
Lambda says its 10,000-GPU Nvidia cluster in Quebec is the first large-scale production deployment of co-packaged optics - MLQ.ai
AI cloud company Lambda has connected more than 10,000 Nvidia Blackwell Ultra GPUs at its Quebec campus using Nvidia Photonics co-packaged optics switches, claiming it's the first large-scale production deployment of CPO technology.
AI cloud company Lambda announced on Monday, Nov. 16, that it has connected more than 10,000 Nvidia Blackwell Ultra GPUs at its Quebec campus with Nvidia Photonics networking switches using co-packaged optics (CPO). The company claims this deployment is the world's first large-scale production implementation of CPO technology. The GB300 NVL72 racks in the cluster are linked by the switches and are being featured in Nvidia's special address at SC26, the supercomputing conference.
ServeTheHome, which has closely followed CPO development for years, confirmed the cluster is operational and described it as the first production CPO cluster. This announcement comes after a vendor keynote at the OCP Summit had suggested CPO technology was still years away from practical deployment.
Performance Advantages
Lambda's tests revealed significant performance benefits of CPO technology compared to conventional pluggable optics. The optics ran at a median temperature of 36 degrees C versus 51 degrees C for pluggable alternatives. Latency was 3.8% lower at rack scale and 10.1% lower at pod scale, while maintaining the same throughput.
Nvidia claims its CPO technology delivers up to 5x better power efficiency, up to 15x higher reliability, and uses 4x fewer lasers than pluggable optics. These figures are vendor-provided specifications rather than actual failure data from Lambda's deployment.
Networking has to keep pace with the demands of increasingly large AI clusters.
Technical Innovation
Co-packaged optics fundamentally changes how AI networks operate. In conventional setups, switch chips send electrical signals over copper to pluggable optical transceivers, which convert them to light for fiber transmission. CPO integrates the optics directly onto the switch package, eliminating the copper path and allowing more power to be directed to GPUs.
Each switch in Lambda's deployment features 144 ports at 800Gb/s and is liquid-cooled. The system uses 18 removable external laser sources on the front of the chassis, which can be swapped in the field. This design reduces potential failure points since lasers are typically the components that fail in pluggable optics.
Lambda had first announced its plans for this deployment in March at Nvidia's GTC conference, stating it was leading one of the largest deployments of Nvidia's Quantum-X InfiniBand Photonics CPO switches in a facility with over 10,000 GPUs. The recent announcement provides the first performance data from that cluster.
Infrastructure and Partnerships
Lambda's Quebec campus has 142 MW of secured, nearly 100% renewable capacity with a guaranteed power usage effectiveness (PUE) of 1.2. The company is a major supplier of compute to Anthropic, though their $35 billion agreement faces uncertainty as it can now be terminated by either party due to Lambda not securing the facility lease.
The deployment represents a significant advancement in networking technology for AI supercomputing, potentially enabling more efficient and reliable large-scale AI clusters. As AI models continue to grow in size and complexity, innovations like CPO will become increasingly critical for maintaining performance while managing power consumption and heat generation.