Palantir and NVIDIA have announced a system that allows US government agencies to run capable AI entirely on their own hardware, with no connection to outside networks. Agencies can train the AI on their own data, customise it for their specific tasks, and own the results outright. Nothing leaves the building.
The AI being used here is called Nemotron, a family of models that NVIDIA makes openly available. Open models are ones where the underlying structure and weights, the core of what makes an AI work, are published for anyone to inspect and modify. This is the opposite of services like ChatGPT or Claude, where you send your data to a provider's server and receive a response back. With an open model, you download it, run it yourself, and your data never touches anyone else's infrastructure.
For most businesses, that distinction has felt theoretical. But it is becoming very practical. Palantir already runs inside the CIA, Pentagon, IRS, CDC, and Army, among others. Its federal contracts nearly doubled in 2025, reaching close to $970 million in a single year. Multi-year contract ceilings awarded in 2025 alone total over $13.7 billion. Palantir is not pitching a concept; it is deeply embedded in the way the US government already operates.
The new Palantir-NVIDIA system sits on top of that existing relationship. Agencies get NVIDIA's Nemotron models running on NVIDIA hardware, managed through Palantir's own software stack, which handles things like who is allowed to access what data and keeps a full audit trail of every action. As the AI is used, agencies can keep feeding it new data to improve it, all within their own walls. The model gets better over time without ever leaving the secure environment.
The broader signal here is not about government technology. It is about where enterprise AI is heading for any organisation that handles genuinely sensitive information. Sovereign AI, meaning AI you own and control rather than rent, has reached top priority on the Gartner Hype Cycle for government services. Analysts predict that by 2028, at least 80% of governments globally will deploy AI agents to automate routine decision-making. Relying on foreign or third-party AI supply chains is increasingly viewed as a security risk, not just a preference.
This same logic is spreading into the private sector. Financial institutions are moving from commercial AI APIs to self-hosted open models, driven not by cost but by the fact that trading strategies, merger documents, and client portfolios cannot safely travel to a third-party server. Legal departments are building isolated AI systems specifically so privileged communications never touch external infrastructure. Manufacturers running operations on factory floors, substations, and remote sites need AI that works without a reliable internet connection.
For the non-technical professional, the practical question is straightforward: does your organisation's most sensitive data currently leave your systems when you use AI tools? If the answer is yes, and if that data includes anything that is subject to privacy law, professional privilege, trade secrecy, or regulatory oversight, then the Palantir model is an example of the alternative architecture that is now available. It is not cheap or simple to set up, but the option exists and is becoming more accessible as open models improve.
Palantir has also recently signed a partnership with Accenture Federal Services, naming Accenture as a preferred implementation partner for US federal customers, with a team of 1,000 professionals being trained and certified on Palantir's platform. That kind of institutional build-out suggests the infrastructure is moving from pilot programmes to standard operating procedure across large agencies. The same progression tends to follow from government into regulated private sectors within a few years.