Skip to main content
NVIDIA CEO Jensen Huang presents a circuit board with a large GPU, highlighting Lambda's 5x power efficiency using NVIDIA pho

Editorial illustration for Lambda Uses NVIDIA Photonics For 5x Power Efficiency In AI Factories

Lambda Cuts AI Power Use 5x With NVIDIA Photonics

Lambda Uses NVIDIA Photonics For 5x Power Efficiency In AI Factories

• 2 min read

Lambda is connecting more than 10,000 GPUs across NVIDIA GB300 NVL72 racks using NVIDIA Photonics co-packaged optics switches, a move the cloud provider says cuts power draw while making its AI factory network harder to knock offline. The company's tests found network latency fell by about 4% at rack scale and roughly 10% at pod scale, with throughput holding steady against conventional pluggable optics across rack, scalable unit, and pod configurations.

The stakes are specific to how AI training and inference actually run. A single interrupted link in a large training job can stall thousands of GPUs at once. In agentic inference setups, one network hiccup can ripple into cascading latency across many agents working in parallel.

Every watt spent keeping a network alive is a watt pulled away from compute, and every interruption burns time that a GPU cluster can't get back. That's the resiliency and efficiency problem Lambda is trying to solve by swapping out pluggable transceivers for co-packaged optics at this scale, according to Dave Ward, president of cloud services at Lambda.

Lambda is tackling the network piece of the puzzle with NVIDIA Photonics co-packaged optics (CPO) switches. The cloud provider is using the switches to connect NVIDIA GB300 NVL72 racks with more than 10,000 GPUs, delivering lower power, higher resiliency and longer mean time between interruptions (MTBI) to an AI factory at scale.

Why this matters

For anyone building or buying AI infrastructure, the bottleneck has quietly shifted from compute to the network stitching that compute together. Lambda's bet on NVIDIA's co-packaged optics isn't a vendor press release footnote, it's a signal that pluggable transceivers are becoming the weak link in clusters running 10,000+ GPUs. A 5x power efficiency gain and 15x reliability improvement, if they hold up at scale, directly change the economics of running GB300 NVL72 racks: less power wasted on networking means more headroom for GPUs, and fewer failed links means fewer stalled training runs.

For founders pricing out AI factory builds, MTBI isn't an abstract spec, it's the difference between a model finishing training on schedule and an expensive restart. We'd want to see independent benchmarks once Lambda's deployment is running at full scale rather than take the 4x-fewer-lasers claim at face value, but the direction is clear: as GPU counts climb past ten thousand, the switch fabric stops being plumbing and starts being a competitive variable worth watching closely.

Common Questions Answered

How much power efficiency improvement does Lambda achieve using NVIDIA Photonics co-packaged optics switches?

Lambda achieves a 5x power efficiency gain by implementing NVIDIA Photonics co-packaged optics (CPO) switches to connect over 10,000 GPUs across NVIDIA GB300 NVL72 racks. This significant reduction in power draw makes the AI factory network more economical to operate at scale while maintaining steady throughput compared to conventional pluggable optics.

What latency improvements does Lambda report when using NVIDIA Photonics CPO switches at different scales?

Lambda's tests found that network latency decreased by approximately 4% at rack scale and roughly 10% at pod scale when using NVIDIA Photonics co-packaged optics switches. These improvements demonstrate the effectiveness of the CPO technology in reducing network overhead across different cluster configurations.

Why is network infrastructure becoming a critical bottleneck in large-scale AI clusters?

For AI infrastructure with 10,000+ GPUs, the bottleneck has shifted from compute capacity to the network infrastructure that connects the compute resources together. Pluggable transceivers are becoming the weak link in these massive clusters, making solutions like NVIDIA's co-packaged optics essential for maintaining efficiency and reliability at scale.

What reliability improvements does Lambda's deployment of NVIDIA Photonics switches provide?

Lambda reports a 15x reliability improvement through the use of NVIDIA Photonics co-packaged optics switches, which deliver higher resiliency and longer mean time between interruptions (MTBI) to the AI factory. This enhanced reliability directly impacts the economics of running large-scale GB300 NVL72 rack deployments by reducing downtime and maintenance costs.

LIVE23:37California Sues OpenAI After It Lost Control of AI Agents 80+ Times