Nvidia Begins Shipping 400 Tbps Spectrum-X CPO Switch at GTC Taipei

"The switch uses Nvidia's most advanced CPO technology," Gilad Shainer, Nvidia's senior vice president of networking, said at the event. He added that the company has already started sending units to partners and expects to expand production capacity in the second half of 2026 .

The NVLink Fusion Ecosystem Opens Up

Nvidia didn’t just ship a product; it also cracked open its proprietary NVLink Fusion interconnect for the first time to outside photonics partners. Both Lightmatter and Ayar Labs announced they had joined the ecosystem, with the goal of making their CPO and near-packaged optics (NPO) products optically and electrically compatible with Nvidia’s SerDes and optical technologies .

Lightmatter: Halving the Fiber

Lightmatter is adapting its bi-directional optical link architecture—the Passage photonic interconnects and Guide laser sources—to Nvidia’s specs. This creates a unified platform for semi-custom AI factories and, crucially, eliminates the need for separate transmit and receive fibers. Lightmatter says the approach cuts fiber and connector requirements by 50%, a dramatic reduction for data centers that can require up to 300 miles of cabling .

The upshot: a customer’s semi-custom XPU can plug directly into Nvidia switch silicon through Lightmatter’s CPO or NPO products, enabling seamless multi-vendor chip connectivity inside the NVLink Fusion fabric .

Ayar Labs: Rack-Scale Optical Fabric

Ayar Labs takes a complementary approach, focusing on connecting thousands of GPUs across multiple racks into a single unified cluster via an optical fabric. Its CPO products target the bandwidth, low latency, and power efficiency that hyperscale AI workloads demand. By joining NVLink Fusion, Ayar Labs makes its technology interoperable with Nvidia’s dominant hardware stack, helping to bring CPO closer to real-world deployment for rack-scale AI infrastructure .

The Strategy: Copper Where You Can, Optics Where You Must

The announcements fit into a longer arc that Nvidia CEO Jensen Huang has been describing since the 2025 GTC. The company’s strategy is pragmatic: use copper interconnects as long as physics allows, then switch to optics when the bandwidth and distance demands outstrip what copper can handle.

"We should use copper as much as we can, for as long as we can, but copper has its limits," Huang said at GTC Taipei. "The right strategy is to scale up with copper as long as you can. After that you scale up further with optics, you scale out with optics and you scale across with optics."

In practice, this means Nvidia’s upcoming Vera Rubin NVL72 and Kyber Ultra NVL144 systems will still rely on copper for their internal scale-up links. Full CPO-based scale-up won’t arrive until the Feynman generation in 2028 . But for the scale-out and scale-across networks that stitch entire AI factories together, the optical transition is happening now.

Nvidia’s combination of shipping a production CPO switch and opening its interconnect ecosystem to third-party photonics startups marks a turning point. The AI industry is moving from debating whether optics will matter to figuring out how quickly they can be deployed.