Omdia: AI Compute Interconnect Drives Optical Network Market to 25-Year High

C114, September 10 (Yue Ming) — "Optical networks have left the 2024 trough behind and entered a supercycle of AI infrastructure buildout, with the market reaching its highest level in 25 years and the growth engine shifting from bandwidth upgrades to AI compute interconnect," said Ian Redpath, research director for optical networks at market research firm Omdia, in a keynote assessment at the Optoelectronic Convergence Technology and Industry Development Forum held alongside the 27th China International Optoelectronic Exposition (CIOE 2026).

According to Omdia estimates, the global optical network market reached an annualized $16.9 billion in the second quarter of 2026, up 13.7% year over year. The second quarter of 2026 alone grew 26.7% from the same period of 2024, with both figures hitting their highest levels in nearly 25 years. Ian Redpath defined 2025 as "the beginning of the recovery" and 2026 as "predictable, AI-backed growth driven by order backlogs."

Two New Types of AI Infra Players

The key variable behind this trend is that AI has reshaped the scale of the data center. Annual capital expenditure across the global data center industry used to be about $100 billion; it is now rapidly approaching and surpassing roughly $1 trillion, a tenfold increase in investment intensity. As NVIDIA CEO Jensen Huang put it, "Optical components used to be an afterthought. Today they have become so, so important."

Notably, Omdia analysts took a close look at two new types of players in the AI infrastructure market.

One type is the NeoClouds — builders and operators of large-scale, centralized AI data centers, many with cryptocurrency backgrounds. They combine the operational know-how to run high-power GPU infrastructure with early partnerships with NVIDIA, and are stepping in to fill the buildout capacity gap that hyperscalers have been unable to keep up with demand.

Another category is NVIDIA itself, which is "vertically integrating" the entire ecosystem through minority stakes rather than control. Its investments span NeoClouds such as CoreWeave and Crusoe, model companies such as Anthropic and OpenAI, Intel (a chip company), and optical fiber and component companies such as Corning, Coherent, Lumentum, and Marvell, including optical fiber, and it has pledged to expand production in solid-core fiber, wafers, and lasers.

Scale Across Concept Rises

"NVIDIA has been using its financial muscle to 'clear' obstacles in the ecosystem that buys its equipment," Ian Redpath said, adding that NVIDIA's distributed training trials are driving the rise of the Scale Across concept, which is expected to become a new high-growth contributor to the market.

He explained that power, space, and cooling limits at a single data center constrain the development of hyperscale AI training models. Scale Across means interconnecting AI xPU resource pools across multiple data center sites, with the goal of maximizing model training efficiency while minimizing latency, power consumption, and cost.

NVIDIA's validation trials showed that Chicago and Ashburn, about 1,000 kilometers apart with 1,536 GPUs deployed at each site, achieved a measured round-trip latency of 21 milliseconds and scaling efficiency above 96%. Based on this, NVIDIA set a target of scaling to 500,000 GPUs. A joint survey by Ciena of more than 1,200 operator experts across 12 countries also confirmed this trend: 95% of respondents believe that hyperscalers' Scale Across demand will be a major contributor to wholesale revenue growth.

Ian Redpath noted that Scale Across is consolidating into three segments: under 20 km, dominated by Direct Detect and Coherent Lite (800LR); 20 to 150 km, requiring no line amplification and serving as the key battleground for 400ZR, 800ZR, and 1600ZR; and beyond 150 km, relying on 800ZR+ pluggable modules with interoperable PCS and 1.6T embedded solutions.

Multiple Technology Innovation Tracks Unfold in Parallel

At the same time, on the technology front, innovation is advancing along multiple tracks in parallel.

First, multi-layer, multi-vendor integration is maturing. OIF's CMIS specification defines management interactions between hosts and modules, and 800G hardware specifications are converging.

Second, multi-rail line systems are bringing innovation to the line system domain. By sharing resources such as OTDRs and OCMs, using C+L band amplification and new pump lasers, they can resolve the conflict between infrastructure built more than 20 years ago and the need to accommodate high-fiber-count optical cables. Hundreds of fiber pairs can now be supported within a 42RU rack.

Third, AI training is spawning a commercial optical circuit switching (OCS) market. Google's in-house 136×136-port OCS has achieved a 30% cost reduction and 40% power savings, while TeraHop, Lumentum and Coherent are following with silicon photonics, MEMS and liquid crystal approaches, respectively.

Fourth, liquid cooling is moving down from supercomputers to pluggable modules. A single XPO MSA module delivers 12.8Tbps of bandwidth, offers four times the front-panel density of OSFP and integrates a liquid-cooled cold plate. The XPO MSA now has more than 100 member companies.

Fifth, the 2nm coherent DSP era has begun. Nokia covers 13 network applications with four DSP building blocks, while 800G/1.6T products from Marvell and others are set to arrive in volume from the second half of 2026 through 2027.

Ian Redpath also outlined the opportunities carriers can seize: AI providers have almost no enterprise last-mile access capability and lack an "AI middle mile," while carriers hold vast enterprise customer relationships, making them best placed to connect enterprise data lakes to AI training centers and build AI middle-mile infrastructure at scale.

He also cautioned that Omdia's forecast rests on two premises: "geopolitical structures functioning normally" and "AI promises realizing their potential." Any change to either would require a reassessment.