Citibank Computex Observation: CPO Technology Reaches a Turning Point, Focus Shifts to Rack-Level Deployment

CPO (Co-Packaged Optics) technology is transitioning from proof of concept to mass deployment at ultra-large scale. According to Citibank's report, the 2026 COMPUTEX exhibition marks a critical turning point for CPO development—the industry discussion has shifted from single optical devices to full rack-level deployment solutions, technical feasibility controversies are essentially settled, and the scale of the ecosystem has become the core issue for the next phase. At the exhibition, NVIDIA showcased a Spectrum-6 SPX scalable CPO rack solution based on Hon Hai/Foxconn, Taiwan’s Wiwynn and U.S. startup Ayar Labs jointly demonstrated the integration of optical I/O in real AI infrastructure, while Marvell positioned itself as the backbone network for future optical clusters. Citibank points out that the industry's current challenge is no longer about proving the technology’s feasibility, but how to drive the ecosystem from prototype deployment to mass production at ultra-large scale AI factories from 2028 to 2030. This assessment means that, for investors, the strategic value of supply chain companies related to CPO is accelerating. Design service providers enter the game, accelerating CPO ecosystem integration Wiwynn showcased its CPO collaboration achievements with Ayar Labs and GUC (Global Unichip Corp.) at the exhibition. Ayar Labs’ optical engine and external laser architecture aim to solve the pain points of high-power electrical SerDes links. The demonstration covered liquid-cooled external laser source modules and high-density fiber wiring solutions, directly addressing reliability and thermal challenges that have long restricted CPO adoption. Meanwhile, MediaTek is deepening its cooperation with Ayar Labs. MediaTek can provide expertise in advanced I/O, chiplet integration, and ASIC design, while Ayar Labs contributes optical engine technology. MediaTek plans to launch CPO solutions with its partners in the next two to three years. Citibank believes that the participation of design service companies GUC and MediaTek signifies that the CPO ecosystem is extending from optical specialist manufacturers to a broader semiconductor design field, which will help accelerate technology standardization and mass production. NVIDIA Kyber Architecture Debuts: GPU Interconnect Scale Doubled, System Component Strategic Value Rises NVIDIA officially unveiled its next-generation rack-level computing architecture Kyber at the exhibition. This architecture is based on the AI factory design concept of "blade + midplane", with the core aim of expanding NVLink interconnect scale to 144 GPUs, doubling the current solution. Kyber adopts a vertical compute blade design, with each chassis accommodating up to 18 compute blades. NVLink switch blades connect directly to compute blades via a cableless midplane, making the midplane a core hub for high-speed signal transmission, power distribution, and system maintainability. In terms of bandwidth, Kyber systems based on Rubin Ultra increase scale-out bandwidth to 14.4Tbps per logical GPU. Achieving full interconnect among 144 GPUs requires 72 NVLink switch chips. Citibank notes that the key investment value of Kyber architecture is that the strategic position of networks, midplane PCBs, connectors, power management, and cooling systems has risen to be as important as the GPU itself. This signifies that the value focus of AI infrastructure is shifting from single compute chips to system-level components, potentially opening new growth opportunities for relevant supply chain companies. Risk Warning and Disclaimer The market involves risk, and investments require caution. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions herein suit their particular circumstances. Investments made based on this article are at one’s own risk.