AI Cluster Cross-Data Center Interconnect: Coherent Optical Modules, OCS, and Hollow-Core Fiber — Who Benefits First?
nashnova research
Nvidia's scale-across architecture links multiple data centers into a single compute pool. Nomura maps three optical technology paths — coherent pluggable orders have already hit record levels, while OCS and hollow-core fiber are still proving out deployment at scale — and the sequencing gap defines who benefits now.
Why does AI training need to reach beyond a single data center?
Power and land constraints at individual sites are pushing compute pools across facility boundaries. Nvidia's 2025 scale-across architecture connects data centers in different regions via high-speed optical networks into one unified pool running the same training job.
This means → optical networking has shifted from "internal data-center accessory" to the binding constraint on whether compute can keep scaling.
Nomura splits the technology stack by distance: 500 m to 20 km (synchronous training) calls for coherent pluggables and OCS; 100 km+ (asynchronous training) calls for long-haul coherent and dense wavelength-division multiplexing (DWDM) — a clear tier above the short-reach schemes used inside a single facility.
Coherent pluggables — how real are the orders?
Cisco's Acacia unit disclosed that March-quarter 2026 bookings exceeded $1 billion, a record, and June-quarter bookings again topped $1 billion, driven mainly by 800G ZR/ZR+. In plain terms = this is not a forecast story — revenue is already flowing.
Cignal AI projects the 800G ZR/ZR+ market will grow at a 145% CAGR from 2025 to 2029. Coherent-lite — a lighter coherent format covering 2–20 km — sits squarely in the campus-to-campus and DC-to-DC range that scale-across needs.
Product timeline: Marvell's 1.6T ZR/ZR+ module on a 2 nm coherent DSP is expected to sample in H2 2026; Infinera and Nokia target a 3.2T coherent-lite prototype by mid-2027, with commercial availability in H2 2027.
Chinese vendors are in a follower position. Accelink and Zhongji Innolight's Terahop unit both use Marvell DSPs for their coherent-lite modules. Whether they enter North American hyperscalers' scale-across procurement depends on DSP readiness and customer qualification progress.
Why are both Google and Nvidia backing OCS?
OCS — optical circuit switches that route light signals directly without electrical conversion — can replace traditional electrical switching in long-distance, high-bandwidth scenarios. Nomura cites two new supporting data points.
First, coherent-lite can recover the link-budget margin OCS consumes, keeping the OCS architecture viable through the 2.4T and 3.2T module generations. Second, Nvidia has explicitly endorsed OCS for all-optical switching at the scale-across layer, supporting 100,000-GPU-class clusters.
Market size: Cignal AI estimates the OCS market at roughly $400 million in 2025, rising to over $2.5 billion by 2029 — a ~58% four-year CAGR. Coherent raised its 2030 global OCS guidance from $2 billion to $4 billion at OFC 2026. This reflects a rapid upward revision in industry expectations.
The market is concentrated among Google, Coherent, and Lumentum.
Hollow-core fiber — how far has deployment gone?
Hollow-core fiber — transmitting through air instead of solid glass — offers a latency advantage: FiberHome says its latest generation cuts transmission latency by 31% versus conventional fiber, directly addressing the tight jitter tolerance of synchronous scale-across training.
North American hyperscalers have moved from testing to deployment. As of 2026, AWS has deployed hollow-core fiber in 10 core data centers; Microsoft acquired UK-based Lumenisity in 2022 and plans to deploy 15,000 km across Azure's global network by end of 2026; Google and Meta have also begun testing.
In China, progress centers on YOFC (长飞): its self-developed hollow-core fiber achieves over 20 km continuous single-span transmission with attenuation of 0.05 dB/km, is in volume production, and has been delivered to all three major carriers. China Telecom's 2026 hollow-core fiber tenders — at the tens-of-millions-of-fiber-kilometers scale — went entirely to YOFC.
How big is the market — what does Ciena's "$1,000 per GPU" mean?
Ciena projects the total addressable market for AI wide-area networking (including networking, DCI, and scale-across) growing from $7 billion to $14 billion, a ~30% CAGR from 2026 to 2029. Scale-across alone could reach $8–10 billion by 2029.
Unit economics: for an 80 km span requiring 400+ fiber pairs, the scale-across market is roughly $500 million; spread across a 500,000-GPU cluster, that works out to $1,000 per GPU (excluding amplifiers). Stretching the span to 1,000 km doubles the figure to $1 billion; at 2,000 km it jumps to $2 billion.
In plain terms = the farther apart the data centers sit and the more fiber they need, the higher the networking cost allocated to each GPU — that is the core variable in scale-across economics.
Caveat: these figures come entirely from Ciena, which is itself a major supplier in this market. Its TAM includes the value of its own line systems and modems. Nomura has not built an independent scale-across market model, nor translated these figures into revenue expectations for Chinese vendors.
Which of the three paths reaches commercial scale first?
Coherent pluggable orders have materialized first — Acacia posted back-to-back billion-dollar quarters, making this the most certain beneficiary today.
OCS carries dual endorsement from Google and Nvidia and sharply upgraded market guidance, but scaled deployment milestones remain to be proven.
Hollow-core fiber is the emblematic technology of scale-across — North American hyperscalers are deploying — but Nomura expects volumes over the next two to three years to stay at only a few million fiber-kilometers, not yet a major driver of overall fiber demand.
This means → in the near term, coherent pluggables capture revenue first; the investment pace for OCS and hollow-core fiber hinges on actual scale-across cluster build-out progress.
市场有风险，内容仅供研究参考，不构成投资建议。