Lambda, a cloud provider specializing in GPU infrastructure, has announced the integration of NVIDIA's co-packaged optics (CPO) switches, marketed as NVIDIA Photonics, into its AI training clusters. The switches connect GB300 NVL72 racks, each housing a large number of GPUs, into a combined system exceeding 10,000 interconnected accelerators. The stakes are significant: training large models requires thousands of GPUs to communicate continuously, and any network failure can slow or interrupt a costly training run.
Co-packaged optics technology differs from traditional networking approaches by placing optical components directly alongside switching silicon, reducing the number of discrete optical transceivers typically required. This design cuts down on potential failure points, lowers power consumption associated with repeated electro-optical conversion, and reduces transmission latency between nodes. For an operator like Lambda, this translates into better cluster uptime and lower operating costs, both critical factors in a market where demand for AI compute keeps accelerating.
The announcement fits into NVIDIA's broader strategy of positioning its networking technology as a strategic component on par with its GPUs for what the company calls 'AI factories' — infrastructure dedicated entirely to training and running large-scale models. Lambda is among the first cloud providers to deploy this technology at a scale exceeding ten thousand GPUs, making it a notable reference case for validating the reliability and energy-efficiency claims of co-packaged optics in production environments.
While the efficiency and resiliency gains cited by NVIDIA and Lambda still require independent, long-term verification, the deployment points to a broader shift: as AI clusters grow larger, networking is becoming as much of a bottleneck as raw compute power. Co-packaged optics could emerge as a standard component for the next generation of large-scale AI infrastructure.