While Silicon Valley scrambles to squeeze more juice out of increasingly expensive digital chips, Chinese scientists at Peking University just pulled off something that sounds like science fiction.
They've built an analogue AI chip that could leave Nvidia's flagship H100 GPU eating dust.
Led by Sun Zhong, the team cracked a problem that's been bugging engineers for over a century. Analogue computing has always been fast but frustratingly imprecise. Think of it as the difference between a race car with no brakes and a precision watch.
These researchers figured out how to have both speed and accuracy.
Their secret weapon? Resistive random-access memory, or RRAM for short. This isn't your grandmother's memory technology. It's non-volatile, meaning it doesn't forget everything when you unplug it. More crucially, it ultimately gives analogue computing the precision it desperately needed.
The numbers are frankly ridiculous. We're talking about processing speeds potentially 1,000 times faster than Nvidia's best GPU. Energy efficiency? Try 100 times better than today's digital processors.
That's not an incremental improvement. That's a complete paradigm shift.
For AI applications, this could be massive. Training those power-hungry large language models suddenly becomes less of an electricity bill nightmare. Edge computing gets a serious enhancement too, since you won't need to ping the cloud for every calculation.
The timing couldn't be more interesting. While the US tightens export controls on advanced chip equipment, China just demonstrated they're not sitting around waiting.
This breakthrough, published in Nature Electronics, shows domestic innovation can sidestep supply chain restrictions entirely. The research was published on October 13, 2025, marking a significant milestone in analogue computing advancement. This development occurs as job displacement from AI automation is expected to affect millions of workers globally, particularly in the service industry.
Six-G communications could be the initial big winner. Massive antenna arrays need real-time signal processing that current chips struggle with. This analogue approach handles those complex matrix operations naturally, like a fish taking to water. The chip effectively solves large-scale MIMO signal detection problems that plague modern communication systems.
The chip processes continuous signals instead of binary states, making it particularly suited for certain mathematical operations. It's not just faster computing. It's fundamentally different computing.
And right now, China's leading the charge while everyone else is still thinking in ones and zeros.