Skip to content
AI.info

The Pulse

Lambda connects more than 10,000 GPUs with photonics switches

Lambda says its Quebec campus now connects more than 10,000 NVIDIA Blackwell Ultra GPUs using co-packaged optics switches. The company reported lower latency and cooler optics in benchmarks against conventional pluggable optics.

Lambda connects more than 10,000 GPUs with photonics switches

More than 10,000 NVIDIA Blackwell Ultra GPUs at Lambda's Quebec campus are now connected using photonics switches with co-packaged optics. Lambda says the installation is the world's first large-scale deployment of the networking technology and reports lower latency than conventional optics in its benchmarks.

The company first announced its early adoption of NVIDIA co-packaged optics, or CPO, networking in November 2025. The deployed NVIDIA Photonics switches have 144 800Gb/s ports, liquid cooling and 18 removable external light sources.

Benchmark results

Lambda reported that CPO delivered the same throughput and more stable signal integrity than conventional pluggable optics across rack, scalable-unit and pod configurations. Latency was lower at each scale, with the reported gap growing from 3.8% at rack level to 10.1% at pod level.

The CPO optics had a median operating temperature of 36°C, compared with 51°C for conventional pluggable optics, according to the company. It also claims that cooler, more stable optics provide up to 15 times higher reliability.

Moving optics closer to the switch

Conventional pluggable optics convert electrical signals to light at a switch's front panel. CPO instead places optical engines alongside the switch ASIC, shortening the electrical path from inches to millimeters. Lambda says that arrangement reduces signal loss, removes the need for retimer chips and increases network resiliency.

In June, the company described an early engineering sample of an NVIDIA Quantum-X InfiniBand Photonics Q3450-LD switch and its installation requirements. The account covered rack fit, power delivery, cooling, fiber routing and installation procedures, with Lambda and NVIDIA working together on those infrastructure details.

That switch supports 18 removable external light-source modules, each feeding eight MPO ports; the lasers remain external so they can be serviced in the field. The rear uses a 48V DC busbar for power and four liquid-cooling connections with dual internal loops.

Campus and deployment

In March, Lambda said a production-scale cluster using NVIDIA Quantum-X InfiniBand Photonics CPO switches was running in an AI factory with more than 10,000 NVIDIA GB300 GPUs. It said its cluster design planned power and liquid cooling at the rack level and tested firmware, drivers and orchestration together before handing over clusters.

The Quebec campus has 142 MW of secured, nearly 100% renewable energy capacity. Lambda says the site enables free cooling for up to 80% of the year, is guaranteed to a power usage effectiveness of 1.2 and has no process water consumption associated with server cooling.

The deployment will be featured in NVIDIA's SC26 special address, where the implementation is due to be presented in more detail.

Sources

Related stories

Explore

More articles