CPO is an emerging market at the intersection of optical transceivers and switching ASICs, driven almost entirely by AI data center bandwidth demands. The TAM is small today but growing exponentially as hyperscalers transition from 400G pluggable optics to 1.6T and 3.2T CPO-based switches. The base case assumes CPO achieves meaningful penetration in AI switch fabric by 2028 and scales with continued AI infrastructure buildout through 2030. Bear case assumes thermal and replaceability challenges delay mass adoption by 2–3 years, keeping pluggable dominant longer. Bull case assumes CPO penetrates not just AI switches but also front-panel optics and DCI applications, materially expanding the serviceable market. Key risk: silicon photonics yield and packaging costs must fall another 40–60% to be cost-competitive at scale — progress is happening but not guaranteed.
“Hyperscaler AI cluster bandwidth requirements are outpacing what pluggable optics can deliver. Co-packaged optics reduce power consumption 30–50% and eliminate SerDes bottlenecks at 1.6T and 3.2T switch speeds — making CPO mandatory for next-generation AI fabric switches.”
“Intel's Silicon Photonics division secured CPO design wins at two hyperscalers for 2027 AI switch deployments. 1.6T CPO modules entering qualification; volume production ramp targeted for H2 2027. CPO is now on the critical path for 400G-per-lane AI data center networking.”
“Marvell's CPO chiplet platform progressing through customer qualification at two Tier-1 hyperscalers. Electro-optic integration with switching ASICs on track. Management guided CPO as a $1B+ revenue opportunity by FY28.”
“Raman amplification and silicon photonics integration maturing faster than consensus expected. Yield improvements on silicon photonic waveguides now commercially viable — key gating factor for CPO cost-competitiveness against pluggable modules is resolving.”
“CPO thermal management and field replaceability remain engineering challenges. Hyperscalers are cautious about lock-in and serviceability — pluggable optics remain dominant through 2027 while CPO qualifications run in parallel. Mass adoption likely 2028+.”
Next-generation AI clusters (NVIDIA GB300, Google TPU v6) require 51.2 Tbps switch capacity. Pluggable optics at this speed are physically impractical — insertion loss, power draw, and PCB area make them non-viable. CPO is the only known solution that meets the bandwidth, power, and density requirements simultaneously. Every AI switch at 1.6T+ is a potential CPO socket.
Stargate's 1M+ GPU clusters require a network fabric operating at unprecedented scale. At this density, inter-GPU bandwidth dominates total cluster cost and power. CPO-based switches reduce networking power by 30–50%, a critical consideration for facilities constrained by power and cooling. If Stargate standardizes on CPO for AI spine switches, it establishes the reference architecture for the entire industry.
TSMC, GlobalFoundries, and Intel Foundry are investing in silicon photonics process nodes to enable high-volume CPO production. A viable foundry ecosystem is a prerequisite for CPO cost parity with pluggables. TSMC's 300mm silicon photonics capacity expansion is the most critical gating factor — if it hits schedule, CPO cost curves drop ahead of demand.