Coordinating data centers to intelligently utilize grid power based upon real-time conditions. Meanwhile in the USA, data centers switching over to backup generators during a brown-out almost destabilized the entire grid recently.

Shanghai has pioneered the synergisation of computing centres with the power grid for peak-load management – a national first – as part of China’s push to enhance computing-grid coordination as its upgrades power infrastructure amid the global artificial intelligence race.

Led by State Grid’s Shanghai subsidiary, the trial achieved a peak load reduction of 97.8 megawatts within a two-hour window, marking the country’s largest single city-level adjustment of computing power load, according to a report on Wednesday by Jiefang Daily, a Shanghai-based official newspaper.

This is equivalent to shutting down the operation of almost 140,000 of Nvidia’s flagship H100 graphics processing units (GPUs) all at once in terms of power capacity. Each unit requires up to 700 watts of power for maximum performance, according to the calculation by the South China Morning Post.

A total of 16 data centre operators took part in the trial, including e-commerce giant Alibaba Group Holding, China Telecom’s Shanghai branch, and data centre firm GDS Holdings. Alibaba owns the SCMP.

Computing-grid coordination aims to transform AI data centres into flexible shock absorbers, which are dispatched by the grid system to ramp up energy-intensive training during green energy surpluses, and scale back operations to shave peak demand during grid strain.

According to State Grid’s Shanghai branch, this latest trial marked the first time a Chinese city had been able to run through the entire technical pipeline of such coordination simultaneously. This involved modulating the data centres’ operational power usage, switching to backup diesel generators when necessary, and shifting computing tasks to other regions to ease the load.

Shanghai’s trial comes as China faces a surge in power consumption driven by the global AI race, which has prompted policymakers to increasingly view computing power as a national utility rather than simply as a resource owned by private firms.