Cambricon's 6th-Gen AI Chip Under Development; H1 Revenue Surges 108% YoY
Nashnova编辑部
Cambricon posted H1 revenue of RMB 5.99 billion and net profit of RMB 2.31 billion, up 108% and 123% year-on-year respectively; its sixth-generation AI processor is now in R&D targeting large-model training and inference — the key window for the market to judge whether it can evolve from an inference-chip maker into a full training-plus-inference platform.
What do the headline numbers actually tell us?
H1 revenue hit RMB 5.99 billion, up 108% YoY; net profit reached RMB 2.31 billion, up 123% YoY.
This means → Cambricon has crossed the "still losing money" stage. Profit is growing faster than revenue, signaling that scale economics are kicking in.
One number stands out: the cloud product line accounted for 99.98% of total H1 revenue. In plain terms = almost every yuan comes from cloud AI accelerators — extreme concentration. When cloud demand is strong, this is a tailwind; if it wobbles, there is virtually no cushion.
What problem is the sixth-gen chip meant to solve?
Chairman Chen Tianshi said the sixth-generation AI processor and its instruction set are under development, targeting improvements across programmability, usability, performance, energy efficiency, and die area.
This means → Cambricon is not just chasing raw speed — it wants to make its chips easier for developers to adopt. That is precisely the moat Nvidia's CUDA ecosystem holds today.
The bigger move: the next-gen platform is explicitly aimed at large-model training and inference. This reflects an attempt to leap from "strong at inference" to "competitive in training, too" — whether training workloads can run at scale on its hardware will be the market's defining test.
How far has large-model compatibility come?
On the inference side, Cambricon has completed adaptation for five mainstream Chinese large models: Zhipu AI's GLM, DeepSeek, Alibaba's Qwen, Moonshot's Kimi, and MiniMax.
On the training side, the company is optimizing for DeepSeek, Qwen, and Tencent's Hunyuan, with plans to scale to clusters of 1,000 to 10,000 cards.
In plain terms = inference already "runs"; training is still being tuned. The real test is whether a 10,000-card cluster can sustain training workloads with adequate communication efficiency, fault tolerance, and long-run stability — those are the hard thresholds for large-scale training.
Is the R&D ramp-up sufficient?
H1 R&D spending was roughly RMB 702 million, up 29.63% YoY; R&D headcount grew from 792 to 1,007, a ~27% increase.
This means → R&D investment is accelerating, but its growth rate (~30%) trails revenue growth (108%) by a wide margin. This reflects a company in the "ship product and earn, while doubling down on the next generation" phase — profits can temporarily fund the R&D expansion.
The software platform is advancing in parallel — covering large-model inference, compute-graph compilation, operator programming, acceleration libraries, and communication. Put simply = the chip is just the hardware base; whether the software ecosystem lets customers plug in and go will determine sixth-gen market adoption.
Content is for reference only, not financial advice.