NVIDIA CPO Switch Enters Mass Production, Supply Chain Roles of Five Key Companies Revealed
Nashnova编辑部
Nvidia's Spectrum-X photonic switch has entered full mass production, packaging the optical engine and switch chip into a single module with five vendors splitting the work — marking AI data-center networking's real shift from pluggable optics to silicon-photonics integration.
What is CPO, and why did mass production take so long?
CPO — co-packaged optics, placing the optical engine and the switch chip inside the same module so signals travel the shortest possible path — had been stuck at the yield-control bottleneck that kept it in the lab-validation stage.
Nvidia and five supply-chain partners have now cleared that bottleneck. This means → CPO has crossed from "can build samples" to "can ship at volume," entering a commercial-deployment cycle.
In plain terms = light and electricity used to travel separate paths, forcing repeated signal conversions between modules — wasteful and error-prone. Now they sit together; conversion distance shrinks to near-zero, cutting both power draw and failure rates.
How big is the performance jump — what do the numbers say?
Compared with conventional pluggable optics, the new architecture delivers a 5× improvement in network power efficiency, 5× longer uninterrupted AI-workload uptime, and a 10× increase in mean time between failures.
The external laser source uses a centralized design, cutting the number of lasers needed by 75% and further reducing power and heat.
This means → the same networking hardware, running AI jobs, fails one-tenth as often — for large-model clusters that need multi-day training runs, that is the most direct efficiency lever.
How do the five vendors split the work?
TSMC fabricates the silicon-photonics chips; ASE Group (矽品精密) handles chip-level packaging and testing — these two steps form the manufacturing core of the CPO module.
Lumentum and TFC supply laser components; Foxconn handles system-level R&D and assembly; Nvidia performs final testing before shipment.
This reflects a structural reality: CPO is not a one-company game. From optical chips to lasers to packaging to system integration, every link requires a specialist — the supply-chain division runs deeper than traditional optical modules.
Switch specs — what do port counts and bandwidth mean?
The 2U chassis model SN6810 holds 128 × 800 Gb/s ports for 102.4 Tb/s total bandwidth; the 5U SN6800 integrates four switch ASICs, offering 512 × 800 Gb/s ports and 409.6 Tb/s aggregate bandwidth.
Larger chassis embed fiber-interconnect modules that enable lateral scaling without adding switch layers. This means → as the network grows, latency does not rise linearly.
Nvidia also launched Spectrum-XGS, extending the same architecture to building- and campus-scale, unifying multiple data centers into a single coordinated system.
How did the market react, and what do brokers say?
Optical-communications stocks rallied after the announcement: Harvest Comms ETF (159695) +3.71%, Cathay Comms ETF (515880) +3.60%, Fullgoal Comms ETF (159583) +3.14% — on a day when all three major Hong Kong indices fell, with comms clearly outperforming.
China Galaxy Securities noted the Spectrum-X solution further validates optical interconnects for next-gen AI data centers, and that rising silicon-photonics penetration will drive rapid growth in upstream optical-chip demand.
Guotai Junan International argued that vertically integrated optical-interconnect technology — CPO and NPO at its core — now has the conditions for scaled replication. This means → the market is shifting from "which technology route wins" to "who can ramp volume first."
What is the key question going forward?
The CPO mass-production bottleneck has been cleared, but whether the technology can sustain cost and efficiency gains across larger-scale data-center deployments is the critical checkpoint for the industry thesis.
In plain terms = the sample stage proved "it can be built"; small-batch production proved "it can be manufactured"; the next proof point is "costs actually come down when you roll it out at scale."
This signals something bigger: the AI-infrastructure bottleneck is migrating from compute chips to network interconnects. Whoever commands optical-interconnect manufacturing at scale holds the entry ticket to the next round of infrastructure buildout.
Content is for reference only, not financial advice.