AI Scaling Bottleneck Shifts from Compute Power to Communication Infrastructure
Yole Group warns that a massive disparity between model growth and compute performance makes photonics a critical system requirement.
The primary obstacle to scaling artificial intelligence has shifted from raw processing power to the infrastructure used to move data. According to a new white paper from Yole Group, the industry faces a critical communication bottleneck that threatens to cap the growth of AI models regardless of chip breakthroughs.
In the report, "Photonics at the Speed of AI," Yole Group highlights a stark divergence in growth rates: AI models scale approximately 100x every two years, while compute performance improves by only about 3.3x in that same window. To bridge this gap, the industry must deploy roughly 70x more interconnects every two years. This shift has transformed optical transceivers from supporting data center infrastructure into a primary system-level requirement. Martin Vallo, PhD, Yole’s senior technology and market analyst for photonics, notes that the bottleneck is now shifting from compute to communication.
The Networking Redesign
Historically, AI acceleration focused on the chip. However, as training clusters expand to include tens of thousands of GPUs for a single model, the communication fabric—specifically latency, bandwidth, and power—now directly limits the utility of available compute. To address this, Yole introduced a four-domain framework consisting of Scale-In, Scale-Up, Scale-Out, and Scale-Across to map the necessary redesign of AI data center networking.
This transition is driving massive capital investment. Hyperscaler CapEx is projected to reach $670 billion in 2026, with a combined total of $5.3 trillion expected between 2026 and 2031. A significant portion of this spending will flow into photonics. The optical transceiver market is forecast to grow from $10 billion in 2021 to $112 billion by 2031, representing a compound annual growth rate (CAGR) of approximately 35%.
Industry Implications
The shift indicates that the future of AI progress depends on solving the "interconnect bottleneck." If the optics module industry cannot sustain the required 70x growth in interconnects, the physical networking layer will become a hard ceiling for model scaling. This has accelerated the adoption of Co-Packaged Optics (CPO). Yole projects that CPO optical-engine revenue will surge from $0.6 billion in 2026 to $112.1 billion by 2031, with scale-up applications expected to capture 94% of that market.
What's Next
As the industry pivots, the critical question is no longer which accelerators to deploy, but whether the necessary photonic technologies will be available at the required scale, time, and cost, according to Yole Group founder and president Jean-Christophe Eloy. The coming years will determine if the supply chain for optical interconnects can keep pace with the exponential demands of generative AI models.