Palantir and Nvidia Team Up to Let Governments Run Their Own AI Independently

For years, building with artificial intelligence meant integrating someone else's model via an API. But for government agencies and operators managing critical infrastructure, this model breaks down. When data cannot legally or operationally leave a secured network—whether it involves intelligence work, power grid management, or patient records—offloading processing to a cloud vendor's endpoint is not feasible.

That's the backdrop for Palantir's recent move. The company unveiled an intelligent engine built on Nvidia technology and Nemotron open models that allows organizations to run, customize, and continuously refine AI systems within air-gapped networks while maintaining complete control over both data and model weights. The key difference: Palantir shipped not a new model but the infrastructure to deploy and own one.

The Shift From APIs to Internal Platforms#

Most organizations will likely mix hosted and self-managed models for the foreseeable future, yet a distinct pattern is crystallizing. Rather than treating models as external services, applications increasingly communicate with an internal AI platform that directs requests to models running on the organization's own GPUs. This approach ensures data stays within the organization's boundaries, giving security and compliance teams direct oversight of governance, auditing, and data retention.

The underlying question is shifting from "which model should we call?" to "which models should we own and operate?" and perhaps more crucially, "what's the operational cost?"

What Nemotron Brings to the Table#

Nemotron represents Nvidia's family of openly available models released in 2026 across three scales—Nano (~31.6B parameters), Super (120B), and Ultra (550B). Each uses a hybrid design combining Mamba and Transformer technology with a mixture-of-experts approach that activates roughly one-tenth of parameters per token, making them substantially cheaper to run than their parameter counts suggest. The models support context windows reaching a million tokens.

Nvidia has published model weights, training data, and recipes under a permissive open license, allowing deployment through various open runtimes including vLLM, SGLang, Ollama, and llama.cpp, as well as Nvidia NIM containerized microservices optimized with TensorRT-LLM. Palantir had already appeared on Nvidia's list of early Nemotron adopters, so this announcement formalizes an existing relationship rather than marking a new departure.