A consistent thread through the discussion was that compute, while critical, is not sufficient on its own.

“Ownership of compute alone does not guarantee meaningful impact or transformation,” noted Vilas Dhar, cautioning against infrastructure-first approaches. He also compared dominant narratives of “AI diffusion” to trickle-down economics, a passive model that concentrates capacity and assumes benefits will follow.

Martin Tisné highlighted the risks of over-indexing on infrastructure, warning that data centres could become “white elephants” without contextual data, local languages, and a well-resourced open-source ecosystem. As he emphasised, “For countries to be able to exercise sovereignty, they need to have contextual AI, and contextual data in their languages with all the diversity.”

Shaun Seow pointed to the skills gap across Asia as a potentially greater constraint than hardware, and underscored the need for more pragmatic collaboration: “Collaboration should prioritise access to compute rather than ownership of it.”

Dr. Shikoh Gitau reinforced the importance of ecosystem readiness: “Without talent, capability, and data, investment in compute infrastructure risks going to waste.”

The panel also challenged prevailing notions of sovereignty. Dhar argued that locating compute within national borders does not remove the interdependencies of the 21st century, calling instead for mutual value exchange. Tisné drew on the example of indigenous data sovereignty in New Zealand’s Māori context to suggest a shift from territorial control toward relational agency.

On making demand more concrete, Dr. Gitau shared that Africa requires approximately 2.5 million GPU hours annually but currently has access to only a small fraction of that capacity. She stressed that investments must be tied to clear, quantified needs and real-world use cases, and that South-South collaboration works best when anchored in such specificity. Seow proposed aggregating demand to negotiate with cloud providers, combined with philanthropic subsidy, as a more realistic pathway than shared infrastructure ownership.

Sushant Kumar reflected on the broader intent of the work: “Opening up computational resources, or more broadly all resources necessary for AI development, in the public interest and for real-world impact is crucial. This research helped us think through and engage more deeply with the Democratising AI Resources Working Group.”

The discussion underscored the need to move beyond infrastructure-centric approaches towards holistic, demand-driven, and collaborative models for building public-interest AI.