Huawei Connect 2026 Unveils 11 AI Chips; Ascend 960 Super Node First to Adopt Near-Package Optics
nashnova research
Huawei unveiled over 10 AI infrastructure chips at Connect 2026, headlined by the Ascend 960 SuperNode — the industry's first supernode to use near-package optics. The full lineup targets Nvidia, Intel, and AMD head-on, signaling Huawei's push from single-chip competition to system-level infrastructure rivalry.
What did Huawei just launch?
Rotating Chairman Wang Tao announced on September 17 in Shanghai a lineup of over 10 chips spanning AI accelerators, CPUs, high-speed interconnects, storage and storage-controller chips, network interface controllers, and a near-package optics (NPO) module with an in-house laser.
This means → Huawei is no longer selling standalone accelerators. It is fielding a full AI infrastructure product line that mirrors the breadth of Nvidia, Intel, and AMD.
In plain terms = the competition has moved from "your GPU vs. my Ascend" to an entire stack — chips, optical links, storage, and networking — laid out on one stage.
When will the Ascend 960 be ready?
The Ascend 960DT is set for mass production in Q1 2027, three quarters ahead of the original schedule, with performance double the previous generation.
The follow-on cadence is locked in: the 960PR arrives in Q3 2027, Ascend 970 in 2028, Ascend 980 in 2029 — one generation per year.
This means → Huawei is trying to shift its product rhythm from "chasing the leader" to a predictable annual cadence, giving customers a roadmap they can plan around.
What is "near-package optics," and why does it matter?
The Ascend 960 SuperNode is the first supernode product to use NPO — near-package optics, a technology that brings optical connections right next to the chip package. Built on Huawei's in-house UnifiedBus architecture and Hi-ONE optical engine, it links up to 4,096 accelerator cards with round-trip latency as low as 2 microseconds.
In plain terms = traditional supernodes move data between racks over electrical signals that slow down over distance. NPO pulls the optical highway right up to the chip's front door.
This reflects a direct technology-path rivalry with the co-packaged optics (CPO) approach led by Nvidia and Broadcom — Huawei published its own NPO standard just last week.
How far along are existing products?
Ascend 910C has completed 1,000 unit deliveries. Ascend 950 is in large-scale commercial deployment.
SuperNode clusters have been shipped to over 370 customers — more than 1,000 units — and the largest super-cluster architecture scales to one million AI processors.
This means → Huawei is not just announcing new products. The previous generation is already running at customer sites — proven delivery at scale is the prerequisite for customers to bet on the next generation.
Which software gap did Huawei just close?
Ascend is now the fourth hardware platform officially supported on the PyTorch website, after Nvidia, AMD, and Intel — and the first Chinese hardware to earn that certification.
Third-party communities including Triton, vLLM, and VRL have completed integration with Ascend. The Linux Foundation has also extended support.
In plain terms = a chip is only as useful as the developers willing to write for it. Official PyTorch support is effectively a ticket into the global AI developer ecosystem.
What is the next milestone to watch?
Huawei plans to integrate its computing and communications infrastructure more tightly, building a next-generation network around AI workload delivery.
Whether the Ascend 960DT hits volume production on schedule in Q1 2027 is the single most important proof point for the entire roadmap.
This means → the blueprint is drawn. From here, the market cares about one thing: on time, at volume, at spec.
市场有风险，内容仅供研究参考，不构成投资建议。