Palantir & NVIDIA launch air-gapped AI for agencies

Palantir has introduced an intelligent engine for U.S. government agencies that uses NVIDIA Nemotron open models in air-gapped environments.

The arrangement brings NVIDIA's open models into isolated computing setups separated from unsecured networks. Agencies will be able to run customised models on their own infrastructure, train them on their own data and retain ownership of the resulting models, including model weights.

The offering centres on Palantir's Sovereign AI Operating System, built on the company's AIP, Ontology, Foundry and Apollo products. The companies said the software stack manages deployment, data authorisation, isolation and auditability for sensitive government environments.

Palantir plans to use NVIDIA Nemotron open models to build custom models for U.S. government work, reflecting demand from agencies operating across commerce, energy, healthcare, agriculture, education and transport.

U.S. government departments and agencies manage a wide range of public services and operational systems, many resembling those of large private-sector organisations. The companies said AI tools can help agencies handle operational complexity in areas such as food safety and interstate highway infrastructure.

Open model push

NVIDIA presented the development as part of a broader U.S. push around open-source and open-model software. It pointed to the country's long history in open-source computing, from early networking work linked to DARPA to projects such as UNIX, C, Linux, GitHub and Docker.

The company argued that open models give government bodies and businesses more ability to inspect, adapt and deploy AI in sensitive settings. It added that open models can help users retain control over proprietary data, deployment environments and model development.

Under the Palantir setup, agencies can continue refining models inside their own environments using new data and feedback from live operations. That keeps the model improvement process within the customer's own systems rather than passing it to an outside provider.

Palantir said its operating system already includes explicit data authorisation, enforced isolation and full auditability. Those features are intended to address concerns that often arise when agencies consider AI systems for classified, regulated or otherwise restricted environments.

Security and control

The emphasis on air-gapped deployment is likely to be central to the offer for national security and other sensitive public-sector uses. In practice, it means the models can operate on infrastructure with no connection to unsecured external networks.

NVIDIA said deployments can also be supported through its AI Enterprise software suite, giving customers a way to manage and run the software around the open models in operational settings.

The companies presented open models as a way to balance transparency, adaptability and cost. NVIDIA said independent review can help researchers identify vulnerabilities, bias and unintended behaviour that a single organisation might miss.

It also said open models can be modified and fine-tuned for specific uses, including in regulated industries where closed models may raise data security or privacy concerns. NVIDIA cited industry research indicating that about two-thirds of companies already use open models and report cost efficiencies from doing so.

The partnership also highlights how major AI suppliers are positioning open-model systems for sovereign and controlled environments, rather than only for public cloud use. That distinction is becoming increasingly important for government buyers seeking modern AI systems without surrendering control over data, infrastructure or model outputs.

NVIDIA said the combination of Nemotron open models with Palantir's AIP, Foundry, Ontology and Apollo products is intended to support secure AI use across government agencies and commercial organisations. It added that running the models on air-gapped NVIDIA-based infrastructure keeps data and models under customer control.