NVIDIA Spectrum-X Ethernet Silicon Photonics Technology Enters Full Mass Production Based on CPO
Alina Collins
Nvidia announced full production of its CPO-based Spectrum-X silicon-photonics switch, delivering 5× better energy efficiency than legacy transceivers and clearing the network bottleneck for million-GPU AI factories.
What problem does this switch actually solve?
Scaling an AI factory means wiring thousands of GPUs together. The network is the highway between them — and legacy Ethernet hits a wall on power, reliability, and deployment speed.
Spectrum-X uses CPO — co-packaged optics, meaning the optical link is built right into the chip package instead of bolted on as a separate transceiver. That cuts both power draw and failure points in one move.
This means → every watt saved on networking is a watt freed for GPU compute. The larger the cluster, the bigger the saving.
How should we read the performance numbers?
Compared with legacy transceiver networks, Spectrum-X silicon photonics delivers three gains: 5× energy efficiency, 5× AI uptime, and 1.3× faster deployment.
In plain terms = the same electricity bill buys more compute, machines go down less often, and a new AI factory's network comes online faster.
This reflects Nvidia shifting the competitive frontier from "single-card horsepower" to "whole-factory network efficiency" — whoever runs the most efficient, most stable network can build the biggest cluster.
Who is building this system for Nvidia?
Production runs on a Taiwan-based supply chain: TSMC fabricates the silicon-photonics chips, SPIL handles chip-level packaging and opto-electronic integration, TFC supplies laser chips rated for round-the-clock AI workloads, and Foxconn assembles the finished rack-scale switch.
This means → CPO is not a lab demo. It has cleared the full "design → fab → packaging → system assembly" pipeline for volume production.
Before shipping, every switch goes through unboxing, installation, and power-on verification inside Nvidia's own AI factory — quality gate before the loading dock.
Who gets it first?
CoreWeave, Lambda, and Oracle Cloud Infrastructure are the first adopters of the Spectrum-X silicon-photonics switch.
All three run large-scale GPU cloud services. Choosing them as launch customers signals the system targets million-GPU-class mega-clusters, not mid-size deployments.
Put simply = Nvidia sells GPUs through Vera Rubin and networking through Spectrum-X, locking the AI factory — from chip to cable — inside its own ecosystem.
Content is for reference only, not financial advice.