Optical Networking Has Become the AI Bottleneck
As model demands outpace compute gains, data centers must shift trillions in capital toward faster interconnects.
Updated on Sept. 21, 2026 in Data Centers
Live Poll
Do you believe the current massive investment in AI infrastructure will continue over the next decade?
Yole Group reports that while AI models scale 100x every two years, compute power only improves 3.3x, forcing a shift in focus to communication infrastructure. This research highlights the critical role of optical transceivers in bridging the widening performance gap.
Why it matters
The massive scaling gap between AI model growth and compute capability forces the deployment of 70x more interconnects every two years to prevent severe bottlenecks. This evolution in the communication fabric is now the primary constraint on latency, bandwidth, and power consumption.
To address scaling gaps, hyperscalers are adopting systems like LPO (Linear-drive Pluggable Optics), XPO, and Open CPX to place optical conversion closer to processing units. The market for CPO (Co-Packaged Optics) engines is forecast to reach $112.1 billion by 2031, representing a massive expansion from $0.6 billion in 2026.
The players
Yole Group
A market research and strategy consulting firm that specializes in semiconductor, photonic, and power electronics supply chains.
The details
The report outlines a four-domain framework: Scale-In, Scale-Up, Scale-Out, and Scale-Across, where Scale-Up applications currently capture 94% of the market. Networking infrastructure is evolving to move optical conversion—the process of changing electrical signals into light for faster transport—closer to the GPU and CPU packages. This minimizes the distance data must travel as an electrical signal, reducing both power consumption and latency.
Timeline
2021: Optical transceiver market reached $10 billion.
2026: Hyperscaler capital expenditure is projected to hit $670 billion.
2026-2031: Total hyperscaler capital expenditure is expected to reach $5.3 trillion.
2031: Optical transceiver market is forecast to reach $112 billion.
The Tech Race
The transition to advanced interconnects marks a departure from the historical era where simply increasing transistor counts solved performance bottlenecks. The industry race is now focused on the communication fabric, with hyperscalers funneling $5.3 trillion into infrastructure to maintain AI model momentum.
Industry analysts expect the 35% CAGR in the transceiver market to dictate hardware procurement cycles for data center operators through 2031. Organizations should track the shift toward CPO and LPO architectures as these technologies become prerequisites for high-performance AI compute nodes.
The takeaway
The bottleneck for future AI scaling is moving from the processor to the network, necessitating a fundamental change in how data center hardware is designed. Watch the 2026 capital expenditure numbers, as they will confirm whether the $5.3 trillion industry investment is successfully closing the interconnect gap.
Further reading
For more on how infrastructure is evolving to meet compute demand, visit /tech/data-centers/.
Source note: This article includes information reported by DQ.
Live Poll
Do you believe the current massive investment in AI infrastructure will continue over the next decade?
