Palantir and NVIDIA pair Nemotron open models with air-gapped deployments for U.S. agencies

NVIDIA said Palantir is using NVIDIA Nemotron open models to help U.S. government agencies build and operate secure, custom AI systems in environments that cannot connect to the open internet. The announcement positions the effort as a bridge between “open models” that can be inspected and tuned and “closed environments” where organizations need isolation, auditability, and control over data and model behavior.

The companies’ focus is on deployments that run in air-gapped settings, which are described as completely isolated from unsecured networks. NVIDIA said Palantir will use Nemotron open models within those setups on NVIDIA accelerated computing, allowing agencies to deploy tailored, frontier-quality capabilities without exposing proprietary operational data to less controlled systems.

NVIDIA also framed the partnership around the ability for customers to retain ownership. In the description of the Palantir approach, agencies and operators can run customized Nemotron models on their own infrastructure, train on their own data, and retain full ownership of the resulting models, including the model weights that encode operational knowledge.

A key element in NVIDIA’s account is Palantir’s Sovereign AI Operating System, which NVIDIA says is built on AIP, Ontology, Foundry and Apollo. NVIDIA characterized that operating system as handling the operational and data authorization layer to support deployment in sensitive environments, emphasizing explicit data authorization, architecturally enforced isolation, and full auditability as central design points.

NVIDIA said agencies can improve those customized models over time using new data and feedback that stays within the customer environment. The company described this as a “data flywheel,” designed to continually optimize model performance while keeping data, model weights, and auditability under customer control rather than outside systems.

The companies’ messaging also ties the collaboration to broader U.S. innovation narratives around open source, from early ARPA-connected networking and foundational software systems to modern tooling. NVIDIA argued that open models can be reviewed by independent researchers to identify vulnerabilities, biases, and unintended behaviors that a single organization might miss, and then be refined based on that visibility to support safety through transparency.

On enterprise deployment, NVIDIA said enterprise-grade deployments can be supported through the NVIDIA AI Enterprise software suite alongside Nemotron open models and Palantir’s critical infrastructure products. NVIDIA’s framing suggests the software layer is intended to make it easier for organizations to operationalize the open-model approach in production, while still maintaining constraints around data security and deployment environments.

While NVIDIA did not provide contract size, named agency customers, or specific performance benchmarks in its blog post, it did connect the initiative to many typical government functions, including commerce, energy, healthcare, agriculture, education, and transportation. NVIDIA also noted that the U.S. government workforce includes about 3 million civilian employees, describing government as comparable to one of the world’s largest enterprises in operational complexity.

What remains unclear from the announcement is the exact scope of deployment timelines and the level of model customization details that will be available agency by agency. NVIDIA also did not specify the technical boundaries of inspection or the extent to which “domain-optimized harnesses” are being standardized across agencies, beyond describing that they can help deliver frontier capabilities while customers retain control over proprietary data, model weights, and deployment environments.

For now, the key developments to watch are how agencies operationalize the air-gapped workflow, whether Palantir and NVIDIA provide follow-on material clarifying deployment reference architectures, and how quickly customized models move from installation to production training loops. The companies’ next indicates will likely focus on practical integration, security controls, and evidence that the “data flywheel” can be sustained under real-world operational constraints.