Inside Nvidia and TSMC's Spectrum-X Co-Packaged Optics Switch

TSMC began mass production of the COUPE platform in April 2026, marking the first time a leading foundry manufactured silicon photonics at scale for AI data center applications .

How Co-Packaged Optics Works—and Why It Matters

Traditional network switches use pluggable optical transceivers inserted into the front panel. Data must travel as electrical signals from the switch ASIC, across a printed circuit board, to a separate transceiver where conversion to light finally happens. This distance introduces significant signal loss—often around 22 dB—and requires power-hungry equalization circuitry, generating substantial heat .

Co-packaged optics eliminate this problem. By placing the optical engine in the same package as the switch ASIC, the electrical path is shortened to the substrate level. Light enters the package directly through fiber and is converted with minimal electrical travel. The benefit is stark:

• Signal loss drops from ~22 dB to ~4 dB .
• Power consumption falls to as low as 9 watts per 1.6 Tb/s port, compared to about 30 watts for traditional pluggable optics .

This is not a marginal improvement. Nvidia and TSMC’s approach delivers a 3.5x to 5x power reduction per port versus conventional interconnect methods . At the massive scale of a modern AI factory with hundreds of thousands of GPUs, these savings translate into megawatts of power reclaimed and a far more resilient, cooler-running network fabric.

Spectrum-X Photonics: Performance and Scale

The new Spectrum-X Photonics line pushes the performance boundary. The SN6800 variant delivers up to 409.6 Tb/s of aggregate bandwidth in a single switch, achieved through 512 ports of 800 Gb/s (or denser configurations scaling up to 2,048 ports at 200 Gb/s) . A more compact version, the SN6810, provides 102.4 Tb/s via 128 ports of 800 Gb/s .

These switches are liquid-cooled and designed for the Ethernet-based AI infrastructures that hyperscalers and large enterprises are assembling to train and run generative AI models. Nvidia positions Spectrum-X Photonics as essential infrastructure for connecting clusters of over one million GPUs across massive single-building “AI factories” .

Shipping and Availability

The technology has moved from announcement to shipping. Nvidia formally began delivering the Spectrum-X CPO switch to select partners in early June 2026, as confirmed by Senior Vice President of Networking Gilad Shainer at GTC Taiwan . The earlier Quantum-X InfiniBand photonics switch, which uses the same underlying CPO technology, had already begun shipping in early 2026 .

Broad availability for the Spectrum-X Photonics Ethernet switches from leading infrastructure and system vendors is expected during the second half of 2026, marking the large-scale commercial rollout of this silicon photonics technology .

The collaboration between Nvidia design and TSMC manufacturing has taken silicon photonics from research into a shipping product, offering a practical path to scaling AI without hitting the power wall.