Temasek Sees AI Breakthroughs Slashing Energy Demands as Data Centre Power Needs Surge
Source: The Business Times
Temasek believes AI's energy equation could shift dramatically as more efficient architectures, better chipmaking, and materials discovery breakthroughs arrive. Speaking at the Bloomberg Sustainable Business Summit, Temasek's Russell Tham called current AI architecture 'fundamentally ineffective' while data centre energy demands continue to rise.
The relationship between AI and energy has become a central tension in the technology's rapid expansion — data centres need more power than ever, but the AI models running inside them are, by Temasek's own assessment, "fundamentally ineffective" in their current architecture. Speaking at the Bloomberg Sustainable Business Summit in Singapore on July 21, Russell Tham, head of emerging technologies at Temasek Global Investments, offered a cautiously optimistic view: the energy equation for AI "may change quite drastically" as more efficient architectures, advanced chipmaking, and materials discovery breakthroughs arrive.
Temasek has direct exposure to this tension. The state-owned investor said earlier this year that it is unlikely to meet its goal of halving the carbon emissions attributed to its portfolio from 2010 levels by 2030 — partly because of AI's rising energy demand. Data centres in the US alone are projected to consume about 20% of the country's electricity by 2035, up from 5.9% today, according to BloombergNEF. For a portfolio that includes significant data centre holdings across Asia, these numbers are not abstractions.
The potential solutions Tham pointed to span multiple fronts. More efficient AI architectures — including alternatives to the dominant transformer-based models — could deliver the same or better results with significantly less compute. Materials innovation, including work by Temasek-backed CuspAI (also backed by Bezos Expeditions), aims to improve semiconductor production by reducing or eliminating the use of rare metals. Better chipmaking techniques, meanwhile, could deliver the kind of node-on-node efficiency gains that have historically driven computing's energy curve downward.
Grid constraints and ballooning electricity prices are already prompting data centre operators across Asia to explore new ways to power their facilities, from small modular nuclear reactors to on-site solar. But Tham's remarks suggest Temasek is betting on the supply-side solution as well — investing in the technologies that make AI itself less energy-hungry, rather than just finding cheaper power to feed the current architecture.
Why it matters for Singapore: Singapore has positioned itself as a data centre hub for Southeast Asia, but land and energy constraints mean it cannot simply build its way out of AI's power demands. Temasek's thesis — that AI's energy problem will be solved by better AI, not just more power plants — aligns with Singapore's broader strategy of investing in R&D and deep-tech innovation. If Temasek's bet pays off, Singapore could emerge as a hub not just for AI consumption, but for the efficiency breakthroughs that make AI sustainable.
