While gas turbines and transformers have drawn attention in the AI arms race, a quieter boom is under way in battery storage. In the first five months of 2026, global shipments of storage for AI data centers reached 10GWh, already exceeding all of 2025. Industry forecasts put annual shipments of AI data center lithium batteries above 300GWh by 2030, and most new orders are going to Chinese manufacturers.
The reason lies in how AI computing consumes power. A single Nvidia NVL72 rack can draw 120 to 140 kilowatts, about 20 times a traditional rack. More troublesome is that thousands of GPUs synchronize during training, alternating between full load and idling, which creates square-wave power swings of tens of megawatts. Grids respond in seconds, while GPU power shifts in milliseconds, and even a 10-millisecond voltage dip can ruin a training run. Traditional lead-acid backup systems are too slow and too far from the chips.
The solution is a three-layer defense: capacitors handle jitter, in-rack battery backup units respond within 2 milliseconds, and campus-scale storage absorbs fluctuations lasting seconds to minutes. Demand is also driven by grid connection queues of three to seven years in some US regions, pushing operators to build their own power supply. The global BBU market, about 3.5 billion yuan ($490 million) in 2025, is projected to reach 35 billion yuan ($4.9 billion) by 2030.
Chinese firms are well placed to deliver. A 19GWh Abu Dhabi project, the largest single storage order globally, was completed by BYD and Sungrow, requiring complex tuning to combine solar variability with GPU load spikes. Sunwoda’s AI data center storage orders grew nearly 30-fold in the first half, while Shuangdeng’s rose over 80 percent. Chinese firms make 83% of the world’s lithium batteries and offer full capability from cells to system integration, an edge built over years in electric vehicles and renewables.
Source: Huxiu.com, September 26, 2026
https://www.huxiu.com/article/4894068.html