China unveils ‘brain-like’ AI server that uses 90% less power: Report
Guangdong Institute of Intelligence Science and Technology (GDIIST) has unveiled the BI Explorer (BIE‑1), a mini‑fridge‑sized AI server it has said can rival a supercomputer while using one‑tenth of the power, with specs including 1,152 CPU cores and 4.8 TB memory.
China’s Guangdong Institute of Intelligence Science and Technology (GDIIST) has revealed a compact, “brain‑like” AI server that it has said can deliver supercomputer‑class performance while consuming around 90 per cent less power than conventional systems.
The BI Explorer computing system, or BIE‑1, were introduced at a forum in the Guangdong–Macao In‑Depth Cooperation Zone on 24 October 2025.
A mini‑fridge‑sized “personal supercomputer”
The BIE‑1 is described as roughly the size of a small, single‑door refrigerator and has been designed to run from a standard household socket with low noise, making it suitable for homes, small offices and mobile environments.
GDIIST said the unit offers training and inference speeds that rival traditional computing clusters while using a fraction of the energy.
Local launch materials listed 1,152 CPU cores, 4.8 TB of DDR5 memory and 204 TB of storage, with noise kept below 45 dB and temperatures reportedly stable under 70°C at peak load.
Who is behind BIE‑1?
GDIIST has incubated the project and two affiliated companies—Zhuhai Hengqin Neogenint Technology and Suiren (Zhuhai) Medical Technology—have jointly brought the device to market.
At the launch, Nie Lei, general manager of Neogenint and head of GDIIST’s Intelligent Computing Systems Joint Laboratory, outlined the goal of “packing the power of a supercomputer into the size of a small refrigerator.”
The institute is led by Professor Zhang Xu, an academician of the Chinese Academy of Sciences, who positioned brain‑inspired computing as a route to lower‑energy, scalable AI infrastructure.
What does “brain‑like” computing involve?
Brain‑inspired (or “brain‑like”) computing draws on principles of neural processing to design algorithms and systems that perform useful work while activating fewer computational elements—thereby saving energy.
Chinese teams have recently demonstrated “spiking” and localised‑attention models that fire only the neurons needed for a task, reporting speed and efficiency gains on long‑context problems and reducing reliance on conventional GPU hardware.
The launch arrives as China seeks to expand AI capacity while curbing energy use and overcoming hardware constraints.
National plans have aimed to knit together distributed computing resources and improve efficiency across training and inference, part of a broader push to close the gap with the United States on AI infrastructure.