AnswersPublished23 sources
CIOE 2026: AI Clusters Push Optics Closer to Compute
CIOE 2026’s clearest signal is architectural: 800G optics are commercial at scale, 1.6T is entering a major ramp, and 3.2T and faster designs are largely still in sampling or validation. The important question is no longer simply which module has the highest headline speed, but which combination of pluggable optics,...
Published byEdited with GPT-5.6 TerraImages generated with GPT Image 2
Research answer
Key takeaways
- CIOE 2026’s clearest signal is architectural: 800G optics are commercial at scale, 1.6T is entering a major ramp, and 3.2T and faster designs are largely still in sampling or validation.
- The important question is no longer simply which module has the highest headline speed, but which combination of pluggable optics, LPO, NPO, CPO and short reach copper best balances power, reach, serviceability and de...
- LightCounting forecasts 65% growth for the Ethernet transceiver market in 2026; CIOE cites a $14.6 billion combined 800G and 1.6T market, or roughly 64% of optical module revenue.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What will China’s 27th China International Optoelectronic Expo (CIOE), opening September 9, 2026, in Shenzhen, reveal about the AI-driven up. Article summary: CIOE 2026 shows that AI networking is moving from a simple “faster pluggable transceiver” cycle to an architecture change: 800G is the shipping-volume product, 1.6T is entering broader deployment, and 3.2T-plus optical e. Topic tags: general, general web, user generated, news, documentation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, water
AI data-center networking is reaching a point where faster front-panel modules alone are not enough. At CIOE 2026 in Shenzhen, the practical market transition was clear: 800G is established at scale, 1.6T is moving into broad deployment, and 3.2T, 6.4T and 12.8T products are primarily road-map, sampling and customer-validation technologies. 2
10
12
That matters because the AI optical up-cycle is fundamentally an interconnect-power story. As bandwidth rises, designers need to reduce the distance that the fastest electrical signals travel before becoming light.
The headline: 800G is the base, 1.6T is the ramp
CIOE says 800G modules have achieved large-scale production and describes 2026 as a key year for the large-scale ramp of 1.6T modules. It also says 3.2T and higher-speed products remain in sampling and customer validation. 2
12
This distinction is more useful than a show-floor list of ever-larger speeds:
- 800G is the commercial foundation for high-volume AI and cloud interconnects.
- 1.6T is the near-term deployment wave, with suppliers building production capacity and customers qualifying platforms.
- 3.2T and beyond show where system design is headed, but they should not be mistaken for current volume demand.
2
10
CIOE cited a LightCounting forecast that 800G and 1.6T modules together could generate about $14.6 billion in 2026, roughly 64% of the optical-module market. It is a forecast rather than a measure of completed sales, but it illustrates how concentrated demand has become at the highest deployed data-center speeds. 4
Why the electrical channel is becoming the constraint
Traditional pluggable optics sit at the front panel. That arrangement means a high-speed signal must travel from a switch or compute ASIC across PCB traces, connectors and, in some designs, additional conditioning components before it reaches the optical module.
At ever-higher lane rates, system designers face a harder trade-off among insertion loss, signal integrity, equalization, thermal headroom and power consumption. CIOE describes NPO and CPO as approaches that reduce high-speed electrical-signal loss by bringing optical engines into close proximity with switching silicon, and identifies them as important paths beyond the 1.6T bandwidth threshold. 12
The result is not a wholesale replacement of pluggables. It is a staged architecture shift in which each option fits a different operating point.
LPO, NPO and CPO: three ways to shorten the path
LPO: preserve pluggability, reduce module power
Linear pluggable optics (LPO) retain a pluggable module but use a linear architecture rather than a module DSP. The attraction is lower power and latency for short-reach applications; the trade-off is that the host electrical channel must be exceptionally clean. CIOE says LPO is gaining acceptance in short-distance interconnects because of its low-power, low-latency profile. 12
NPO: move a detachable optical engine closer to the ASIC
Near-packaged optics (NPO) place an optical engine close to the switch ASIC, sharply reducing the electrical distance while retaining a more modular approach than full co-packaging. CIOE exhibitors showed NPO engines at 3.2T and 6.4T, including a 6.4T NPO engine from Accelink listed in the event guide. 1
13
CPO: minimize electrical reach, accept a tougher operations problem
Co-packaged optics (CPO) integrates optics with, or immediately beside, the switch package. It offers the most aggressive reduction in high-speed electrical reach, but thermal design, manufacturing yield, repair and field service become more complicated. This is why CPO is best understood as a system-architecture option rather than a universal successor to pluggable transceivers.
The likely outcome is coexistence: conventional DSP-based pluggables, LPO, NPO and CPO will serve different reach, power, serviceability and operational-risk requirements.
What Chinese suppliers are signaling
CIOE’s product announcements show that Chinese suppliers are advancing from presentations of future concepts to a mix of shipping products and increasingly concrete next-generation engines.
- Eoptolink has begun small-scale 1.6T shipments, is identified by CIOE as a major 800G silicon-photonics producer and is developing higher-speed NPO and XPO products. The show floor also included 12.8T XPO samples.
1
- Accelink displayed a 6.4T silicon-photonics single-mode NPO optical engine, according to exhibition reporting and the CIOE guide.
- HGTECH is cited by CIOE as having bulk supply of 1.6T modules, while the broader show featured silicon-photonics NPO and CPO demonstrations aimed at higher speeds.
5
- Silicon photonics and thin-film lithium niobate both appeared in supplier road maps for modules ranging from 400G to 1.6T. These are complementary technology paths toward denser integration and lower-power high-speed transmission, not evidence that one platform will immediately displace all others.
The key diligence point is product status. A public demonstration, a sample, customer validation and volume shipment are materially different milestones. CIOE’s own framing places most 3.2T-and-faster products in the earlier stages of that sequence. 2
Copper still matters—inside tightly controlled reaches
The move toward optics closer to compute does not mean copper is disappearing. It means copper is being used where its short-reach economics and power profile remain compelling.
NVIDIA describes the Vera Rubin NVL72 rack as using a modular copper-cable spine/backplane made of up to four preintegrated cable cartridges that connect the trays. 49 Luxshare is developing 224G and 448G electrical-interconnect technologies, including CPC, NPC, cable-cartridge and high-speed-backplane products. It says its 448G CPC PCIe DAC technology can reduce overall system power consumption by more than 30%; that is a company claim, not an independently verified system-wide result.
52
This division of labor is central to the AI-rack design trend: keep copper extremely short and engineered, then use optics where distance, aggregate bandwidth or power make electrical links less practical.
Fibre is also becoming an AI-infrastructure issue
The CIOE story extends beyond the rack. As AI workloads spread across buildings, campuses and metro areas, fibre performance matters for data-center interconnect as well.
YOFC says it has produced a 91.2-km continuous hollow-core fibre and deployed hollow-core fibre across more than 10 commercial and pilot projects. 38 Separate reporting says the company has delivered more than 10,000 fibre-kilometres and reported attenuation of 0.04 dB/km.
34 At CIOE, Hengtong introduced hollow-core anti-resonant fibre.
12
Hollow-core fibre is not a substitute for the short electrical links inside a rack. Its relevance is different: it is a developing option for lower-loss, lower-latency links across longer data-center and metro distances.
What to watch after CIOE
The most meaningful indicators will be less glamorous than a new terabit number on a booth sign:
- 1.6T shipment and qualification progress: Is deployment widening beyond early adopters?
- Power per bit: Are LPO, NPO and CPO delivering a material system-level advantage, not just an optical-engine claim?
- Operational practicality: Can near- and co-packaged designs meet thermal, service and repair requirements?
- The electrical-channel budget: How much high-speed copper trace can a given switch or compute platform tolerate before optics must move closer?
- Longer-reach AI connectivity: Does hollow-core fibre progress from field deployments to repeatable, economic data-center-interconnect use?
LightCounting forecasts 65% growth in the Ethernet-transceiver market for 2026, following strong growth in the preceding two years. 23 But CIOE’s deeper message is not simply that AI will buy more optics. It is that AI is forcing a redesign of where the electrical link ends and the optical link begins.
For now, 800G is the commercial anchor and 1.6T is the next major ramp. The surge of NPO, CPO and XPO demonstrations points to the next constraint: when very fast electrical channels cannot meet loss, power and thermal budgets, the optics have to move closer to the compute and switching silicon.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "CIOE 2026: AI Clusters Push Optics Closer to Compute"?
CIOE 2026’s clearest signal is architectural: 800G optics are commercial at scale, 1.6T is entering a major ramp, and 3.2T and faster designs are largely still in sampling or validation.
What are the key points to validate first?
CIOE 2026’s clearest signal is architectural: 800G optics are commercial at scale, 1.6T is entering a major ramp, and 3.2T and faster designs are largely still in sampling or validation. The important question is no longer simply which module has the highest headline speed, but which combination of pluggable optics, LPO, NPO, CPO and short reach copper best balances power, reach, serviceability and de...
What should I do next in practice?
LightCounting forecasts 65% growth for the Ethernet transceiver market in 2026; CIOE cites a $14.6 billion combined 800G and 1.6T market, or roughly 64% of optical module revenue.
