On June 29, 2026, Palantir introduced what NVIDIA’s announcement calls an “intelligent engine” for U.S. government agencies that runs NVIDIA Nemotron open models in air-gapped environments — networks fully isolated from unsecured systems. Agencies and operators can run customised Nemotron models on their own infrastructure, train them on their own data, and keep full ownership of the resulting models, including the weights.
The engine sits within Palantir’s Sovereign AI Operating System, built on its AIP, Ontology, Foundry and Apollo products, which handles deployment, data authorisation and isolation. Palantir also plans to use Nemotron to build custom models for government work.
/ 01What was announced
According to NVIDIA’s blog post and SecurityBrief’s report:
- Open models, closed networks. Nemotron models can be deployed in classified and air-gapped settings on NVIDIA accelerated computing. NVIDIA said running models on air-gapped infrastructure “keeps data and models under customer control.”
- Ownership of weights. Agencies retain the customised models and the weights that encode their operational knowledge.
- Continuous improvement in place. Models can be refined with new operational data and feedback inside the agency’s own environment, which NVIDIA describes as a data flywheel.
- Target users. No specific agencies were named. The announcement cites demand across commerce, energy, healthcare, agriculture, education and transportation.
- Support. Enterprise deployments can be supported through the NVIDIA AI Enterprise software suite.
The companies did not publish benchmarks, deployment numbers or pricing.
/ 02Why it matters
The announcement reflects a shift in how governments think about AI risk. Two months ago the Pentagon signed agreements with eight vendors to bring frontier AI to its classified networks, and its CTO argued publicly against relying on any single provider. Open-weight models take that logic further: if an agency holds the weights, a vendor dispute, pricing change or policy decision cannot switch off its capability.
Air-gapped deployment also addresses a practical constraint. Much government data cannot leave the networks it lives on, and many sensitive workloads have no route to a commercial API at all. Models that run entirely inside the perimeter are the only option for those workloads, and open models make that easier to arrange.
There are trade-offs. Open models may trail the best proprietary systems on some tasks, and a model that is continually retrained on operational data needs disciplined evaluation, versioning and rollback. An agency that owns its weights also owns the job of knowing what changed between versions and why the model now behaves differently.
/ 03Weights are not the whole of the knowledge
The phrase “weights that encode operational knowledge” deserves a second look. Fine-tuning does put some institutional knowledge into a model, but it puts it there in a form nobody can read, cite or correct line by line. If a fine-tuned model has absorbed an outdated assessment or a misattributed report, the only remedy is to retrain and hope.
For high-stakes work, the knowledge an agency depends on — entities, relationships, sources, assessments and the rules that govern them — is better held in an explicit structure alongside the model. A knowledge graph can be inspected, corrected and versioned. Each fact can carry its source and its date. A model reasoning over that graph can cite the facts it used, and when a fact changes, the change takes effect immediately without retraining.
Palantir’s own platform includes an ontology layer, which suggests the industry broadly agrees that models need structured context. The open question for buyers is how portable that context is. Knowledge held in open standards such as RDF and OWL, aligned to a widely used upper ontology, can move between platforms. Knowledge held only in one vendor’s proprietary format can create the same lock-in that open weights were meant to avoid.
Owning your weights is a good start. Owning a readable record of what your AI knows is the rest of it.
/ 04What agencies should ask
- What is owned, and in what format? Weights, training data, evaluation sets, and the structured knowledge the system relies on should all be exportable.
- How are model updates evaluated? A data flywheel needs a regression suite and a clear record of what each version was trained on.
- Can answers be traced? For analysis and decision support, each output should point to the sources and facts behind it, not only to the model version that produced it.
Open models in closed networks are a welcome development. They work best when the knowledge they reason over is as open to inspection as the weights are to ownership.
Owning the weights protects an agency from a vendor. It does not let anyone audit what the weights believe, and a data flywheel that keeps retraining them makes that audit harder with every turn.
Keep the operational facts outside the weights. An agency’s entities, assessments and rules sit in a BFO-aligned graph inside the air gap, each fact carrying its source, its date and the analyst who approved it. When an assessment is withdrawn, the node is superseded and every answer that depended on it can be found by following the graph; no retraining is needed. The model, open or proprietary, reasons over the graph and cites the nodes it used, so a new model version can be tested by replaying a fixed question set and comparing the evidence chains, not only the answers. Because the ontology exports as OWL or RDF, the knowledge outlives any single platform, which is the same independence the open weights were meant to secure.
How we’d model it
Assessment—supersededBy→AssessmentFact—hasSource→Source documentAnswer—dependsOn→FactModel version—trainedOn→Dataset
- Open Models, Closed Environments: Palantir Brings Secure AI to US Agencies With NVIDIA Nemotron — NVIDIA Blog blogs.nvidia.com ↗
- Palantir & NVIDIA launch air-gapped AI for agencies — SecurityBrief US securitybrief.news ↗
See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.