Nvidia's Spectrum-X switches ramping to 100,000 units by 2027

Nvidia's push into co-packaged optics reflects a structural shift in AI infrastructure economics. As AI clusters scale, the interconnect and networking layer—not just GPU compute—becomes a bottleneck that constrains overall system performance and efficiency. By integrating optical switching more tightly with its own networking silicon, Nvidia is extending its platform control from chips into the fabric that connects them. This move also signals Nvidia's strategy to capture a larger share of hyperscaler spending beyond GPUs themselves.

The timeline matters. Spectrum-X CPO entered mass production in August, and GF Securities projects a 6.7× increase in shipments year-over-year (15,000 to 100,000 units). That ramp depends on supplier readiness—TSMC is already expanding its testing and inspection footprint—and on customer adoption at scale. The next inflection point arrives in 2027, when Nvidia's NVL576-based systems using CPO are expected to begin adoption in the second half of that year.

Equally significant is the possibility that Nvidia's interconnect technology—particularly NVLink Fusion—could appear in custom AI infrastructure outside Nvidia's own GPU platforms. GF Securities expects Amazon's Trainium 4 chip to potentially adopt NVLink Fusion, which would extend Nvidia's technology footprint into competitors' silicon. Such a move would benefit optical and semiconductor component suppliers and reinforce Nvidia's position as an AI infrastructure vendor, not just a GPU maker.