Nvidia, in partnership with Lambda, has introduced a new photonics co-packaged optics switch designed to handle the soaring data demands of artificial intelligence workloads. The announcement, made earlier this week, marks a step toward optical interconnects that could slash power consumption while cranking up bandwidth inside AI clusters.
How the switch works
The switch blends optical components directly with silicon electronics—a technique known as co-packaged optics. Instead of relying on traditional electrical cables that generate heat and limit speed, the design uses light to move data between chips. That means less energy wasted and faster communication, especially critical when training large language models or running inference at scale. Nvidia says the approach can significantly reduce the physical footprint of network gear in a data center.
Why photonics matters for AI
AI models are growing fast, and the networks that connect thousands of GPUs struggle to keep up. Electrical interconnects hit a wall on bandwidth and power. Photonics sidesteps those limits. By integrating optics at the switch level, the new hardware aims to deliver higher throughput with lower latency. For companies running massive training jobs, that could translate into shorter training times and lower electricity bills.
Lambda’s role
Lambda, a cloud provider known for selling GPU-heavy servers and offering cloud instances, joined Nvidia to develop the switch. Lambda’s engineers worked on the integration and testing. The company has been focused on making AI infrastructure more accessible, and the new switch could eventually appear in Lambda’s own data centers—though no timeline has been shared.
What comes next
Nvidia and Lambda haven’t released pricing or a specific availability date. The switch is expected to go through qualification with hyperscaler partners before hitting the broader market. For now, the two companies are positioning it as a piece of the puzzle for next-generation AI networks, where every millisecond and every watt count.
