CREDIT: NVIDIA
NVIDIA’ s response operates on two fronts. The first is efficiency. Since 2016, the company has improved energy efficiency for AI training by 3,000 times and for inference by 45,000 times.
The firm’ s forthcoming Vera Rubin architecture is expected to deliver up to 100 times more performance per watt than its current Blackwell generation, while a full transition to liquid cooling will reduce water consumption by a factor of 300 compared to traditional air-cooled systems. What’ s more, it will have the added benefit that the water used can be 45 ° C, eliminating the need for chillers entirely.
The second front is infrastructure design. Together with Emerald AI and a group of major generators including AES, Constellation, Invenergy, NextEra Energy and Vistra, NVIDIA is developing a new class of AI facilities designed to be flexible in how they consume and store energy.
Research from Duke University suggests that if data centres flex their demand for just a few hours each year, the US could absorb around 100 GW
energydigital. com 29