Optical AI data centre networks are becoming essential as clusters grow beyond the limits of electrical wiring. Inside a large AI facility, thousands of processors must exchange enormous streams of data. The connections between chips can determine how much expensive computing hardware is actually productive.
Copper remains useful over short distances, but higher bandwidth raises power loss and signal-integrity challenges. Fibre optics moves information as light, allowing longer reach and high data rates with less electrical resistance.
Why AI needs so much network bandwidth
Training a large model divides calculations across many accelerators. Those chips repeatedly exchange model parameters, gradients and intermediate results. If one part of the cluster waits for data, the entire system can lose efficiency.
Inference is also becoming more network-intensive. Large mixture-of-experts models, long context windows and AI agents that use multiple services can move data among processors, memory and storage at high speed.
From pluggable optics to co-packaged optics
Traditional optical transceivers plug into the edge of a network switch. Electrical traces still carry signals from the switching chip to those modules. As speeds rise, that electrical distance consumes more power.
Co-packaged optics brings optical components much closer to the switch silicon. Shorter electrical paths can reduce energy per bit and increase the total bandwidth leaving a package. The engineering is difficult because lasers, photonics, electronics and heat management must work together with high manufacturing yield.
Why laser supply has become strategic
In March 2026, Nvidia and Coherent announced a multiyear optics partnership. Nvidia committed to purchases and invested in research and manufacturing capacity for advanced laser and optical-network products.
The agreement is significant because it treats optics as a foundational part of AI infrastructure, not a commodity accessory. It is also a company announcement; deployment volumes, costs and technical results will need independent evidence.
Optical networking trade-offs: power, repairs and cost
Optical links can reduce energy and extend reach, but they are not free. Lasers consume power, photonic components must be precisely aligned, and repairs may be more complex than replacing a cable. Co-packaging can also make thermal design harder because optical components sit near hot switching silicon.
Data-centre builders must balance bandwidth, reliability, serviceability and cost. Different distances inside and between racks may use different technologies.
What this means for the AI race
The public conversation often treats GPUs as the whole AI supply chain. In practice, useful computation depends on memory, networking, storage, power and cooling. Faster processors can magnify a network bottleneck rather than solve it.
Optics may become one of the quiet technologies that decides how efficiently future AI factories scale. The winning systems will not simply contain the fastest chips; they will keep those chips fed with data.
Sources and further reading
This article reports on industry infrastructure and does not provide investment advice.
Leave a Reply