Nvidia co-packaged optics switch reduces datacenter power consumption by 3kW per rack while maintaining performance.

The era of Silicon Photonics and Co-Packaged Optics (CPO) has arrived as AI firm Lambda receives and deploys one of the first NVIDIA Quantum-X InfiniBand platforms. Networking has become a critical component of AI factories, delivering interconnectivity at unprecedented speeds that move data between massive GPU clusters at the speed of light. These solutions reduce switch power consumption, minimize failure points, and enhance token throughput efficiency across AI infrastructure deployments.

NVIDIA has begun delivering its Quantum-X Infiniband solutions to leading AI firms including Lambda AI, which has integrated the platform into its AI ecosystem. The Quantum-X AI networking solution leverages silicon photonics co-packaged optics and delivers 800G capabilities in GB300 NVL72-scale racks. According to Lambda, the back-end fabric now accounts for 86% of their networking power in a three-layer cluster, representing a fundamental shift in how data center architecture is designed.

Lambda is among the first to deploy NVIDIA's Co-Packaged Optics solution, built around the Q3450-LD switch. The power efficiency gains are substantial: a standard switch solution consumes approximately 7.0kW of power, while NVIDIA's Silicon Photonics solution consumes 3.95kW, delivering 3.05kW of savings on GB300 "Blackwell Ultra" platforms. Beyond power efficiency, CPO dramatically reduces failure points—Lambda notes that a 128,000-GPU data center requires 655,000 discrete transceiver modules across its switching fabric, each representing a potential point of failure. CPO eliminates many of these optical components, leading to significantly fewer failure modes.

The Q3450-LD hardware features 18 removable light source modules feeding 144 MPO ports, with fiber-array connections that connect directly into the silicon photonics engine rather than using traditional OSFP cages. Power is supplied through 48V DC with DGX-compliant busbar connectors, while cooling is provided via four UDQ4 liquid cooling connections with dual internal loops—design choices familiar to organizations already running GB300 NVL72 racks.

As Agentic AI pushes data centers toward greater token throughput and computational efficiency, elastic and resilient data movement has become essential. CPO technology enables more compute within the same physical data center footprint, addressing the infrastructure demands of next-generation AI workloads. NVIDIA has positioned itself at the forefront of this transition, with other vendors working to match the scale and efficiency of the industry leader.