先说结论：光通信不是 AI 外围，而是规模化以后绕不开的问题Main Point: Photonics Is Not Peripheral Once AI Scales
GPU 越多，机器之间要搬的数据越多；当铜线功耗、延迟和距离撑不住，光通信就从配角变成关键约束。More GPUs mean more data movement. When copper cannot handle power, latency, and distance, photonics moves from side role to constraint.
AI 数据中心不是把 GPU 堆得越多越好。GPU 要一起训练、一起推理，就必须高速交换数据。规模越大，连接越像瓶颈：距离更长，带宽更高，功耗更重，故障点更多。An AI data center is not simply “more GPUs.” GPUs need to train and infer together, which means high-speed data exchange. As clusters grow, interconnects become a bottleneck: longer distance, higher bandwidth, more power, and more failure points.
这就是为什么市场会从 NVDA 这种大明星，顺着网络设备、光模块、硅光、激光器一路往下看。小公司被放进显微镜，不是因为它突然变大，而是因为它可能站在一条变窄的通道上。That is why the market moves from a large anchor like NVDA down through networking, optical modules, silicon photonics, and lasers. A small company enters the microscope not because it has become large, but because it may sit on a narrowing passage.
先认机器：为什么铜线会被光逼到墙角Machine First: Why Copper Gets Pushed into a Corner
传统连接里，电信号要在芯片、板卡、接口和线缆之间跑。速度越高、距离越长，损耗和功耗越明显。到了大规模 AI 集群，很多连接不再是几厘米的小问题，而是跨机柜、跨交换层的系统问题。In traditional connections, electrical signals travel across chips, boards, connectors, and cables. The faster and farther they go, the more loss and power matter. In large AI clusters, links are no longer centimeter-level board problems; they become rack and network-fabric problems.
CPO 的想法，是尽量把光电转换靠近交换芯片，减少电信号在板上的长距离奔跑。这样做可能降低功耗、提升信号质量，也会改变供应链里谁更重要。The idea behind CPO is to move optical conversion closer to the switch silicon, reducing long electrical runs on the board. That can cut power, improve signal quality, and change which suppliers matter.
供应链地图：从 GPU 往下看，会看到哪些层Supply-Chain Map: What You See When You Look Downstream from GPUs
| 层级 | 角色 | 读者要问 |
|---|---|---|
| GPU / 加速器 | 制造算力需求 | capex 是否继续 |
| 网络 / 交换芯片 | 连接机器 | 带宽和功耗是否成为约束 |
| 光模块 / CPO | 把数据变成光传输 | 路线是否进入量产 |
| 硅光 / 激光器 | 提供关键光源和集成 | 谁能通过客户认证 |
| 封装 / 测试 / 材料 | 决定交付能力 | 良率和产能是否跟得上 |
第一层是 GPU 和加速器，它们制造算力需求。第二层是网络和交换芯片，负责让机器互相说话。第三层是光模块和 CPO，决定数据能不能更快、更省电地移动。再往下，是硅光平台、激光器、外部光源、封装测试、衬底和材料。The first layer is GPUs and accelerators, which create compute demand. The second is networking and switch silicon, which lets machines talk. The third is optical modules and CPO, which determine whether data moves faster and with less power. Below that are silicon-photonics platforms, lasers, external light sources, packaging and testing, substrates, and materials.
Serenity 式分析会问：哪一层如果停了，下游路线图会被迫延期？答案不一定是最有名的公司，反而可能是体量小、技术窄、替代慢的供应商。This analysis asks which layer would delay downstream roadmaps if it stopped. The answer is not always the famous company; it may be a small, narrow, hard-to-replace supplier.
当时股票环境：大公司确认方向，小公司放大弹性Market Context: Large Companies Validate Routes; Small Companies Amplify Elasticity
截至 2026-06-04，NVDA 市值约 5.19 万亿美元，它的路线图和生态动作会影响整个 AI 基建链条。GFS 约 464 亿美元市值，POET 约 8.5 亿美元市值。体量越往下，价格越容易被单条证据推动。As of 2026-06-04, NVDA is around $5.19T in market value, and its roadmap shapes the broader AI infrastructure chain. GFS is around $46.5B, while POET is around $850M. As market value gets smaller, a single evidence point can move price more.
这不是说小公司一定更好，而是说它们对信息更敏感。一个合作声明、一个客户线索、一次财报措辞变化，都可能让市场重新想象它的位置。读者要做的是把想象和证据分开。That does not mean small companies are automatically better. It means they are more sensitive to information. A partnership, customer clue, or change in financial wording can reshape market imagination. The reader's job is to separate imagination from evidence.
反方：光通信大方向对，也不代表每个小公司都能赢Bear Case: The Photonics Direction Can Be Right While Many Small Companies Lose
技术路线正确，不等于标的正确。CPO、硅光、外部光源都可能变重要，但具体赢家还要看客户、封装、良率、成本、资本开支和交付周期。A correct technical route does not mean every ticker is correct. CPO, silicon photonics, and external light sources may all become important, but winners still depend on customers, packaging, yield, cost, capex, and delivery timing.
小公司还要面对更现实的问题：现金能不能撑到量产，客户会不会集中，合作是否只是研发阶段，收入出现时规模够不够。只要其中一个关键问题没有回答，文章就不能把故事写成确定。Small companies also face practical issues: can cash last until volume, are customers concentrated, is the partnership still in R&D, and is revenue large enough when it arrives? Until key questions are answered, the story cannot be written as certain.
学习 takeaway：先画机器，再看股票Learning Takeaway: Draw the Machine Before Looking at the Stock
看 AI 光通信推文，不要从 ticker 开始。先画机器：GPU 在哪里，交换芯片在哪里，光模块在哪里，光源在哪里，谁负责封装和测试。When reading AI photonics posts, do not start with the ticker. Draw the machine first: where are the GPUs, switch chips, optical modules, light sources, packaging, and testing?
机器画清楚以后，再问哪一层最窄、哪家公司最难替代、证据到了哪一道门。这样你看到的就不是热闹，而是一条可以复核的产业链。Once the machine is clear, ask which layer is narrowest, which company is hardest to replace, and which proof gate has been reached. Then the story becomes a verifiable chain, not just market noise.