Key Takeaways:
Co-packaged optics have moved beyond the viability debate. The question now is which supply chain players will capture value as the technology scales from prototypes to hyperscale AI factories by 2028.
The 2026 COMPUTEX trade show in Taipei marked a turning point for co-packaged optics, or CPO, a technology that embeds optical transceivers directly alongside compute silicon to overcome the bandwidth and power limits of traditional electrical interconnects. Discussions shifted from whether CPO works to how to deploy it at rack scale, according to a Citi research note covering the event.
"CPO has reached an inflection point where the industry conversation has moved from individual optical components to complete rack-level deployment solutions," the Citi analysts wrote. "The technology feasibility debate is largely settled."
Nvidia provided the show's most concrete demonstration of CPO's system-level potential. The company unveiled Kyber, its next-generation rack-scale computing architecture built around a blade-and-midplane design that doubles the NVLink interconnect scale to 144 graphics processing units. Each Kyber chassis holds as many as 18 vertical compute blades, with NVLink switch blades connecting directly through a cable-free midplane that serves as the central hub for high-speed signal transmission, power distribution and serviceability.
The bandwidth requirements are staggering. A Kyber system based on Nvidia's upcoming Rubin Ultra architecture delivers 14.4 terabits per second of scaling bandwidth per logical GPU, requiring 72 NVLink switch chips to achieve full 144-GPU interconnectivity. The midplane — a printed circuit board that was historically a passive structural component — has become a mission-critical element of the system's electrical and optical signal integrity.
The midplane becomes the motherboard
Kyber's architecture signals a structural shift in where value accrues in AI infrastructure. The network fabric, midplane PCB, connectors, power management and thermal systems now carry strategic importance on par with the GPU itself, according to Citi. That expands the addressable market for suppliers of high-speed connectors, advanced PCB laminates and liquid cooling solutions — components that previously commanded lower margins than compute silicon.
Nvidia's Jensen Huang recognized Ayar Labs, a silicon photonics startup, as an optical interconnect partner during his GTC Taipei keynote. Ayar Labs demonstrated its TeraPHY optical engine alongside Wiwynn, a rack-scale server vendor, and GUC (Global Unichip), a design services firm, in a live CPO integration at COMPUTEX. The demo addressed practical deployment challenges including cooling of external laser source modules and high-density fiber routing — two longstanding reliability concerns that had slowed CPO adoption.
MediaTek is also deepening its involvement. The Taiwanese chip designer plans to co-develop CPO solutions with partners within two to three years, contributing expertise in advanced I/O, chiplet integration and application-specific integrated circuit design, while Ayar Labs provides the optical engine technology.
Investment implications for the supply chain
Citi's report frames 2028 to 2030 as the window for hyperscale CPO deployment, giving investors a concrete timeline for when the technology could meaningfully affect semiconductor and networking equipment revenue. The entry of design service firms such as GUC and MediaTek into the CPO ecosystem suggests the technology is moving from specialist optical companies into the broader semiconductor design infrastructure, which typically accelerates standardization and cost reduction.
For investors, the key question is which parts of the supply chain capture the most value. Nvidia's Kyber architecture suggests that midplane PCB makers, connector suppliers and thermal management vendors will see their content per rack rise significantly as optical interconnects replace copper. Companies with exposure to advanced PCB laminates, high-speed connectors and liquid cooling — such as Taiwanese supply chain vendors — stand to benefit as hyperscalers begin qualifying CPO-based systems for production.
Nvidia shares trade at roughly 35 times forward earnings. If CPO adoption follows the trajectory Citi outlines, the total addressable market for optical interconnects in AI data centers could expand well beyond current estimates, though no specific revenue projections have been disclosed by any of the companies involved.
This article is for informational purposes only and does not constitute investment advice.
