News

Lambda deploys NVIDIA Photonics switches across 10,000-plus GPUs

Lambda has deployed NVIDIA co-packaged-optics switches at its Quebec campus to connect GB300 NVL72 racks spanning more than 10,000 Blackwell Ultra GPUs.

D
Nov 16, 2026 · 2 min read

Lambda has deployed NVIDIA Photonics switches at its Quebec campus to connect GB300 NVL72 systems spanning more than 10,000 Blackwell Ultra GPUs, the companies said. The installation brings co-packaged optics into a large AI cluster by moving optical components closer to a network switch’s main processor.

According to Lambda’s deployment announcement, each liquid-cooled switch has 144 ports operating at 800 gigabits per second and 18 removable external light sources. Neither company named the switch model, identified the network as InfiniBand or Ethernet, or disclosed how many switches the campus uses. NVIDIA’s technical specifications associate a 144-port configuration with Quantum-X Photonics InfiniBand, but that product name does not appear in the deployment announcements.

With conventional pluggable optical networking, electrical signals travel from a switch chip across a circuit board to optical modules at the front of the switch. The companies say their co-packaged design places optical engines beside the switch ASIC, shortening the electrical path from inches to millimeters and eliminating retimer chips. External modules supply light to the engines and can be replaced in the field.

Lambda said its tests maintained the same throughput as conventional pluggable optics while cutting latency by 3.8% at rack scale and 10.1% at pod scale. It also reported a median optics temperature of 36 C, versus 51 C for pluggable optics, and said signal integrity was more stable. The companies did not provide absolute latency values, workload details, sample sizes or a full test methodology. The sources reviewed for this story contain no independent validation of the results.

Separately, NVIDIA claims the Photonics platform can deliver up to five times better power efficiency, up to 15 times higher reliability and four times fewer lasers than pluggable optics. These are vendor product claims, not measured results disclosed for Lambda’s Quebec cluster. The deployment announcements report no cluster-level power savings.

NVIDIA’s own materials also use different comparisons. An August 2025 technical description cited 3.5 times the power efficiency and 10 times the resiliency for its Photonics platforms, rather than the newer figures of up to five times and 15 times. The opened materials do not explain whether the figures use different products, test conditions or definitions.

More news