On November 16, 2026, AI infrastructure company Lambda announced a large-scale deployment of NVIDIA Photonics switches in the network connecting more than 10,000 NVIDIA Blackwell Ultra GPUs at its data center in Quebec, Canada. The switches use CPO (Co-Packaged Optics) technology, which places the switch chip and the optical engines in the same package. In Lambda's comparison tests, network latency fell by 3.8% at rack scale and 10.1% at pod scale versus a conventional pluggable optical module setup, while throughput (the amount of data transferred per unit of time) stayed the same.
Lambda announced its early adoption of CPO technology in November 2025, and it has now disclosed the scale of the deployment and the results of its performance comparison. The equipment in question is not the GPUs' compute processing itself but the networking gear that handles communication between racks holding large numbers of GPUs. In large-scale AI systems, overall performance depends not only on GPU compute power but also on how efficiently data can move between GPUs, which makes improvements in network latency and power efficiency important.
Latency down by up to 10.1% at the same throughput, with optical components running 15°C cooler
According to Lambda's official announcement, the company compared conventional pluggable optical modules with CPO at three scales: rack, scalable unit and pod. A scalable unit is a building block that sits between the rack and the pod, and a pod is a group of multiple racks. In the tests, signal quality remained stable and network latency fell at every scale while throughput was maintained. The median temperature of the optical components was 36°C with CPO and 51°C with the conventional setup, a difference of 15°C.
The table below organizes Lambda's measurements alongside the design characteristics and product performance figures published by NVIDIA for CPO technology.
| Item | Result / published spec | Source type |
|---|---|---|
| Network latency | Down 3.8% at rack scale and 10.1% at pod scale at the same throughput | Lambda comparison test |
| Median optical component temperature | 36°C with CPO, 51°C with conventional setup | Lambda comparison test |
| Electrical signal travel distance | Reduced from inches to millimeters | Lambda and NVIDIA design explanation |
| Power efficiency | Up to 5x | NVIDIA product description |
| Reliability | Up to 15x | NVIDIA product description |
It is important to distinguish what Lambda actually measured from the performance figures NVIDIA cites as product features. The latency reductions and optical component temperatures come from Lambda's comparison tests. The up to 5x power efficiency and up to 15x reliability, on the other hand, are figures NVIDIA presents for the product in general, and they were not measured in Lambda's environment.
Lambda's announcement does not give details such as absolute latency values, the specific workloads used, the number of measurements or margins of error. It is therefore not possible to directly calculate from the reported latency reductions how much AI model training time or effective GPU compute performance would improve. Still, the result of lower network latency and cooler optical components at the same throughput is a useful reference for evaluating the effects of CPO in a real deployment.
CPO moves the conversion to optical signals closer to the switch chip
In the conventional pluggable optical module approach, electrical signals travel from the switch ASIC to optical modules attached to the front of the equipment, where they are converted to optical signals. CPO, by contrast, places the optical engines that handle this conversion in the same package as the switch ASIC. The major difference is that the distance electrical signals must travel is greatly shortened.
According to NVIDIA's technical explanation, CPO shortens the electrical signal path from inches to millimeters. This reduces signal loss and eliminates the need for retimer chips, which reshape signals degraded over long electrical traces. With fewer components needed in the signal path, power consumption in that portion can be reduced. The stable signal quality and lower optical component temperatures Lambda reported are likely related to this structural difference.
The NVIDIA Photonics switches deployed here each have 144 ports of 800Gb/s and are liquid-cooled. According to NVIDIA, the cooling system is designed to remove heat not only from the switch ASIC but also from the optical engines in the same package. The design concentrates the components needed for high-speed optical communication near the chip while managing the heat they generate through liquid cooling.
Light sources placed externally, with replacement in case of failure in mind
In CPO, the optical engines sit near the switch ASIC, but the laser light sources that supply the light are installed outside the package. According to NVIDIA, each switch has 18 removable external light sources, and field technicians can replace a light source module if it fails. In the conventional pluggable approach, a laser failure requires replacing the entire optical module, whereas CPO separates the optical engines from the replaceable light sources.
This configuration takes maintainability in large-scale AI data centers into account. In environments where many switches are operated over long periods, replacement work and equipment downtime caused by light source failures can become an operational burden. NVIDIA also says CPO can reduce the number of lasers required to one-fourth of that needed in conventional approaches. Combining fewer light sources with a replaceable design is intended to improve the reliability and maintainability of the communication equipment.
The Quebec data center where the switches were deployed is also a facility built with large-scale AI system operation in mind. According to Lambda, it has 142MW of power capacity and runs on nearly 100% renewable energy. It also uses outside-air free cooling for up to 80% of the year and guarantees a PUE (Power Usage Effectiveness, a measure of a facility's overall power efficiency) of 1.2. These are conditions of the data center as a whole and should be distinguished from any power-saving effect of the CPO switches alone.
The comparison tests showed that CPO reduced network latency and optical component temperature. In deciding whether to adopt it for large-scale AI infrastructure, the absolute latency values and power consumption under real AI workloads will matter, as will failure rates and light source replacement frequency over long-term operation. Once such measured data is available, it will be possible to assess concretely how far CPO can improve not just communication performance but also data center operating costs and maintenance burden.
