산업 분석에 따르면 Nvidia의 CPO 대량 생산 확대에 따라 플러거블 광 모듈이 코패키지 광학에 입지를 잃고 있습니다.

On the early morning of August 14, NVIDIA officially announced that its Spectrum-X Ethernet silicon photonics switch has entered full mass production. This marks the world’s first mass-produced 200G-per-lane co-packaged optics (CPO) Ethernet switch, with initial shipments already delivered to CoreWeave, Lambda, and Oracle for deployment.

The full impact becomes clear when viewing three sets of official data together: compared to traditional pluggable solutions, laser counts are reduced by a factor of four, power consumption drops by roughly fivefold, and mean time between failures (MTBF) improves by approximately tenfold. Link loss falls from a peak of around 22dB to roughly 4dB, while signal integrity increases by about 64 times. The SN6810 packs 128 800Gb/s ports into a 2U liquid-cooled chassis, delivering a total switching capacity of 102.4Tb/s; the larger SN6800 integrates four switching chips to achieve 409.6Tb/s.

For the optical module industry, this day carries weight comparable to the shift from feature phones to smartphones. Yet most reports have mischaracterized the narrative—CPO does not mean “optical modules are dying”; it means the “pluggable” form factor within optical modules is dying. Light will still be light, fiber will still be fiber; only the location of the optical engine has shifted from the switch panel to directly beside the switching chip. What is truly being rewritten is the entire supply chain division of labor built around the “pluggable” paradigm.

Below, we break down the physics-based first principles, performance metrics, and three layers of the supply chain. By the end, you will see who risks being pushed out of the game and who has secured greater leverage.

Why Must Optical Engines Be Soldered Directly to Chips?

Let us start with the underlying physics: the catalyst for this shift is that electrical signals can no longer travel efficiently over PCB traces.

In traditional switches, the signal path runs as follows: the switching ASIC generates high-speed electrical signals, which travel along internal PCB traces to the front-panel pluggable optical modules, where they are converted to optical signals before entering the fiber. This PCB trace segment spans roughly tens of centimeters.

At lower speeds, tens of centimeters were inconsequential. But once rates reach 200Gb/s per lane (using PAM4 encoding, pushing Nyquist frequencies into the 50GHz range), signal loss on PCBs escalates exponentially—dielectric loss, skin effect, and reflections all degrade sharply with frequency. NVIDIA’s test results show that link loss in traditional setups peaks at approximately 22dB. By placing the optical engine directly adjacent to the switching ASIC, this electrical transmission path shrinks from tens of centimeters to mere millimeters, dropping loss to around 4dB.

What does the gap between 22dB and 4dB actually mean? Every 3dB reduction doubles signal power. An 18dB improvement translates to roughly a 64-fold increase in signal integrity. This is not a marginal process optimization; it is an architectural reconstruction. The old model was “electrical signals run tens of centimeters, then convert to light.” The new model is “electrical signals convert to light just a few steps after leaving the chip.”

Fundamentally, CPO moves the “electro-optical conversion” step from the panel edge to right next to the chip. That single shift eliminates the loss across the entire high-frequency PCB trace and removes the entire mechanical structure designed solely to enable pluggability—gold-finger connectors, removable housings, panel cutouts, and associated thermal designs.

Here lies a counterintuitive point: a significant portion of the cost and reliability degradation in pluggable optical modules is not spent on “transmitting light,” but on “enabling removal.” Contact resistance of gold fingers, mechanical tolerances of connectors, insertion/extraction lifespan and repeatability, and panel heat dissipation—all these losses stem directly from the pluggable form factor. CPO eliminates them entirely.

The figures for laser reduction, power drop, and MTBF improvement each follow distinct logic.

Laser counts drop fourfold due to externalized light sources. In traditional setups, every pluggable transceiver houses its own laser, resulting in panels densely packed with dozens of lasers for a 128-port switch. CPO adopts an External Laser Source (ELS) module for centralized illumination, allowing one external source to replace multiple transceivers’ lasers. Lasers are the most expensive, delicate, and failure-prone components in optical modules. Cutting their usage by three-quarters simultaneously improves both cost and reliability.

Power consumption drops fivefold due to shortened electrical paths. The closer electro-optical conversion is to the switching chip, the shorter high-frequency electrical signals travel on the PCB, and the lower the drive power required. Traditional setups demand high-current drivers at the panel end; CPO drastically reduces both the distance and intensity of this drive. Lower power eases thermal pressure, which in turn feeds back into MTBF.

The tenfold MTBF improvement is the result, not the cause, of the previous two points: fewer lasers, elimination of mechanical failure points like pluggable connectors, and reduced thermal stress naturally elevate reliability. When NVIDIA stated that “fewer lasers and failure-prone components” are used, the “failure-prone components” primarily refer to the gold-finger connector assemblies.

Viewed together, these three metrics form a clear causal chain: externalizing light sources cuts laser volume → shortening electrical paths slashes drive power → simultaneous reduction of failure points and heat sources pushes reliability up an order of magnitude. These are not three independent selling points; they are three facets of a single architectural shift.

This also explains why market rumors repeatedly claimed “CPO would face delays.” CPO must simultaneously clear three hurdles: silicon photonics manufacturing yield, external laser source coupling, and wafer-level packaging testing. Any bottleneck would push back mass production. However, NVIDIA’s August 14 mass production announcement was immediately followed by confirmation from two major optical communication giants, Lumentum and Coherent, that CPO progress remains on schedule. Coherent noted that rack-to-rack scale-out CPO will begin contributing revenue in the second half of this year, while Lumentum confirmed scale-up products will enter mass production and shipment in the second half of 2027. The rumors have been debunked.

Who Gets Kicked Off the Table, Who Gains Leverage

This is the core of the discussion.

CPO does not eliminate optical modules; it redistributes value across the supply chain. To assess who wins and who loses, apply a single criterion: Are you making a living from “pluggable” integration, or from “optical” components?

Those reliant on “pluggable” face direct disruption. Companies whose core value lies in pluggable optical module assembly, standardized packaging, and mating interfaces bear the brunt of the shock. NVIDIA is not merely “procuring” optical modules here; it is orchestrating a custom-built optical interconnect supply chain tailored to its architecture. The five suppliers explicitly named in its announcement reveal this: TSMC handles silicon photonics manufacturing, Foxconn manages system assembly, Lumentum supplies laser chips, SPIL provides packaging and testing, and TFC manufactures laser module sub-assemblies. Notably absent from this list are traditional pluggable optical module assemblers.

As the world’s largest optical module buyer, purchasing roughly five million units annually, NVIDIA’s decision to build its own supply chain and retain definition rights over the optical engine shifts bargaining power away from module manufacturers and toward the platform provider. This goes beyond “NVIDIA taking optical module business”; it represents a transfer of definition rights over the optical interconnect layer itself.

Those focused on “optical” components benefit. Passive devices, laser chips, and packaging/testing volumes do not decline in CPO architectures—they increase. TFC’s FAU fiber arrays and ELS external laser sources are mandatory for CPO, directly landing on NVIDIA’s official supplier list. Silicon photonics processes, external light sources, and wafer-level packaging represent the true value-add and new technical barriers in CPO.

A clearer picture emerges as layered coexistence: CPO captures the densest, most critical short-reach interconnects within racks and between switches, while pluggable optical modules retreat to mid-to-long-distance applications and building-to-building links requiring flexible maintenance. The day when pluggable modules become entirely obsolete is far off, but their primary battlefield is contracting—from “high-speed core interconnect” to “peripheral elastic interconnect.”

Who Stands Where in This Chain?

Mapping the supply chain layers to specific companies reveals divergent positions.

TFC (300394.SZ) holds the clearest position among A-share listed firms. It is the sole mainland Chinese company on NVIDIA’s official supplier list, producing laser module sub-assemblies. With its 1.6T optical engine already in mass production and active development of CPO-required FAU and ELS products, TFC benefits from a fundamental structural shift: passive devices and optical engine sub-assemblies see higher utilization in CPO than in the pluggable era. This advantage stems not from “riding NVIDIA’s coattails,” but from its products occupying the highest-value segments of the CPO stack.

The international suppliers each hold distinct roles. TSMC (2330.TW) provides the silicon photonics manufacturing foundation; Foxconn (2317.TW) handles system integration and assembly; Lumentum (LITE) supplies external laser sources and chips, representing one of the highest technical-barrier segments in CPO; and SPIL (3711.TW, part of ASE Group) performs wafer-level and chip-level packaging and testing, directly benefiting from increased CPO packaging complexity.

Traditional optical module leaders face a bifurcated outlook. Leaders like Zhongji Innolight (300308.SZ) and Eoptolink (300502.SZ) will not see immediate disappearance of short-term orders and revenue—mid-to-long-range interconnects, legacy markets, and customer transition periods remain intact. Their true test lies ahead: once NVIDIA internalizes the highest-speed interconnect layer, the long-term ceiling for pluggable modules is compressed. Their value anchor shifts from “core NVIDIA supplier” to “provider of peripheral interconnects and legacy market transitions.” The valuation logic for these two positions differs fundamentally.

The beneficiary thesis is straightforward but requires directional clarity: CPO reshapes the trend of “bringing optical engines closer to the chip.” Beneficiaries are passive devices, laser chips, silicon photonics processes, and packaging/testing segments adjacent to the chip. Losers are the “pluggable standardization” segment, which sits furthest from the chip. Traditional optical module manufacturers will not collapse overnight; they will be gradually squeezed toward the periphery.

NVIDIA’s announcement formally invalidates the decade-old rules of the optical module industry—“whoever makes it smaller, faster, and cheaper”—for the high-speed interconnect layer inside server racks. The new rule is simple: whoever secures positioning in silicon photonics manufacturing, external light sources, and packaging integration earns a seat at NVIDIA’s table. And this time, NVIDIA is the dealer.

For manufacturers that have spent a decade perfecting pluggable optical modules, the hardest question is not “when will CPO scale?” but a more pressing concern: When the world’s largest optical module buyer begins soldering optical engines directly onto its own boards, who will be your next customer?