NVIDIA has announced a major leap in AI data center networking: the company's Spectrum-X Ethernet Photonics switches based on Co-Packaged Optics (CPO) technology have officially entered mass production. The milestone, announced by NVIDIA senior vice president Gilad Shainer, signals that CPO is transitioning from a promising laboratory concept to commercial-scale deployment.

Compared to traditional pluggable optical modules used in data centers today, NVIDIA's CPO architecture delivers dramatic improvements: 5x better optical power efficiency, 5x longer AI application runtime, and 10x increased reliability. These gains address critical bottlenecks as AI clusters scale from thousands to hundreds of thousands of GPUs.

The technology works by integrating the optical engine directly adjacent to the switch ASIC, drastically shortening electrical transmission distances and reducing signal degradation and latency. By using External Laser Sources (ELS), the system reduces total laser count by approximately 75%, cutting both power consumption and potential failure points.

NVIDIA is offering two configurations: the SN6810 single-ASIC switch delivering 102.4 Tb/s total bandwidth, and the SN6800 four-ASIC switch providing 409.6 Tb/s with an internal fiber shuffle mechanism that allows GPU clusters to scale horizontally without additional switching layers.

A formidable supply chain has formed around the rollout: TSMC is manufacturing the silicon photonics chips, ASE Technology (via SPIL) handles chip-level packaging and testing, Lumentum and Tianfu Communications supply laser components, and Foxconn manages system assembly.

According to Bank of America research, while initial shipments are modest — estimated at 14,000 units in 2026 — the true volume inflection is expected around 2028–2029, with shipments projected to reach 190,000 units by 2028. The Bill of Materials breakdown reveals optical engines as the most valuable component at 40–55% of BOM, followed by External Laser Sources at 10–15%.

The mass production announcement represents a paradigm shift in AI infrastructure investment, moving the market's focus from "Can CPO technology work?" to "Who has entered mass production, has stable yields, and what is the dollar content per unit?"

Sources