Nvidia's Next AI Infrastructure Bet Is Scaling Fast
I'm LongbridgeAI, I can summarize articles.Nvidia has entered mass production of its Spectrum-X co-packaged optics (CPO) switches, addressing networking bottlenecks in large AI clusters. GF Securities projects shipments will surge from 15,000 units in 2026 to 100,000 in 2027, supported by TSMC's capacity expansion. Amazon may adopt similar near-packaged optics in its Trainium 4 systems using Nvidia's NVLink Fusion, potentially benefiting optical and semiconductor suppliers like Lumentum and Marvell.
Nvidia is moving another critical piece of its AI infrastructure strategy into mass production, with its Spectrum-X co-packaged optics, or CPO, switches now ramping for scale-out deployments. The shift matters because networking is becoming an increasingly important bottleneck as AI clusters grow larger, and GF Securities expects Nvidia's CPO switch shipments to surge from 15,000 units in 2026 to 100,000 in 2027.
CPO integrates optical components more closely with networking silicon, helping reduce power consumption and improve bandwidth as data moves across massive AI systems.
Nvidia disclosed on Aug. 14 that Spectrum-X CPO switches had entered mass production. GF Securities analyst Alicia Xia said supplier preparations are accelerating alongside that ramp.
To support robust capacity ramp, TSMC has substantially expanded its CPO testing and inspection equipment footprint, notably adding Insertion 2/3 capacity, Xia said.
The next expansion could arrive in 2027. Nvidia's NVL576-based scale-up systems using CPO or near-packaged optics are expected to begin adoption in the second half of that year.
Amazon (AMZN) could broaden the trend further. GF Securities expects Trainium 4 to include multiple configurations, with two likely using near-packaged optics. One would rely on PCIe Gen7, while another is expected to use UALink switching.
The bank also expects Trainium 4 to adopt Nvidia's NVLink Fusion, potentially extending Nvidia's technology footprint into custom AI infrastructure outside its own GPU platforms.
That ecosystem shift could benefit optical and semiconductor suppliers including Lumentum , Coherent , Semtech , Marvell and Tower Semiconductor .
Investor Takeaway
The key metric is whether the projected 15,000-to-100,000 unit shipment ramp actually materializes.
A successful rollout would reinforce Nvidia's push to capture more AI-system spending beyond GPUs, including networking and interconnects.
Investors should also watch TSMC capacity expansion, supplier readiness and adoption of NVLink Fusion by hyperscalers such as Amazon.
The bull case strengthens if CPO becomes standard in larger AI clusters and Nvidia can bundle networking more tightly with compute. Delays in optical integration, slower AI infrastructure spending or competing interconnect standards would weaken that opportunity.