Lambda Deploys Photonic Switches for 10,000 GPUs
Lambda connected more than 10,000 NVIDIA GB300 NVL72 GPUs using co-packaged optics switches, reporting lower latency with unchanged throughput.
Updated on Nov. 17, 2026 in Data Centers

Lambda deployed NVIDIA Photonics switches to connect more than 10,000 NVIDIA GB300 NVL72 GPUs. In its tests, the company reported lower latency at rack and pod scale while throughput stayed the same as conventional pluggable optics.
Why it matters
Latency and power draw are growing constraints as AI clusters scale to thousands of GPUs. The results suggest co-packaged optics could reduce interconnect bottlenecks in large data center deployments.
The switches have 144 ports running at 800 gigabits per second and 18 removable external light sources, with liquid cooling holding median optics temperatures at 36°C versus 51°C for conventional optics. Lambda reports using one-quarter as many lasers, though the company's figures for power and reliability gains have not been independently confirmed.
The players
Lambda
Lambda is a cloud computing company that provides GPU infrastructure for artificial intelligence workloads.
NVIDIA
NVIDIA is a semiconductor company that designs GPUs widely used for AI training and inference.
The details
The co-packaged optics design shortens the electrical path from inches to millimeters and removes the need for retimer chips. Lambda reports the approach delivers five times the power efficiency and 15 times the reliability of conventional optics.
Timeline
Lambda's deployment and test results were reported on November 17, 2026.
The Tech Race
Co-packaged optics is emerging as a competitive lever in the AI infrastructure race, as NVIDIA Photonics switches move interconnects closer to the GPU. Early deployments like Lambda's show whether this approach can displace pluggable optics at cluster scale.
For users of AI cloud services, lower interconnect latency could translate into faster training and inference times. Adoption will depend on whether providers choose co-packaged optics over conventional pluggable hardware.
The takeaway
Interconnect design is becoming as important as GPU performance in large AI clusters. Watching independent benchmarks of co-packaged optics will show whether these gains hold outside vendor-reported tests.
Further reading
Explore more on Data Centers to see how AI cluster networking is evolving.



