systems with 8 GPUs per node, paired with Spectrum-X networking infrastructure. The software stack includes Nvidia’s CUDA-X libraries and Palantir’s Rubix platform, which enforces zero-trust security principles across the entire deployment.
Key capabilities go beyond just running models in isolation. The system supports explicit authorization protocols, customer isolation, data portability, erasure rights, and full auditability. There’s also a continuous feedback loop built in, allowing models to improve over time based on operational use without sending data outside the secure perimeter.
The Nemotron model catalog gives agencies access to a range of model sizes and configurations, while Nvidia’s cuOpt optimization engine adds supply chain and logistics capabilities on top of the core AI infrastructure.
Supply chain optimization as a proving ground #
The partnership’s expansion on September 10, 2026 added supply chain optimization to the mix, and Nvidia put its money where its mouth is by deploying the system internally first. The company used the sovereign AI stack to manage constraints across its global supplier network, dealing with the complexity of 1.3 million parts per AI rack.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our