Temasek is betting on next-generation chips and AI model architectures, predicting disruptive changes in AI energy demand

The Zhitong Finance App learned that Singaporean state-owned investment company Temasek Holdings Pte. (Temasek Holdings Pte.) pointed out that continued progress in the field of artificial intelligence (AI) is expected to significantly change the energy demand pattern of the industry — and currently, this industry is putting increasing pressure on the global electricity network.

Russell Tham, head of Temasek's Global Investment Emerging Technology Division, said that with the application of more efficient AI architectures, material science innovation and continuous optimization of chip manufacturing processes, the overall impact of the AI industry on the environment is expected to be drastically reduced.

“In the future, the energy structure required to generate AI tokens could change dramatically,” Tham said at the Bloomberg Sustainable Business Summit in Singapore on Tuesday. He further stated that the existing AI architecture “is still fundamentally inefficient, despite being amazing and growing at an extremely rapid pace.”

Temasek said earlier this year that it may be difficult to meet its goal of halving the carbon emissions of its investment portfolio from 2010 levels by 2030, in part because the AI industry's demand for energy continues to rise.

Tight grid loads and rising electricity prices are prompting data center operators to actively explore new power supply solutions. According to forecasts, according to current trends, the electricity consumption of US data centers will account for about 20% of the country's total electricity consumption by 2035, far higher than the current 5.9%.

Companies such as Cuspai, a British startup funded by investors such as Temasek and Bezos Expeditions, are working to improve semiconductor production processes to reduce or even replace the use of some rare metals.

Tham revealed, “We have invested in highly innovative materials and chip architectures that are significantly more energy efficient than existing technology. At the same time, we have also laid out some new AI model architectures that have not been fully verified but are very promising. These architectures are also expected to achieve a significant leap in energy efficiency.”