Per the September 13 AI Daily, Cambricon's new flagship cloud chip, the Siyuan 690, entered mass production in early 2026: dual-die packaging, FP16 compute above 700 TFLOPS, 196GB HBM3, and measured interconnect bandwidth above 890 Gbps—nearly double the previous 590.
With its MagicMind + NeuWare unified stack, it achieved "Day 0" adaptation to DeepSeek V4 and V3.2 (chip adaptation and open-source code on the model's release day), with Zhipu's GLM-5 synchronized on day one—upgrading the China compute–model relationship from "can adapt" to "synchronized at launch."
H1 results: revenue 5.996 billion yuan (+108.13% YoY), net profit 2.311 billion yuan (+122.61%), a seventh consecutive profitable quarter; Q1 contract liabilities surged to 396 million yuan as orders piled up.
As Nvidia's China share fell from 40% to 8%, Cambricon is capturing a sizable share of training and inference via "Day-0 software-hardware synergy," cementing the domestic AI industry's closed loop.
Source: AI Chinese Community
