Sovereign AI is a tagline used to describe efforts by different nations to establish their own AI supercomputers, entirely within their countries, to allow them to develop their own artificial intelligence capabilities without depending on foreign (US-based) cloud service providers.
Here are a few key sovereign AI efforts:
Notably, there is no US sovereign AI effort because all AI effort is US-sovereign by default. However, Genesis is positioning itself as US-sovereign in the sense that its infrastructure is not owned by commercial interests like cloud providers.
Precise definition
The Canadian sovereign AI effort defines “sovereign” in loose terms:1
a Canadian-located, Canadian-governed system that ensures data residency, operational control, and decision-making authority and agency remain in Canada
Specifically,
- Data should prioritize residency in Canada - but not all data must reside in Canada.
- Compute and storage must be owned or leased from Canadian entities.
- The physical design and access control must rest in the hands of a Canadian entity
- Components should “maximize” Canadian integration, installation, and support. Foreign-made components are OK.
HPC dressed as AI, or real AI?
I’ve noticed that the biggest sovereign AI efforts are being run much more like national HPC efforts than true AI infrastructure projects. They are commonly attached to traditional HPC organizations (universities) and have procured AI systems from traditional HPC vendors (Cray, Eviden) building traditional HPC system platforms (Cray EX, BullSequana XH3000) rather than systems optimal for AI (see HPC vs AI). This leads me to believe that “sovereign AI” is really masquerading as “HPC that can technically be used for AI as well.”
Andrew Jones presented this highly relevant slide at the HPC-AI Advisory Council UK Conference in October 2026:2

Drivers
Cohere found that Qwen3.8-Flash-Next was intermittently probing the system on which it was running, then uploading the results to a bucket hosted in Beijing.3 Their blog post had this great set of lines:
CrowdStrike saw DeepSeek-R1’s rate of severe vulnerabilities climb when prompts mentioned Tibet. Booz Allen found three of four Chinese coding models wrote more vulnerable code for a government customer, with Qwen3-Coder roughly 130 percent worse. Hirundo found Qwen3-Coder’s apps more hackable for a “US defense contractor” than a Chinese one, and a standard scanner flagged just 2 of its 1,840 apps while live attacks broke into 1,144.
These sound dubious, but I haven’t scrutinized the statistics to say one way or another. However, the fact that models can be trained to behave this way is a good reason to not trust models trained by other countries.