Sovereign AI isn't optional — it's survival, says NUS AI Institute director

Countries that outsource their foundational AI models risk permanent technological dependence and "digital colonisation," according to Professor Mohan Kankanhalli, Director of the NUS AI Institute and Deputy Executive Chairman of AI Singapore. Speaking at the ATxSummit in Singapore, he argued that true AI sovereignty goes far beyond data residency laws — it must extend to the model layer itself.

Kankanhalli defined sovereign AI as the ability to own and govern the full AI value chain: generating and safeguarding data, building models, deploying them, and continuously improving them while retaining the economic and societal upside. He warned that nations relying exclusively on foreign models create a self-reinforcing cycle where their citizens' data trains overseas systems while domestic ecosystems fall further behind. "If a foreign corporation or government suddenly decides to restrict access to their model, as occurred when Meta altered its Llama strategy, local enterprises that built on that tech are left stranded," he said.

The professor outlined three pathways for nations navigating sovereign AI: buying a model (fast but creates long-term dependency), building from scratch (full control but requires sustained investment), or a hybrid "blend" approach that starts with buying and systematically transitions to building. He pointed to Singapore's own approach as an example — through AI Singapore, the country is developing the Sea-Lion family of Southeast Asian language models while also investing in specialised medical AI trained on local healthcare data. Smaller, mid-sized models tailored to specific sectors can deliver significant value without requiring frontier-scale investment.

Kankanhalli also framed the debate in national security terms. As frontier AI models become increasingly capable of discovering software vulnerabilities and mounting cyberattacks, countries without comparable defensive capabilities will find themselves at a strategic disadvantage. "Countries which possess these frontier models can attack the cyber infrastructure of other countries not having such a model," he cautioned. To defend against such models, nations need equally powerful ones — either built domestically or acquired through deep trust-based partnerships.

Why it matters for Singapore: This isn't abstract theory — Kankanhalli's argument directly validates the approach Singapore has been taking through AI Singapore, the Sea-Lion models, and the NAIS framework. The tension between building versus buying is the central strategic question for every mid-sized economy navigating AI today. Singapore's bet on a hybrid model — targeted domestic models where it has data advantages, plus strategic access to frontier systems from trusted partners — may well be the pragmatic blueprint for other nations facing the same choice.