US government agencies are swapping proprietary AI for open-weight models, and Palantir CEO Alex Karp told CNBC's Squawk Box on July 1 that multiple clients have already made the move. The transition, driven by a need for full control over data and compute, marks a shift in how classified environments approach artificial intelligence.
Karp's comments came two days after Palantir and Nvidia formally expanded their collaboration. The integration pairs Nvidia's Nemotron family - open-weight models whose parameters can be freely modified and deployed on local hardware - with Palantir's AIP, Ontology, Foundry, and Apollo platforms. The companies describe the combined system as an "intelligent engine" built to operate inside sovereign and classified networks.
"Many of our US clients are already using these models, including multiple supporting critical US infrastructure," Karp said during the interview. Neither company has disclosed which agencies are involved, how many contracts exist, or the financial terms of the arrangement.
How the partnership actually functions
The June 29 announcement builds on work the two firms began in October 2025, when they outlined a Sovereign AI Operating System Reference Architecture. That blueprint is now moving into operational deployments on classified networks. Agencies can customize and run AI models entirely on their own infrastructure, training and inference on sensitive data never leaves their controlled environment.
This approach sidesteps the token-based pricing that closed-model providers rely on. For defense and intelligence work, sending classified information to a third-party cloud - even one run by an American company - is often unacceptable from both a security and a compliance standpoint.
Why closed models struggle to win government deals
Karp didn't hold back on the critique. Token-based pricing, he argued, simply can't meet the operational requirements of enterprise and government customers. Proprietary models from labs like OpenAI and Anthropic require data to be sent out, processed remotely, and returned. In classified settings, that pipeline introduces risk that most agencies won't accept.
Nvidia's Nemotron models eliminate the dependency. Because they're open-weight, organizations download and run them on local hardware, keeping data in-house. That control extends to the analytics pipeline as well - everything from model fine-tuning to inference stays inside the agency's own infrastructure. The Palantir platform provides the operational layer to make that possible at scale.
What the shift signals for the vendor landscape
For government technology leaders, the move toward open-weight models is not just a procurement preference. It reflects a strategic bet on supply chain independence and data sovereignty. The hardware and software stack becomes something the agency owns and operates, rather than a service it subscribes to.
This trend has implications for the broader AI ecosystem. As more classified workloads shift to open-weight deployments, Nvidia deepens its position as the infrastructure provider behind those deployments. Palantir solidifies its role as the integration layer that connects raw model capability to regulated operating environments.
Closed-model vendors are not sitting still. OpenAI and Anthropic have both been pursuing government contracts. But their architectures introduce friction in environments where data must never leave agency control, a requirement that is unlikely to relax in the near future. This shift in AI for Government procurement signals a long-term challenge for any model provider whose business relies on remote processing.
Why this matters for Government
The operational reality for government agencies is straightforward: data control is non-negotiable in classified and critical infrastructure work. Open-weight models let you deploy AI on your own hardware, with your own data, on your own terms. If your agency is evaluating AI providers, the ability to run models locally - without metered pricing or external data flows - is quickly moving from a desirable feature to a baseline requirement. The vendors that recognize this shift will have a structural advantage in future procurement cycles.
Your membership also unlocks:
