Lambda Deploys NVIDIA Photonics Across a 10,000-GPU AI Factory

Lambda is using NVIDIA Photonics co-packaged optics switches to connect NVIDIA GB300 NVL72 racks containing more than 10,000 GPUs. NVIDIA says the deployment delivers lower power use, higher resiliency and longer mean time between interruptions for AI training at scale.
- Lambda is connecting more than 10,000 GPUs with NVIDIA Photonics switches.
- Latency fell by about 4% at rack scale and up to about 10% at pod scale in NVIDIA tests.
- NVIDIA claims up to 5x better power efficiency and up to 15x higher reliability.
- Each switch has 144 ports running at 800Gb/s and 18 removable light sources.
- Median optics temperature was 36°C, compared with 51°C for conventional pluggable optics.
Co-packaged optics connect Lambda’s GPU racks
The cloud provider’s deployment uses NVIDIA Photonics switches to connect the racks in an AI factory. Each switch has 144 ports running at 800Gb/s, while liquid cooling removes heat from the switch chip and optical engines.
Lambda president of cloud services Dave Ward said, “Networking has to keep pace with the demands of increasingly large AI clusters.” He described bringing co-packaged optics into a deployment of this scale as “an important step toward building the infrastructure needed to support the next generation of AI.”
According to NVIDIA, tests with the switches reduced network latency, the time needed to move data, by about 4% at rack scale and by up to about 10% at pod scale. Across rack, scalable unit and pod configurations, the network maintained the same throughput as conventional pluggable optics.
NVIDIA targets power use and interruptions
NVIDIA said the network’s power efficiency and resiliency are important because it sits at the center of AI factory productivity. During a large training job, one interrupted link can stall thousands of GPUs. The company also said that, for agentic inference, a network disruption can create cascading latency effects across many agents.
NVIDIA Photonics replaces pluggable transceivers with co-packaged optical engines on the switch package. According to NVIDIA, this provides up to 5x better power efficiency and up to 15x higher reliability than conventional pluggable optics. The design also uses 4x fewer lasers, reducing the number of components that can fail.
NVIDIA said the lower power demand matters because every watt used for networking is unavailable for compute, while interruptions can add time to AI workloads.
Replaceable light sources support field maintenance
The optical engines sit alongside the switch chip on the same package. They convert electrical signals into light, which carries data over fiber between parts of the cluster. Co-packaging shortens the electrical path from inches to millimeters, according to NVIDIA, reducing signal loss and removing the need for retimer chips.
Each switch has 18 removable external light sources that supply light to the optical engines. Technicians can replace these light-source modules in the field. With traditional pluggable optics, a laser failure requires replacing the entire module.
NVIDIA reported a median optics temperature of 36°C in Lambda’s deployment, compared with 51°C for conventional pluggable optics, and said its tests found more stable signal integrity. For Lambda, the combination of lower latency, power efficiency and serviceable light sources is intended to support the operation of a large GPU network while reducing the effects of failures and heat.


