The global AI “computing power arms race” has focused largely on breakthroughs in chip manufacturing processes, but the real bottleneck constraining computing capacity may well be electricity.
A single AI data centre with 100,000 GPU cards can draw instantaneous power equivalent to the entire electricity load of a small city. As computing clusters grow ever larger, the strain on power systems rises in tandem.
At present, Chinese AI chips still lag behind international leaders in per-card performance. Closing that gap takes time, yet the explosive demand for AI computing waits for no one. In this context, deploying a super-node system architecture, one that interconnects and coordinates tens of thousands of AI accelerators at high speed to boost the system efficiency and deployment capability of large-scale computing clusters, has emerged as a preferred pathway. Global tech giants are also pursuing this route, but China relies on it more deeply and with greater urgency.
Of course, the super-node approach comes with costs: vastly greater power demand, stricter requirements for power supply stability, and more sophisticated energy scheduling capabilities. In other words, whoever solves the “where, how stable, and how efficient” questions of electricity supply in the most computing-dense situation will seize the initiative in the next round of the AI race.
On August 6, Envision Group announced the official commissioning of its “Envision Ulanqab Star River Base,” while simultaneously unveiling what is currently designed to be the world’s largest AI computing super-single-unit at the same campus. Envision described this as the first flagship project under its global “Gobi Mission” initiative—and an engineering answer to the very challenge posed above.
Unlike conventional data centres, this super-single-unit has tightly coupled computing and energy from the ground up. According to Envision, it features a high proportion of direct green-power connections, i.e. wind-solar-storage plants linked via self-built or dedicated lines that feed renewable electricity directly to the data centre, bypassing the public grid to reduce both electricity costs and carbon footprint. With a total floor area exceeding 120,000 square meters (equivalent to 20 standard soccer pitches), once fully operational it is expected to become the world’s single-largest AI data centre in terms of token output capacity.
In terms of energy architecture, the super-single-unit leverages its high-percentage direct green-power connection and an AI-driven power system, with a total planned campus capacity of 2 GW. Based on public industry data, a 1 GW data centre operating at full load consumes roughly the same amount of electricity annually as a city of one million people.
In terms of the computing scale, the unit supports the simultaneous operation of one million GPU cards, with a computing power scale of one million petaflops. (One “petaflop” denotes one quadrillion floating-point operations per second, roughly equivalent to the combined computing power of 500,000 mainstream laptops.)
According to Envision, the “Envision Ulanqab Star River Base” achieves the world’s highest computing-density, super-single-unit layout, precisely addressing China’s need for interconnection of one million domestic cards, and aims to provide a replicable “China solution” for indigenous computing clusters.
Source: WeChat Official Account——电池中国
Disclaimer: The content provided above is intended solely for educational and reference purposes. The original author retains the copyright to the articles, and their reproduction is aimed at facilitating learning and knowledge-sharing. If there are any concerns regarding copyright infringement, please contact us for immediate removal. We value the intellectual property rights of the original creator and encourage readers to engage with the original work.
