Artificial intelligence infrastructure is moving into a new constraint: communication. As GPU clusters grow to tens of thousands of accelerators, optical interconnects (photonic light-based communications) are becoming a system-level requirement rather than a supporting data-center component, according to Yole Group’s latest photonics analysis.
Yole’s white paper, Photonics at the Speed of AI, argues that the industry’s main challenge is shifting from simply deploying enough compute to connecting that compute efficiently. In massive AI-training clusters, bandwidth limitations restrict accelerator utilization, latency slows synchronization, and interconnect power consumption competes directly with the power budget available for processing.
“AI makes photonics a system requirement, not an optional enhancement,” said Martin Vallo, senior technology and market analyst for photonics at Yole. “The bottleneck is shifting from compute to communication.”
A Four-Layer Interconnect Challenge
Yole frames the opportunity through four interconnect domains: Scale-In, Scale-Up, Scale-Out and Scale-Across. Each represents a different distance, performance requirement and technology transition within AI infrastructure.
The distinction matters because no single technology is expected to replace all existing interconnect approaches. Copper remains important, particularly within racks and over very short reaches, and is being pushed further through increasingly sophisticated, megawatt-class rack designs. Meanwhile, optical technologies are moving closer to processors and switches where electrical links reach practical limits in bandwidth, signal integrity and power efficiency.
The report rejects the idea of an abrupt “copper is dying” transition. Instead, it describes a staged migration in which copper is deliberately extended while optics are introduced where their bandwidth-density and reach advantages justify the additional cost and integration complexity.
Demand Outruns Conventional Scaling
The demand math is especially challenging for the optical supply chain. Yole estimates that AI models are increasing in scale by roughly 100 times every two years, while compute capability advances by only about 3.3 times over the same period. Closing that gap requires approximately 70 times more interconnects every two years—a growth rate that conventional optical-module manufacturing is not currently structured to meet.
That pressure is expected to fuel a major expansion in the optical-transceiver market. Yole projects the sector will grow from approximately $10 billion in 2021 to $112 billion by 2031, representing an estimated 35% compound annual growth rate. The forecast is tied to exceptional hyperscaler investment: capital expenditures are projected to reach $670 billion in 2026 alone and $5.3 trillion cumulatively from 2026 through 2031.
Optics Moves Closer to Silicon
Pluggable transceivers remain central to today’s deployments, but Yole sees the industry evolving through linear pluggable optics (LPO), XPO and the Open CPX ecosystem. The common direction is to move optical conversion progressively closer to GPU and CPU packages, reducing the electrical distance that high-speed signals must travel.
Co-packaged optics, or CPO, is expected to become a significant part of that transition, although initially in narrow applications. Rather than a universal near-term replacement for pluggable modules, CPO is entering commercial deployment primarily through proprietary switching platforms operated by the limited number of companies with sufficient vertical integration to manage manufacturing, serviceability and supply-chain risk.
Yole forecasts CPO optical-engine revenue will rise from $0.6 billion in 2026 to $112.1 billion in 2031. Scale-up applications—high-bandwidth links connecting accelerators within tightly integrated AI systems—are expected to account for about 94% of that market.
For the photonics sector, the implication is clear: AI infrastructure is creating an optical-connectivity challenge measured not only in speed, but in manufacturability, power efficiency, packaging and deployment scale. The competitive question is increasingly whether network and photonic technologies can be delivered in sufficient volume, on schedule and at an economically viable cost.
More information is available here.
Top image courtesy of t1nexus.com.
You must be logged in to post a comment.