Palantir has unveiled a new platform that runs secure AI for U.S. government agencies using NVIDIA's open models, NVIDIA Nemotron[1]. It brings frontier-grade models into "air-gapped" environments that are fully isolated from outside networks, letting each agency run, train, and own the models on its own infrastructure. The idea is deliberately paradoxical: open models prove most valuable inside closed environments.

Bringing open models into classified environments

What Palantir introduced is a way to build agency-specific models on top of NVIDIA Nemotron open models[1]. The models run on NVIDIA accelerated computing inside air-gapped environments, completely cut off from unsecured external networks[1].

Government operations overlap heavily with private-sector business across commerce, energy, healthcare, agriculture, education, and transportation[1]. With roughly 3 million civilian employees, the U.S. federal government is, in NVIDIA's words, one of the world's largest enterprises[1]. The aim is to use AI to streamline this sprawling complexity and support frontline decisions.

NVIDIA Nemotron is a family of open models that publishes its weights (a model's parameters), training data, and training techniques, designed for autonomous AI agents[2]. The latest Nemotron 3 comes in three variants, Nano, Super, and Ultra, to match different use cases and workloads[2]. Because the open weights can be reviewed before deployment, organizations can bring them into sensitive environments and fine-tune them on their own data, which fits this use case well.

A "sovereign AI" OS that holds the data and ownership

The foundation for operating these models securely is Palantir's Sovereign AI Operating System[1]. Built on the company's AIP, Ontology, Foundry, and Apollo, it handles the operational and data-authorization layer needed for deployment in sensitive environments[1]. Explicit data authorization, architecturally enforced isolation, and full auditability sit at its core[1].

Each agency can run a customized Nemotron model on its own infrastructure, train it on its own data, and fully own the resulting model, including the weights[1]. Since those weights encode an organization's operational knowledge, keeping them in-house rather than handing them to a third party carries real weight. Continually refining the model with fresh data and feedback creates a data flywheel that improves performance while keeping data, models, and audit logs under the customer's control[1]. For enterprise-grade production, the NVIDIA AI Enterprise software suite provides additional support[1].

Open models for trust, control, and lower cost

NVIDIA frames the value of open models around three points: trust, control, and cost[1].

Trust comes from transparency. Because third parties can independently review the models, vulnerabilities, biases, and unintended behaviors that a single organization might miss are easier to find and fix[1]. Control means companies, governments, and developers can modify and fine-tune models for their own use cases, deploying them even in regulated fields such as finance where closed models might run afoul of data protection rules[1]. On cost, NVIDIA notes that about two-thirds of companies already use open models and value their cost efficiency[1]. Cost efficiency has become an important factor in building AI that can scale without breaking.

Open source has long been a pillar of U.S. technology leadership, running through ARPANET, UNIX, C, Linux, GitHub, and Docker[1]. This effort extends that lineage of open models into one of the most sensitive arenas of all, national security.

Summary

Palantir and NVIDIA have laid out a platform that brings Nemotron open models into air-gapped environments, letting U.S. agencies run, train, and fully own models, weights included, on their own infrastructure. Palantir's sovereign AI OS handles data authorization, isolation, and auditing, while the pitch centers on trust through transparency, control over use cases, and the lower cost of open models. That said, this is the vision Palantir and NVIDIA are painting, and how well it can balance security with results will depend on real-world deployments to come.

出典:https://blogs.nvidia.com/blog/palantir-secure-ai-us-agencies-nemotron-open-models/

出典:https://www.nvidia.com/en-us/ai-data-science/foundation-models/nemotron/