Nvidia's Spectrum-X Ethernet Photonics switches — the first volume-produced 200G/lane co-packaged optics system — are now in full mass production, doubling per-switch bandwidth to 102.4 Tb/s while cutting laser count by 75 percent.
返回
Nvidia CPO switches enter mass production, cutting laser count 75%
Nvidia's Spectrum-X Ethernet Photonics switches entered full mass production, the first volume-produced 200G/lane co-packaged optics (CPO) system. Per-switch-chip bandwidth doubles to 102.4 Tb/s from the Blackwell-generation Spectrum-4's 51.2 Tb/s, with laser count down 75 percent and power consumption cut fivefold.
"The bandwidth required for Scale-up, roughly speaking, can reach about 10x that of Scale-out, while also demanding extremely low latency," Gilad Shainer, Nvidia's senior vice president of networking, said. The Scale-up network — the interconnect linking GPUs within a compute domain — is where the AI server bottleneck has shifted from raw compute to high-speed data movement.
The CPO architecture integrates optical engines directly alongside the switch ASIC, eliminating the DSP and shortening electrical signal paths from 15-30 centimeters of PCB trace to millimeters. Optical loss drops from roughly 22 dB to about 4 dB, signal integrity improves 64x, and mean time between failures extends tenfold. The system ships in two form factors: a 2U SN6810 with 128 800 Gb/s ports and a 5U SN6800 with 512 800 Gb/s ports at 409.6 Tb/s aggregate bandwidth. CoreWeave, Lambda, and Oracle are the first-wave customers, with production ramping from May to July 2026.
The mass production milestone shifts the competitive battleground from rack-to-rack interconnects into GPU core interconnects, cementing the "optical in, copper out" trend. For the supply chain, it means TSMC, Foxconn, and ASE gain silicon photonics and advanced packaging revenue, while upstream laser and InP suppliers face a demand inflection as AI data center buildouts accelerate.
CPO's Yield Math and the NPO Alternative
CPO has historically struggled with manufacturing yield. A soldered optical engine cannot be reworked — removing a reflowed, underfilled engine requires solder-melt temperatures of 220°C to 260°C millimeters from the ASIC, and sub-micron fiber alignment doesn't survive a second thermal excursion. SemiAnalysis calculated that a hypothetical 32-engine package at 95 percent attach yield per engine lands near 19 percent compound yield. Nvidia designed spare engines into Spectrum-X for redundancy, and Broadcom reported more than 1 million cumulative 400G-equivalent port-hours in Meta testing without a single link flap.
Near-packaged optics (NPO) has emerged as an interim architecture that keeps the optical engine socketable while moving it off the faceplate. Broadcom showed a 3.2T VCSEL-based NPO product line at OFC 2026, and six connector and optics firms formed the Open CPX MSA to standardize a socketed optical-engine interface. LightCounting CEO Vladimir Kozlov projected annual port shipments topping 100 million within five years, against fewer than 1 million co-packaged and near-packaged ports in 2025.
Upstream Optical Chip Demand Accelerates
China Galaxy Securities said the mass production of Spectrum-X Ethernet Photonics for Vera Rubin confirms the value of optical interconnect in new AI data centers, and rising silicon photonics penetration will drive high-speed growth in upstream optical chip demand. Several A-share optical chip companies — Yuanjie Semiconductor, Shijia Photonics, and Focuslight Technologies — have announced capacity expansions and private placements, moving from planning into substantive capital expenditure.
Taiwanese suppliers span the CPO stack: TSMC handles silicon photonics manufacturing, Foxconn assembles the switch systems, SPIL manages packaging and testing, Browave supplies fiber optic passive components, FOCI produces fiber array units, EZconn builds external laser source modules, ShunSin Technology handles advanced packaging, and LandMark Optoelectronics supplies InP light sources with 1.6T products already in mass production.
Nvidia invested $2 billion each in Lumentum and Coherent in March and partnered with Corning in May to expand U.S. optical connectivity capacity tenfold. The next battleground is whether optical interconnect can penetrate GPU core interconnects — the Scale-up domain traditionally dominated by copper. As GPU compute domains expand from single racks to cross-rack configurations, the "optical in, copper out" trend is expected to accelerate.
For investors, the CPO ramp benefits TSMC (2330.TW), Foxconn (2317.TW), and ASE Technology Holding (3711.TW) through advanced packaging and silicon photonics revenue. Lumentum and Coherent gain from external laser source demand, while upstream InP suppliers like LandMark Optoelectronics (3081.TW) stand to capture margin as light source supply becomes a gating factor for volume ramp. The question is whether NPO's socketable architecture extends its run through the end of the decade — SemiAnalysis pushes full CPO scale-up volume to 2028 or 2029 — or whether Nvidia's production ramp compresses that timeline.
This article is for informational purposes only and does not constitute investment advice.
[1] Nvidia CPO switches enter full mass production, AI data centers usher in the 'Silicon Photonics Era'![2] Nvidia's CPO Switches Enter Full Mass Production; TSMC and Foxconn Lead Supply Chain Positioning[3] NVIDIA officially announces full-scale mass production of ...[4] NVIDIA’s CPO switches enter full-scale mass production, ushering AI data centers into the “silicon photonics era”! - Moomoo[5] Near-packaged optics (NPO) gains ground as the industry hedges against CPO's growing pains — analysts say volume for NPO silicon photonics products will extend until the end of the decade - Tom's Hardware
