LightCounting LightTrendsCodex · 实盘不做把握 6/10不做原因:消息已过时
北京时间·源无时间·以抓取为准·⟳ 抓取 22:50:13·⚡ 决策 08-25 23:45:31 +55m18s·62.7 天前
LightCounting复盘华为Atlas 950 SuperPoD采用4096个800G LPO光模块
本条不是2026-08-25的新催化，而是7月21日LightCounting对WAIC展品的研究复盘；即使方向上利好国产AI光互联链条，也无法确认未被市场消化，且缺少可量化订单和可执行价格。
📄 原文LightCounting LightTrends↗ 原文
Atlas 950 SuperPoD: Achieves All-Optical Interconnection with Superior Cost and Power...
LightCounting reports on highlights from WAIC 2026 by Carol Cao July 21, 2026 The 2026 World Artificial Intelligence Co…
不做 华工科技（000988.SZ）
不做华工科技（000988.SZ）：LPO供应链逻辑方向为正，但本次材料发布已约35天，缺少一手订单金额、收入弹性和真实行情验证，price-in风险高。
展开详情(成交明细 / 研究档案 / 原文全文 / 决策记录链)
研究档案
结论摘要
事件事实指向华为Atlas 950 SuperPoD提升800G LPO光模块需求，但原文日期为2026年7月21日，核心展会发生在7月17日至20日，当前已滞后约35天；华为未上市，二阶映射到华工科技等光模块链条需要一手客户关系、订单金额和最新行情支持，现有公开材料不足以形成正期望交易。
消息传播链
- 华为官方披露昇腾950超节点真机在WAIC 2026首次公开亮相。
- LightCounting发布研究复盘，补充4096个800G光模块、LPO和机柜结构等细节。
- 中文产业分析继续传播Atlas 950 SuperPoD和800G LPO光模块供应链叙事。
关键理由
- Atlas 950 SuperPoD采用4096个800G光模块，语义上利好国产AI光互联和LPO产业链。
- 华为官方确认Atlas 950 SuperPoD在WAIC 2026真机亮相，说明超节点方案处于公开展示阶段。
- 但本次材料是一个月前研究复盘，不是新订单或新增量产公告。
主要风险
- 未披露订单金额、供应份额和量产节奏，无法估算对华工科技收入和利润的增量占比。
- 事件已在7月WAIC及后续产业文章中传播，2026年8月25日再交易存在明显price-in风险。
反向证据
先例如何影响了这次决策
对本次决策的影响:先例不足叠加新鲜度不足，最终不做。
方向:没有完成可审计event_study，不能用历史样本支持入场。
涨幅量级:缺少同市场、同机制、同相位且首发锚点干净的高相似样本，无法校准tp/sl。
形态:本事件属于展会技术复盘和供应链读透，滞后传播后容易出现见光死或无反应风险。
原文全文
展开原文 · 3,120 字
LightCounting reports on highlights from WAIC 2026
by Carol Cao
July 21, 2026
The 2026 World Artificial Intelligence Conference (WAIC) was held in Shanghai from July 17th to 20th. It centered on intelligent computing infrastructure, general-purpose large models, embodied intelligence and industrial AI applications, featuring technology showcases and professional academic exchanges.
Huawei’s Atlas 950 SuperPoD made its first public appearance at WAIC 2026, serving as the conference’s main highlight. The showcased system combined 1,024 Ascend 950 NPUs into a SuperPoD, using UnifiedBus (UB) 2.0 communication protocol and a hybrid mesh/Clos network architecture.
Atlas 950 SuperPoD also brings substantial optimizations at chip, rack, and system levels, with major improvements in performance, energy and cost efficiency. It delivers 1 EFLOPS (FP8) / 2 EFLOPS (FP4) peak compute, supports a 256TB global unified memory address space, provides TB scale inter NPU bandwidth, and achieves end-to-end RTT as low as 3μs. In comparison, CloudMatrix 384, released last year, features 48TB global unified memory and 392 GB/s one way bandwidth per NPU.
The system’s network relies on both copper and optical interconnects:
The system on display and shown in the photo above, included 16 compute racks and 4 central switch racks interconnected with 4,096 of 800G optical transceivers. It is the first supernode design adopting Linear-drive Pluggable Optics (LPO). Built on Huawei’s 100G VCSEL technology and opto-electronic co-design capabilities, the LPO solution optimizes the system costs, latency and power consumption.
Compared with CloudMatrix 384, Atlas 950 SuperPoD adopts a more compact cabinet layout.
It can be seen from the front view of Atlas 950 SuperPoD, each compute rack is configured with 8 NPU trays and 8 CPU trays. Each NPU tray carries 8 NPUs, reaching a total of 64 NPUs per compute rack.1024-NPU cluster is composed of 16 compute racks plus 4 central switch racks. By comparison, CloudMatrix 384 is built with 12 compute racks and 4 switch racks. Each compute rack contains 4 compute nodes, and each node is equipped with 8 NPUs together with host CPUs.
From the rear view, the quantity and layout of ports have changed. The compute rack of CloudMatrix 384 has 256 QSFP-DD modules in 4 vertical columns, while Atlas 950 SuperPoD has 128 modules in 8 columns. The configuration has evolved from 8×400G modules per NPU to 2×800G modules per NPU. Calculating from the on-site displayed compute racks, the total number of optical modules required for the compute part of SuperPoD is 4,096 (including server and UB switch sides). On average, each NPU is associated with four 800G OSFP modules.
An orthogonal electrical interconnection is deployed between the front tray and the rear plug-in boards of the compute rack of Atlas 950 SuperPoD.
LightCounting will publish Cloud Data Center Optics report at the end of July, covering network infrastructure and optical interconnects across major cloud companies.
Full text of the research note is available to subscribers at: https://www.lightcounting.com/login.
by Carol Cao
July 21, 2026
The 2026 World Artificial Intelligence Conference (WAIC) was held in Shanghai from July 17th to 20th. It centered on intelligent computing infrastructure, general-purpose large models, embodied intelligence and industrial AI applications, featuring technology showcases and professional academic exchanges.
Huawei’s Atlas 950 SuperPoD made its first public appearance at WAIC 2026, serving as the conference’s main highlight. The showcased system combined 1,024 Ascend 950 NPUs into a SuperPoD, using UnifiedBus (UB) 2.0 communication protocol and a hybrid mesh/Clos network architecture.
Atlas 950 SuperPoD also brings substantial optimizations at chip, rack, and system levels, with major improvements in performance, energy and cost efficiency. It delivers 1 EFLOPS (FP8) / 2 EFLOPS (FP4) peak compute, supports a 256TB global unified memory address space, provides TB scale inter NPU bandwidth, and achieves end-to-end RTT as low as 3μs. In comparison, CloudMatrix 384, released last year, features 48TB global unified memory and 392 GB/s one way bandwidth per NPU.
The system’s network relies on both copper and optical interconnects:
The system on display and shown in the photo above, included 16 compute racks and 4 central switch racks interconnected with 4,096 of 800G optical transceivers. It is the first supernode design adopting Linear-drive Pluggable Optics (LPO). Built on Huawei’s 100G VCSEL technology and opto-electronic co-design capabilities, the LPO solution optimizes the system costs, latency and power consumption.
Compared with CloudMatrix 384, Atlas 950 SuperPoD adopts a more compact cabinet layout.
It can be seen from the front view of Atlas 950 SuperPoD, each compute rack is configured with 8 NPU trays and 8 CPU trays. Each NPU tray carries 8 NPUs, reaching a total of 64 NPUs per compute rack.1024-NPU cluster is composed of 16 compute racks plus 4 central switch racks. By comparison, CloudMatrix 384 is built with 12 compute racks and 4 switch racks. Each compute rack contains 4 compute nodes, and each node is equipped with 8 NPUs together with host CPUs.
From the rear view, the quantity and layout of ports have changed. The compute rack of CloudMatrix 384 has 256 QSFP-DD modules in 4 vertical columns, while Atlas 950 SuperPoD has 128 modules in 8 columns. The configuration has evolved from 8×400G modules per NPU to 2×800G modules per NPU. Calculating from the on-site displayed compute racks, the total number of optical modules required for the compute part of SuperPoD is 4,096 (including server and UB switch sides). On average, each NPU is associated with four 800G OSFP modules.
An orthogonal electrical interconnection is deployed between the front tray and the rear plug-in boards of the compute rack of Atlas 950 SuperPoD.
LightCounting will publish Cloud Data Center Optics report at the end of July, covering network infrastructure and optical interconnects across major cloud companies.
Full text of the research note is available to subscribers at: https://www.lightcounting.com/login.
决策记录链 · 只追加
每次领取产生一条记录,重新决策追加一条并以 supersedes 指向被替代的记录(ADR 0014)。
这条决策早于决策账本(2026-10-08 起逐条追加),没有追加记录。