News: Optoelectronics
7 September 2026
As AI models scale 100x every two years, Yole maps shift from compute to communication
In Yole Group’s White Paper ‘Photonics at the Speed of AI’ (part of the firm’s AI White Papers Collection), it argues that optical transceivers — once a supporting piece of data-center infrastructure — have become a system-level requirement for scaling artificial intelligence.
For most of the past decade, optical transceivers were infrastructure: important, technically demanding, but infrastructure nonetheless. According to Yole Group, that era is over. When a cluster of tens of thousands of GPUs is built to train a single model, compute is only as useful as the communication fabric that holds it together: latency stalls synchronization, bandwidth throttles utilization, and interconnect power competes directly with compute power. The bottleneck has migrated to the wire, the fiber, and the components that convert between them.
“AI makes photonics a system requirement, not an optional enhancement,” says Martin Vallo PhD, Yole’s senior technology & market analyst, Photonics. “The bottleneck is shifting from compute to communication.”
To make sense of that shift, Yole’s White Paper introduces a four-domain framework, Scale-In, Scale-Up, Scale-Out, Scale-Across, reflecting how interconnect engineers think about the problem today. Each layer has a different technology entry point, a different incumbent to displace, and a different commercialization timeline.
The picture is neither a simple ‘copper is dying’ narrative nor a ‘CPO will solve everything’ announcement: copper is being extended deliberately through megawatt-class rack engineering, while co-packaged optics (CPO) is entering commercial deployment narrowly, in proprietary switching platforms, for the handful of players vertically integrated enough to absorb its manufacturing risk.
The scaling math is stark. AI models grow roughly 100x every two years while compute improves only about 3.3x, pushing the industry to deploy about 70x more interconnects every two years – a rate the optics module industry is not currently equipped to sustain. That pressure flows directly into the transceiver market: from about $10bn in 2021 to a forecast $112bn in 2031 (~35% CAGR), fueled by hyperscaler CapEx projected to reach $670bn in 2026 alone, and a combined $5.3 trillion between 2026 and 2031.
The White Paper also maps where pluggable optics are headed next, evolving through LPO, XPO and the Open CPX ecosystem toward optical conversion progressively closer to the GPU and CPU package. Yole’s new market & technology report ‘Co-Packaged Optics for Data Centers 2026’ quantifies that shift: CPO optical-engine revenue is forecast to grow from $0.6bn (2026) to $112.1bn (2031), with scale-up applications alone expected to capture about 94% of that market.
“As AI infrastructure scales globally, the industry’s primary constraint is shifting,” notes founder & president Jean-Christophe Eloy. “The question is no longer simply which accelerators can be deployed, but whether the networking and photonic technologies required to connect them will be available at the right scale, at the right time, and at the right cost.”
