--- title: "Huawei's New Supernode Puts 4,096 Chips in One \"Computer\" — But the Win Is in the Glass, Not the Silicon" date: 2026-09-16 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-ascend-cloud language: en --- # Huawei's New Supernode Puts 4,096 Chips in One "Computer" — But the Win Is in the Glass, Not the Silicon > At Huawei Connect 2026 (17 September) Huawei launched the Ascend 960 supernode (昇腾960超节点), billed as the world's first to use NPO optical interconnect, scaling to 4,096 accelerators. With Cambricon's trillion-yuan market cap, China's domestic compute build-out is moving from chip races to system races. Someone figured out that the bottleneck in AI is no longer the chip — it is the wire between the chips. Huawei's answer, unveiled this month, puts the optical transceiver next to the silicon and calls the whole thing one computer. On **17 September 2026**, at Huawei Connect 2026 in Shanghai, Rotating Chairman **Wang Tao (汪涛)** launched the **Ascend 960 supernode (昇腾960超节点)** — described as the world's first supernode to use **NPO (Near-Packaged Optics, 近封装光学)**. The pitch is simple to state and hard to build: connect up to **4,096 accelerators** so they behave like a single machine, and do it with light instead of copper at the edges. ## What "supernode" actually means A supernode is a middle layer between one server and a whole data centre. Huawei's design uses the **LingQu (灵衢)** high-speed interconnect plus a new optical engine called **Hi-ONE** to give every component equal access to memory across physical servers. The result: - **Scale:** up to **4,096 cards** in one supernode, up from **1,024** in the previous Ascend 950 generation. - **Compute:** **8 EFLOPS FP8** and **16 EFLOPS FP4** per supernode. - **Optics:** **5,500 Hi-ONE modules** replace what would have been **48,000 800G optical modules**, cutting power by more than **550 kW** and lifting availability to **99.8%**. - **Cluster:** multiple 960s link via LingQu or RoCE into clusters of **512,000 cards**, and with multi-rail topology up to **1,000,000 cards**. The headline claim is that NPO — turning electricity to light centimetres from the chip — breaks the distance and bandwidth wall that copper hits. Hi-ONE is billed as the first mass-produced NPO product, at **7.2T** capacity, and the only one with a built-in light source. Huawei says it has opened the LingQu 2.0 protocol to the industry and is joining a MIIT effort to standardise domestic interconnect — a tell that the real prize is an ecosystem, not a single box. For context, Nvidia's GB200 NVL72 packs 72 GPUs; Huawei's argument is not that one 960 beats one NVL72 on efficiency — it is that 4,096 domestic accelerators wired as one machine clear a scale that Western export rules block. The win is in the system, not any single chip. ## The context: China's compute build-out The 960 is the front edge of a stack already in the field. - The **Ascend 910C** (CloudMatrix 384) supernode, 384 NPUs plus 192 Kunpeng CPUs, has been deployed in **more than 1,000 sets** across internet, telecom, finance and other sectors. - The **Ascend 950** supernode (Atlas 950 SuperPoD) — **8,192 Ascend 950DT chips**, **1 EFLOPS FP8**, **256 TB** unified memory — is slated for commercial launch in Q4 2026. - Huawei says the air-cooled and liquid-cooled 960 supernodes will ship in **Q2 and Q3 2027**. The strategic logic is explicit: if you cannot buy the fastest chip, build the largest, best-wired machine from domestic parts. ## The domestic rival: Cambricon's quiet quarter Huawei is not the only one. **Cambricon (寒武纪)**, the other flagship of Chinese AI silicon, posted a breakout **2026 first quarter**: revenue of **2.885 billion yuan (≈ US$406M / HK$3.14B)**, up 159.6%, and net profit of **1.013 billion yuan (≈ US$143M / HK$1.10B)**, up 185%. Its **MLU590** chip — 7 nm, about **256 TFLOPS FP16** and **512 TOPS INT8** — is in volume delivery and pitched at roughly 80% of Nvidia's A100. On **30 June 2026** Cambricon's market cap crossed **1 trillion yuan (≈ US$140B / HK$1.09T)**, the first STAR Market company to do so. The synergy is the part worth watching: **DeepSeek's V4-Pro** achieved Day-0 adaptation to the Ascend platform in 2026, and several reports say DeepSeek chose Ascend 950PR as a flagship compute base. Model and silicon, both domestic, closing the loop. ## The honest limits The numbers are impressive and almost entirely Huawei's or Cambricon's own. - **Benchmarks are vendor-stated.** FP8/FP4 throughput, the 550 kW saving and 99.8% availability come from Huawei's launch, not an independent lab. - **"World's first NPO supernode" is Huawei's framing.** NPO is an industry-wide research area; the claim is about mass production and integration, not the concept. - **Scale needs yield.** Building 5,500 optical engines per supernode at volume is a manufacturing bet, not yet a track record. - **Export controls are the backdrop, not the cure.** Domestic parts solve availability, not the software-ecosystem gap versus CUDA. ## What readers can do now - **If you buy AI compute in China, ask for cluster-level numbers, not chip-level peaks** — MFU, latency, and real token throughput per supernode. - **Track Cambricon's MLU590 adoption** as the canary for non-Huawei domestic silicon. - **Watch the 950/960 commercial deployments in Q4 2026–2027** before treating "million-card clusters" as deployed reality. ## Honest limitations Ascend 960 figures come from Chinese state and technical media (Xinhua, Science and Technology Daily, CGTN, Tencent/Universal coverage of Huawei Connect 2026) and Huawei's own disclosures. Cambricon financials are from Eastmoney and the company's filings. Deployment counts (1,000+ 910C supernodes, 750+ 384 systems) vary slightly by source and date and are Huawei-reported. The "world's first NPO supernode" and all performance figures are vendor claims not independently audited. Exchange rates use roughly 7.1 RMB/USD and 1.09 RMB/HKD. Reporting is current to late September 2026. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-ascend-cloud Free to quote with attribution and a link to the original.