Sovereign AI Systems: Why Governments Must Build Their Own Models

Governments are shifting from AI experimentation to deployment, but a critical question looms: who controls the models and data that power national decision-making? According to Alan Webber, Program Vice President for National Security, Defense and Intelligence at IDC, the answer increasingly points toward sovereign AI systems—models and datasets developed, hosted, and controlled entirely within national borders.

Why it matters

When governments use commercial AI platforms like ChatGPT or Claude, queries and data flow to U.S. servers, creating dependencies that conflict with national security requirements and data sovereignty laws. For defense, intelligence, and highly regulated sectors like healthcare and financial services, that external reliance is untenable. Sovereign AI represents a fundamental architectural choice: build domestic capability or accept strategic vulnerability.

What sovereign AI actually means

Sovereign AI differs from commercial cloud AI in three dimensions: where the model is developed, where the data resides, and where processing occurs. China's domestic AI ecosystem exemplifies the approach—models built in-country, trained on local data, deployed behind national boundaries. Europe is following suit, with efforts like Italy's Leonardo program creating regional alternatives.

Webber, speaking with Federal News Network's Terry Gerton, emphasized that sovereignty doesn't necessarily mean air-gapped systems. Instead, physical and digital controls ensure data never leaves defined jurisdictions. "Europeans don't want their data coming over to the United States, then answers coming back," he explained. Some U.S. hyperscalers are attempting to offer sovereign boundaries within their platforms, but true sovereignty requires domestic development.

The approach applies beyond government. Healthcare systems handling patient records and financial institutions managing citizen transactions increasingly deploy sovereign architectures to meet regulatory requirements.

The data problem comes first

Webber identified data quality as the primary obstacle governments face. "They don't know what data they have. They don't know the quality of that data," he said. Federal agencies possess vast archives—the U.S. Department of the Interior maintains land records dating to 1876—but lack systematic understanding of what's usable for AI training.

The challenge extends beyond inventory. Agencies must determine which historical records remain relevant, establish access controls within sovereign systems, and verify accuracy before training models. Getting mineral rights or water rights determinations wrong because of poor data curation carries real consequences.

Three more critical barriers

Beyond data, Webber outlined three additional challenges:

Regulatory frameworks remain incomplete. Questions about proper use, access levels, and accountability for AI-generated decisions lack clear answers. A recent executive order directing the Treasury Secretary to review vulnerabilities in advanced AI models represents early foundation work, but comprehensive governance is still under construction.

Skills gaps are severe. While querying ChatGPT resembles a Google search, building repeatable AI capabilities requires specialized training most government employees haven't received. The technology evolves faster than workforce development programs can adapt.

Cybersecurity for AI systems is being invented in real time. "We're literally building the airplane as we're flying," Webber said. The National Institute of Standards and Technology and other agencies are developing standards, but keeping pace with AI advancement remains uncertain.

The speed question

When asked whether governments can maintain pace with AI evolution, Webber acknowledged the difficulty but rejected the possibility of failure. "We can't answer no, we have to figure out a way and it has to be a priority," he said, noting that even tech CEOs are requesting constraints on AI applications in biotechnology due to security risks.

The deployment picture remains uneven. National security and defense organizations lead adoption globally, while civilian agencies show spotty implementation. Webber characterized the overall government position as "really at the beginning" despite pockets of progress.

These details were first reported by Federal News Network in an interview with Alan Webber of IDC.