NVIDIA announced Spectrum-X Photonics in March 2025, alongside Quantum-X Photonics for InfiniBand; by 2026 it calls the Ethernet family Spectrum-X Ethernet Photonics. Two switch models are named: the SN6810, at 102.4 Tb/s (“128 ports of 800Gb/s or 512 ports of 200Gb/s”), and the SN6800, at 409.6 Tb/s (“512 ports of 800Gb/s or 2,048 ports of 200Gb/s”).

On 31 May 2026 NVIDIA described them as “the world’s first co-packaged-optics (CPO)-based switches with 200Gb/s SerDes — now in production”. This entry is listed with its original 2025 date; the production update is the 2026 development.

In a large GPU cluster the optics, not the switch chips, can draw a large share of network power, and every pluggable module is a part that can fail. NVIDIA’s technical blog puts a pluggable interface at “often 30W”, against “as low as 9W” with co-packaged optics.

Removing hundreds of thousands of pluggable modules saves power that can go to GPUs instead, and NVIDIA argues it removes a common failure point.

In a conventional switch the chip sends electrical signals across the circuit board to each front-panel module, and each module has its own signal processor and lasers. With co-packaged optics, small optical engines are mounted in the same package as the switch chip, so the electrical signal travels millimetres instead of tens of centimetres and needs far less power to drive. The lasers are moved out to separate, replaceable laser modules that feed several channels each.

The trade-off is serviceability: a failed optical engine cannot be swapped like a pluggable module. NVIDIA argues the parts are more reliable, which is the claim to watch as deployments grow.

Power & Cooling

Lower network power frees budget for compute. Check the switch’s cooling requirement with the supplier at design time: it is not given in the sources read here.

Networking

Fewer parts to buy, install and clean, but a new design to qualify, and a switch that is replaced as a unit if its optics fail. The cabling plan changes: fibre attaches to the switch, not to modules.

Operations

Spares change from optical modules to laser modules and whole switches; failure handling and vendor support terms need to be agreed before go-live.

1. AnnouncedDescribed publicly; no date for customers to buy it.
2. Planned (current stage)A stated availability window that has not arrived yet.
3. ShippingIn production and reaching early customers or partners.
4. Generally availableOrderable by any customer through normal channels.
5. DeployedRunning in customers’ production environments, confirmed by them or the manufacturer.

In March 2025 NVIDIA said the Ethernet switches were “coming in 2026”, later the second half of 2026. On 31 May 2026 it said they are “now in production” and named CoreWeave, Lambda and OCI among the first adopters. No NVIDIA page read for this entry says they are shipping to customers or generally available.

• Scale-up versus scale-out networkingThe scale-out fabric these switches build.
• RDMA, RoCEv2 and InfiniBandThe Ethernet transport Spectrum-X carries.
• The building blocks: networkWhere optics sit in the network.

1. NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUs(opens in a new tab)NVIDIA Newsroom · 18 March 2025 · Manufacturer
2. Scaling AI Factories with Co-Packaged Optics for Better Power Efficiency(opens in a new tab)NVIDIA Technical Blog · 18 August 2025 · Technical documentation
3. NVIDIA Vera Rubin Ramps Into Full Production(opens in a new tab)NVIDIA Newsroom · 31 May 2026 · Manufacturer

Figures are as the manufacturer states them unless marked otherwise. None has been independently tested by BSS/OSS Academy.