Government AI Security
Palantir and NVIDIA Deploy Nemotron Open AI Models in Secure US Government Agencies
Key Points
Nemotron 3 Ultra features a 550-billion-parameter model as part of the Nemotron 3 family, and NVIDIA announced the Nemotron Coalition on March 16, 2026, a global consortium of AI labs to advance open frontier models.
Customers will own self-improving models specific to their mission, with the intelligent engine collecting telemetry and using it to post-train and align the model to mission-relevant tasks.
Data rights are emphasized, including data portability and the right to erasure, alongside architecturally-enforced customer-specific isolation and full auditability.
The offering integrates Palantir's infrastructure stack—AIP, Ontology, Foundry, and Apollo—alongside NVIDIA Nemotron open models to enable on-premise deployments in sovereign environments.
NVIDIA frames this move within a historical arc of open-source leadership in American AI, citing milestones such as DARPA-driven Internet origins, UNIX, Linux, GitHub, and Docker to highlight the move toward transparent, open frontier AI.
Palantir and NVIDIA formally launched a new "intelligent engine" on June 29, bringing NVIDIA's Nemotron 3 Ultra — a 550-billion-parameter AI model — into air-gapped U.S. government networks that never touch the public internet. The move gives federal agencies a way to run frontier-level AI entirely on their own hardware, with full control over their data. Yahoo Finance reported that Palantir stock jumped 4.6% to $118.09 following the announcement.
The partnership expands a deal first announced on March 12, when the two companies unveiled a "Sovereign AI Operating System Reference Architecture" in Miami. The goal is simple: let agencies use powerful AI without sending sensitive data to outside servers. Palantir CEO Alex Karp said the system removes "the underlying security risks around proprietary insights migrating into the weights of closed models," according to Investing.com.
Nemotron 3 Ultra Runs Entirely Inside Secure Walls
Nemotron 3 Ultra uses a hybrid architecture with 550 billion total parameters but only 55 billion active at any one time. That design gives it 5x to 9x higher throughput than older models, according to TechBuzz AI. It also supports a context window of up to 1 million tokens — meaning it can process enormous documents in one pass. The model runs on NVIDIA's Blackwell Ultra GPUs using 4-bit precision, so fewer chips are needed.
The "air-gapped" setup means the network is physically cut off from the internet. No data leaves the building. Palantir's AIP and Apollo software collect local usage data and use it to retrain the model on agency-specific tasks — a process NVIDIA calls Multi-Teacher On-Policy Distillation. Agencies own the improved model. The data never leaves their walls, Crypto Briefing reported.
Open Weights Beat Closed Models for National Security
NVIDIA CEO Jensen Huang called open-source AI "foundational to national security, public safety, and U.S. technology leadership." The argument is that agencies can inspect every layer of an open model. With a closed, proprietary model — like those from OpenAI or Anthropic — the government cannot see what is inside. That creates a "black box" risk that Palantir and NVIDIA say they are eliminating.
There is also a cost driver. High prices for proprietary AI API calls — charged per token, every time the model runs — led agencies to seek cheaper on-premise options, according to Investing.com. By owning the model weights outright, agencies avoid long-term vendor lock-in and control their own "tokenomics." The War Department's GenAI.mil platform already serves 1.3 million personnel who have deployed hundreds of thousands of AI agents in just five months.
NVIDIA's Nemotron Coalition Builds a Western AI Bloc
On March 16 at GTC 2026 in San Jose, NVIDIA launched the Nemotron Coalition — a $26 billion, five-year global consortium. Founding members include Mistral AI, Perplexity, LangChain, and Thinking Machines Lab. The goal is to keep Western open models ahead of high-performing Chinese alternatives. Rivals like China's Qwen 3.5 (397 billion parameters) and Kimi K2.6 (1 trillion parameters) have raised alarms in Washington about losing the AI edge.
The Nemotron 3 Ultra weights were released on June 4 under the Linux Foundation's OpenMDW-1.1 license. That open license lets any agency inspect, customize, and redistribute the model. Crypto Briefing noted the architecture is designed for "long-running agents" — AI that can run complex, multi-step tasks like logistics planning or legal discovery, not just answer chat questions.
Deployment Reaches Beyond Defense Into Key Infrastructure
The initial focus is on the War Department and the Intelligence Community. But Yahoo Finance reported that Palantir and NVIDIA are actively marketing the engine to healthcare, transportation, and energy sectors. All of these rely on strict data privacy rules. A hospital cannot send patient records to a cloud server. A power grid operator cannot risk a data breach. The on-premise model fits both use cases.
The U.S. War Department had already signed "Classified Networks AI Agreements" with eight companies — including NVIDIA and Oracle — on May 1 to deploy AI on IL6 and IL7 secure networks, the highest classification tiers. Palantir's Foundry, Ontology, and AIP platforms sit underneath all of it, providing the data governance layer. Palantir's Global Head of Commercial, Ted Mabrey, said the same "Sovereign AI" stack is also expanding into private aviation through a deal with Surf Air Mobility, according to TechBuzz AI.
Sentiment
Mostly PositivePublishers
20
Articles
40
Reach
60
