Chinese Chip Start-up Biren Bets on Light-Based 'Supernodes' to Match Nvidia

At WAIC, Biren unveiled optical-interconnect systems it says can link up to 1,024 AI chips in one cluster, far beyond the roughly 128 GPUs that copper links effectively cap.

Jul 19, 2026Source: South China Morning Post

Chinese GPU start-up Biren Technology used WAIC in Shanghai to stake out a different path to scale. On July 18, it unveiled next-generation "supernode" systems that use optical data transmission — light rather than copper — to lash large numbers of AI chips together into a single cluster.

Breaking the copper ceiling

The core claim is about interconnect. Traditional copper links effectively cap a server at around 128 GPUs before signal loss and power draw become prohibitive. Biren's design, built on near-packaged optics (NPO) that move optical fibers closer to the chips, is engineered to scale up to 1,024 processor cards in one cluster — an order-of-magnitude increase in how many accelerators can act as one machine.

'An inevitable choice'

Biren's VP of AI framework architecture, Ding Yunfan, framed the shift bluntly: "optical interconnect has become an inevitable choice." His argument is that single-GPU performance is no longer the bottleneck; the hard problem is stitching thousands of chips into one coherent system, and copper simply cannot carry the bandwidth that requires.

Why it matters for China

The pitch is inseparable from geopolitics. With US export controls limiting access to Nvidia's most advanced accelerators, Chinese firms are trying to compensate at the system level — winning on how chips are connected rather than on the raw performance of any single die. If clustering can offset a per-chip disadvantage, interconnect becomes a way around the sanctions math.

The funding backdrop

Biren is scaling this ambition against a fundraising push; the company has separately been reported seeking around $900 million to fund its GPU roadmap. The supernode reveal is its clearest technical answer yet to the question of how a sanctioned ecosystem builds frontier-scale compute.

Why It Matters

Optical interconnect is where China is trying to turn a chip disadvantage into a systems advantage: if you cannot buy Nvidia's fastest GPUs, link far more of your own into one cluster. Nvidia itself is moving the same direction with rack-scale designs, so Biren is racing a moving target — but a credible 1,024-card optical supernode would give Chinese labs a way to assemble frontier-scale compute despite export controls.

Filed underNvidia AI Chips #Biren #China AI #optical interconnect