DR HELENA DIEZ-Y-RIEGA & DR. JON PUGH OPTICA INTERVIEW
THE NEXT FRONTIER: PHOTONICS IN THE AGE OF AI With its high‑bandwidth, low‑loss and energy‑efficient data transmission, photonics are well positioned with the performance needs of AI and data‑centre systems. Optical Connections Editor Brian Dolby spoke with Optica’s Technical Director for Laser and Life Science Technologies Dr Helena Diez-y-Riega and Technical Director for PICs and Quantum Technologies Dr. Jon Pugh on the key role of photonics in the new optical infrastructure era.
From your perspective, how well is photonics positioned to meet the bandwidth demands driven
and redrivers and making it increasingly difficult to build larger, denser AI systems. Photonics provides the bandwidth density and reach needed to overcome these physical limitations and enables AI infrastructure to continue scaling. The conversation has therefore shifted from whether photonics will be deployed at scale to how deeply it will penetrate AI infrastructure. Scale-out is already optical, scale-up is now becoming optical, and scale-across will rely on optical networking to interconnect the next generation of multi-campus AI factories as individual data centres reach the practical limits of available power and land. The challenge for the industry is no longer proving the need for photonics, but manufacturing and deploying these technologies at the scale and cost required by hyperscale AI infrastructure. BD Do you see photonic interconnects as the key solution for reducing latency and energy consumption in hyperscale data centres?
They also help reduce latency by simplifying the communication path. As electrical links become shorter at higher data rates, maintaining signal integrity requires additional processing stages, each introducing additional delay. By moving optics closer to the switch and compute silicon, data can traverse the network more directly with fewer intermediate electrical devices. However, photonic interconnects should not be viewed in isolation. Their full benefits are only realised alongside advances in switch silicon, advanced packaging, co-packaged optics and optical I/O. Together, these technologies are enabling a new generation of AI systems that are both more energy efficient and capable of delivering lower-latency communication at unprecedented scale. Ultimately, we see the future as a heterogeneous networking architecture. Copper will remain the most practical solution over very short distances, while photonic interconnects will increasingly be deployed wherever latency, energy efficiency and scalability become the limiting factors. The industry is no longer evaluating whether optics should be used, but where it delivers the greatest system-level benefit. BD What do you think are the biggest hurdles in integrating photonic components into existing data centre architectures? We believe the biggest hurdles have shifted away from demonstrating the technology itself and towards manufacturing and integrating photonic components at the scale, cost and reliability demanded by hyperscale AI infrastructure. The first challenge is advanced packaging. Bringing photonic and
BD
by AI and cloud workloads?
Photonics is exceptionally well positioned to meet the bandwidth demands of AI because the
OP
networking architecture of AI infrastructure is evolving across three distinct domains: scale-up, scale-out and scale-across. Scale-up connects GPUs within a high-bandwidth “super-node”, enabling a few to a few hundred processors to operate as a single system with shared memory. Scale-out connects many of these super-nodes together, allowing hundreds of thousands of GPUs to communicate across an AI factory. Scale-across extends this concept further by connecting multiple AI factories across different campuses or geographic locations into one logical computing resource. The key point is that scale-out has already become an optical networking problem. With the introduction of platforms such as NVIDIA’s Spectrum-X Ethernet architecture, optics is now fundamental to connecting large GPU clusters. The next major transition is scale-up, where optical technologies are moving closer to the compute itself through co-packaged optics, optical I/O and advanced photonic integration. This represents a significant shift, as optics is no longer confined to rack-to-rack or building-to-building connections, but is moving into the GPU interconnect architecture itself. As data rates continue to increase to 224 Gb/s per lane and beyond, copper interconnects suffer from increasing insertion loss, signal integrity degradation and crosstalk. The practical reach of high-speed electrical links continues to shrink, requiring additional retimers
Yes, we do. Photonic interconnects are becoming a key technology for reducing both
OP
latency and energy consumption, particularly as AI clusters continue to grow in size. From an energy perspective, moving data has become an increasingly significant contributor to the overall power budget of an AI factory. Optical interconnects require fewer high-speed electrical components over longer distances and, as the industry moves towards co-packaged optics and optical I/O, they reduce reliance on power- hungry DSPs, retimers and redrivers. This enables data to be transported more efficiently, allowing a greater proportion of the available power to be devoted to computation.
18
www.opticalconnectionsnews.com
| ISSUE 44 | Q3 2026
Made with FlippingBook Annual report maker