Huawei Outlines AI Strategy Focused on Hardware Monetization and Ascend Chip Roadmap
At Huawei Connect, Huawei announced that its Ascend 950 supernodes have achieved large-scale commercial deployment with over 1,000 sets deployed, while revealing an accelerated release schedule for its Ascend 960 chips. The Ascend 960DT is set to launch in Q1 2027 with doubled performance, alongside the industry's first Near-Packaged Optics (NPO) supernode architecture. Huawei's focus on custom HBM memory and supernode scaling demonstrates China's push to build an independent, end-to-end AI computing infrastructure. By accelerating its chip roadmap and adopting novel interconnects, Huawei aims to position its Ascend ecosystem as a primary alternative to Western AI hardware suppliers. The Ascend 950 series integrates Huawei's self-developed HiBL 1.0 HBM memory, offers 2.5x higher interconnect bandwidth, and splits into 950PR (optimized for LLM prefill) and 950DT (optimized for decode and training). The upcoming Atlas 960 liquid-cooled supernode is scheduled for Q3 2027, followed by annual generation updates including Ascend 970 in 2028 and Ascend 980 in 2029.
## BACKGROUND
Near-Packaged Optics (NPO) is an optical interconnect architecture that places optical transceivers on the circuit board right next to the processor ASIC, drastically reducing latency and power consumption compared to traditional pluggable optical modules. High Bandwidth Memory (HBM) is a high-speed memory interface crucial for modern AI accelerators to process large neural network parameters efficiently.