Lambda Scales a 10,000-GPU AI Factory with NVIDIA CPO Switches
Summary:Lambda has deployed NVIDIA Photonics co-packaged optics switches in a production AI factory connecting more than 10,000 NVIDIA GB300 NVL72 GPUs. Each switch provides 144 ports rated at 800 Gb/s. Lambda reports that network latency fell by about 4% at rack scale and by up to 10% at pod scale while sustaining the same throughput as conventional pluggable optics. The CPO design places optical engines beside the switch ASIC, shortening electrical paths from inches to millimeters, removing retimers
Deployment at production scale
Lambda has deployed NVIDIA Photonics co-packaged optics switches to connect GB300 NVL72 racks in an AI factory spanning more than 10,000 GPUs. Each switch exposes 144 ports rated at 800 Gb/s. Summing the nominal one-way port rates yields 115.2 Tb/s of raw bandwidth, although this is not the same as application-effective throughput after protocol overhead, congestion and overlap with computation are removed.
The important point goes beyond port speed. Lambda says the deployment spans rack, scalable-unit and pod configurations and is already operating in a production AI factory. The public information does not disclose the full topology, switch count, fiber reach, oversubscription ratio, congestion-control implementation or NCCL version, so it is not possible to reconstruct the network's exact contribution to job completion time.
How CPO changes the signal path
In a conventional pluggable-optics switch, electrical signals travel from the switching ASIC through the package substrate, printed circuit board, high-speed connectors and an external optical module before being converted for transmission over fiber. At 800 Gb/s, the longer electrical path increases insertion loss, reflections and equalization difficulty, while retimers add power and potential failure points.
CPO places the optical engines directly beside the switching ASIC, reducing the critical electrical path from inches to millimeters. The reported Lambda configuration includes 18 removable external light sources that feed the packaged optical engines over fiber. The engines perform the electrical-to-optical conversion close to the ASIC.
This architecture has three direct consequences:
- A shorter electrical path reduces high-speed signal loss and allows retimers to be removed;
- Co-packaging reduces component count across the package, board and optical-module boundaries;
- External light sources can be replaced independently when a source fails, instead of requiring replacement of the entire pluggable module.
Replaceable light sources do not, however, mean every CPO failure is hot-swappable. A failure in a packaged optical engine or the switching ASIC may still require package-level or complete-switch service. The report also does not state whether the 18 light sources provide redundancy, automatic failover or hot-swap support.
Reported results: gains center on latency, power and reliability
All figures below come from internal tests disclosed by NVIDIA and Lambda and have not been independently benchmarked:
| Metric | Reported result | Engineering interpretation and limitation | |---|---:|---| | Rack-scale network latency | About 4% lower | A relative improvement; absolute latency, traffic model and distributions were not disclosed | | Pod-scale network latency | Up to about 10% lower | The source of the gain at a particular switch tier or traffic class is not identified | | Throughput | Equal to pluggable optics | CPO does not directly increase bandwidth; its main benefits are efficiency and availability | | Power efficiency | Up to 5× better | The denominator is unclear: optical engine, switch chip or complete switch | | Reliability | Up to 15× higher | No MTBI value, failure definition, sample size or confidence interval is provided | | Laser count | 4× fewer | Indicates a different light-source architecture, but sharing and redundancy details are undisclosed | | Median optics temperature | 36°C versus 51°C for pluggable optics | Liquid cooling removes heat and Lambda reports more stable signal integrity |
The 4% and 10% latency improvements cannot be translated directly into training-throughput gains. Collective operations such as All-Reduce amplify tail latency and the straggler effects caused by failed links. NCCL collective operations, gradient sizes, message fragmentation and end-to-end job completion time were not included, so the data do not yet demonstrate an end-to-end training improvement.
Why reliability and power matter to AI-factory output
In a 10,000-GPU training job, one failed critical link can leave many GPUs waiting for a collective operation to complete, while the cluster continues consuming power and cooling capacity. Lambda also notes that bursty agentic-inference traffic can turn a local network incident into cascading latency across multiple agents.
CPO therefore offers more than lower energy use. It can reduce failure points in the optical path and shorten mean repair time. Lower switch power releases capacity for GPU computation, while a longer MTBI can reduce large-scale training restarts and agent-service degradation.
Reliability still needs fault-injection validation. Operators should test external light sources, packaged optical engines, power supplies, firmware upgrades, fiber contamination and port reconfiguration, then record detection time, failover time, retransmission volume, job interruption time and MTTR. The reported 15× maximum does not replace these production metrics.
Engineering trade-offs
Packaging and supply-chain exposure
Co-packaging the optical engine with the switching ASIC more tightly couples semiconductor fabrication, optical packaging and system thermal management. Lower interface and connector counts are offset by the risk that package yield, assembly yield or component shortages affect complete-switch delivery. Lambda's deployment indicates that CPO has moved beyond a laboratory trial, but supply redundancy has not been disclosed.
External light sources create a new maintenance object
In pluggable optics, the laser usually fails with the module. CPO concentrates laser-lifetime and source-management issues in an external subsystem. Sharing sources can reduce laser count, but it may also introduce a common failure domain. The power, coupling, monitoring and replacement mechanisms for the 18 sources—and whether they provide N-plus-one redundancy—are critical to real reliability.
Ecosystem and compatibility costs
CPO switches, optical engines, external sources and fiber connectors form a tightly integrated vendor ecosystem. This enables signal-integrity optimization for a specific platform but increases upgrade, cross-vendor compatibility and spare-parts complexity. Operators must verify support across switch firmware, NCCL, congestion control, telemetry and cluster-management tools before migration.
Efficiency gains depend on utilization
The 5× figure is an upper bound and should not be applied to the entire AI factory. Significant absolute savings require sustained high-bandwidth training or high-utilization inference traffic. Procurement comparisons should therefore include watts per bit, networking power per GPU and facility PUE, rather than relying only on optical-component specifications.
Deployment guidance
- Create an application-level baseline. Compare CPO and pluggable systems on equivalent GB300 NVL72 clusters, covering All-Reduce, parameter-server or hybrid-parallel traffic. Record mean latency, P95 and P99 latency, synchronization stalls, job throughput and training completion time.
- Separate energy measurements. Measure the switching ASIC, optical engines, external light sources, retimers, liquid-cooling system and complete switch independently, then convert the results into Tb/s, watt-hours per GPU and cost per training token.
- Run fault injection. Simulate light-source failures, disconnected fibers, port errors, firmware upgrades and switch maintenance. Validate redundant paths, traffic rerouting, collective-communication recovery and the blast radius of agent requests.
- Scale the topology in stages. Validate temperature, signal integrity, fiber-contamination controls and spare-parts procedures in one pod before expanding across multiple pods. The 144-port design offers a high-radix option, but whether it reduces network tiers and hop count depends on the actual topology and traffic distribution.
- Complete the TCO model. Include switch and optics acquisition, liquid-cooling integration, external-source spares, field-replacement labor, failure downtime and next-generation upgrade cycles.
Lambda's deployment demonstrates that CPO can operate inside a 10,000-GPU production network. The available evidence is best interpreted as directional validation rather than independent performance certification. CPO's business case ultimately depends on whether it lowers energy per GPU, reduces downtime from network faults and keeps total cost of ownership within an acceptable range under real training workloads.
— END —Source: NVIDIA AI News (2026-11-16)
Tags: #NVIDIA Photonics #Co-Packaged Optics #Lambda #AI Clusters #GB300 NVL72
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments