
中国的光模块制造商和PCB生产商正在报告强劲需求,因为国内AI芯片集群正在扩大规模。正是这层不起眼的基础设施使得华为昇腾集群在无需英伟达的情况下可行——供应链正在形成。
简单来说:要在中国构建无需英伟达芯片的AI集群,就需要自有版本的每一个组件——芯片、内存、互连、网络、电源。而光模块和PCB层则是目前实际交付的部分。
据Pandaily报道,中国的光模块制造商和PCB生产商正在因国内AI芯片集群部署而报告强劲需求。光互连——即连接芯片与服务器间的高速光纤链路——是大规模AI训练集群中的关键层,其中GPU间的数据传输是主要性能瓶颈。
英伟达出口管制的故事聚焦于芯片。而较少被关注的则是其下层结构:互连、交换结构、电源传输和PCB。一台华为昇腾910C集群若要在性能上与H100相匹敌,其光互连和交换层必须达到同等水平。中国光模块制造商——如Innolight、Eoptolink等——已在过去两年悄然在这一层级建立产能。国内AI部署的强劲需求意味着这些集群正在实际组装和连接,而非仅停留在公告阶段。主权计算栈正从底层开始构建,而光层现已成为生产现实,而非依赖风险。
中国光模块供应商是否开始赢得非中国超大规模数据中心的合同——这将表明主权计算的构建已产出全球竞争力的组件供应商,而非仅满足内部需求。
本文由人工智能撰写,并经人工编辑审核。
China’s focus on domestic optical modules and PCB supply for AI clusters is smart-undervalued but critical. The real test will be scaling this without cutting corners on reliability.
Building the whole stack vertically makes sense, but will China manage to scale without cutting corners on reliability? I worry about long-term stability when pushing domestic alternatives so hard.
China's vertical integration might speed up development, but will their optical modules and chips hold up at hyperscale? Reliability is everything when you're betting the farm on domes
China’s push for vertical integration in compute hardware is bold but risky-can they really match the precision of TSMC’s 3nm process with their own stack without sacrificing efficiency at scale?
China’s push for domestic optical modules makes sense, but will they prioritize raw performance over long-term durability? Hyperscale AI workloads don’t forgive failures.
If this vertical integration pays off, it could give China an edge in specialized AI workloads. But will the global market actually adopt domestically built stacks, or are they just another set of closed ecosystems?
Fascinating how China’s building its own compute stack to reduce dependencies. Wonder if they’ll hit bottlenecks before hitting escape velocity on AI workloads.
It’s smart to secure the whole stack, but how long before foreign alternatives in optics or software force China to keep innovating just to stay in the race?
Compute souverain chinois : nodes legacy, clusters massifs