CIOE 2026 Puts the AI Optical-Interconnect Upgrade in Focus

CIOE 2026 is a useful snapshot of how AI infrastructure is reshaping optical communications. Taking place September 9–11 at the Shenzhen World Exhibition & Convention Center, the 27th China International Optoelectronic Exposition is pairing its exhibition with sessions on AI compute networks, high-speed optical interconnects and optical transmission. 3613

The central theme is clear: as AI systems connect ever-larger numbers of accelerators and switches, networking must deliver more bandwidth while containing power use and latency. That is driving a transition from established 400G links toward broad 800G deployment, 1.6T modules, and new ways of placing optics closer to the networking silicon. 316

From 400G to 800G and 1.6T

CIOE’s program presents 800G and 1.6T as technologies moving toward commercialization, while its AI-cluster forum explicitly links large intelligent-computing clusters to a move to 1.6T and beyond. 616

That does not mean every data-center connection immediately moves to the highest available speed. Different reaches, switch designs and upgrade cycles will continue to support a mix of 400G, 800G and 1.6T products. But the direction of travel is toward higher bandwidth density as more data must move among compute, storage and switching systems.

LightCounting’s April forecast calls for 65% growth in the Ethernet-transceiver market in 2026. Its earlier AI-cluster outlook estimated a $26 billion Ethernet-optics market for 2026, up from $16.5 billion in 2025. 2123

A widely reported LightCounting forecast further places combined 800G and 1.6T module revenue at about $14.6 billion, roughly 64% of the overall 2026 optical-module market. That is a forecast rather than realized market demand, and it should be read accordingly. 1922

Why module speed alone is not enough

At these data rates, short electrical paths and power efficiency become as important as the optical module’s headline speed. CIOE’s materials therefore emphasize several architectures rather than a single successor to conventional pluggable optics.

• Pluggable optics remain the mainstream choice for scale-out links between systems because of their standardization and compatibility advantages. 16
• LPO (linear-drive pluggable optics) removes the DSP from the module, an approach CIOE describes as especially suited to low-power, short-reach intra-rack interconnects. 8
• NPO (near-packaged optics) places optics nearer to the switching package. CIOE describes it as highly integrated and simpler to deploy than some alternatives; its AI-cluster forum suggests NPO may reach end deployments ahead of CPO because of ecosystem compatibility. 816
• CPO (co-packaged optics) integrates optical engines with compute or switch ASICs on a common packaging substrate, shortening the electrical connection between electronics and optics. CIOE describes the architecture as reducing interconnection length to about 1–2 cm and targeting lower latency, lower power use and higher bandwidth density. 5

The practical takeaway is that pluggables are not disappearing. Instead, the market is diversifying: standardized modules remain important where interoperability and serviceability matter, while NPO and CPO target the densest, most power-constrained parts of the AI network. 16

What CIOE says about the Chinese supply chain

The Information and Communication Expo is designed to cover the optical-communications chain from chips and components through modules, equipment, semiconductor materials and solutions for AI data centers. 39

Organizers have highlighted Accelink’s 800G and 1.6T ecosystem, while CIOE’s materials also identify Eoptolink among companies working on high-speed optical-module products and next-generation packaging roadmaps. 18 The AI optical-interconnect forum agenda includes Accelink and HGTECH-related optical businesses among the participating companies, alongside device, chip, equipment and cloud players. 9

These appearances are meaningful as evidence of active product and technology development, but they should not be confused with proof of broad production volumes, customer wins or long-term market share. Those require separate shipment and deployment evidence.

Copper, photonics and fiber will coexist

The AI networking upgrade is not an “optics replaces copper” story. CIOE’s show materials include high-speed connectors alongside modules, optical engines and switching systems, reflecting the need for different interconnect media at different reaches. 315

Likewise, module and packaging advances are only one part of the network. CIOE’s transmission agenda includes hollow-core fiber and space-division multiplexing as emerging options for high-capacity, latency-sensitive optical transport. 6 The expo’s forum materials say hollow-core fiber is progressing toward larger-scale deployment, but this remains a different and earlier technology transition from the widespread adoption of high-speed pluggable modules. 2

The constraint investors and buyers should watch

Demand is not the only determinant of optical-module growth. LightCounting cautions that shortages of XPUs and switch ASICs could limit the pace of AI-cluster expansion in 2026, even amid strong demand for connectivity. 20

That caveat matters. CIOE’s most important message is not that every emerging optical architecture has already won, but that AI infrastructure is forcing the entire interconnect stack to evolve. The competitive question is increasingly how efficiently suppliers can deliver bandwidth—from electrical traces and copper links to photonic packaging, optical modules, switching and fiber—at the scale required by AI clusters.