MANY MINDEDBiren pitches light-linked chip clusters to scale AI past copper limits
Shanghai-based Biren Technology unveiled next-generation 'supernode' systems designed to link thousands of AI chips into a single cluster. The announcement was made at the 2026 World Artificial Intelligence Conference in Shanghai on 18 July 2026. Biren's approach uses optical data transmission and near-packaged optics to move data between chips faster than conventional wiring allows.
The company's rationale, laid out by AI framework architecture vice-president Ding Yunfan, is that single-GPU performance is no longer enough for the workloads AI now demands. Ding said traditional copper-wired electrical connections cap current server architectures at 128 GPUs. Biren's distributed, decoupled architecture is designed to scale up to 1,024 cards across a cluster, though that figure is described as a potential capability rather than a demonstrated one.
Scaling out matters for abundance because the bottleneck in AI compute is increasingly about connecting many chips efficiently, not just making each chip faster. If clusters can grow eightfold past today's copper limits, more compute becomes reachable from the same generation of hardware. For a company working around chip export limits, that also means squeezing more capability from the silicon that is available.
The honest caveat is that this is an architectural bet, not a shipped product with proven results. The 1,024-card figure is a target. What to watch is whether Biren can demonstrate the supernode at scale and whether optical interconnect delivers the efficiency gains it promises in real deployments.
Source: SCMP Tech