Lambda Brings Co-Packaged Optics Into a Cluster of More Than 10,000 GPUs
Lambda says it has connected more than 10,000 NVIDIA Blackwell Ultra GPUs at its Quebec campus using switches that bring optical networking closer to the chips handling the traffic.
Published November 17, 2026Intelligence / Frontier Tech / Experimental Systems

Lambda says it has connected more than 10,000 NVIDIA Blackwell Ultra GPUs at its Quebec campus using switches that bring optical networking closer to the chips handling the traffic. Announced November 16, the deployment puts co-packaged optics into the working infrastructure of a large AI cluster. Lambda announcement
The company reports network latency 10.1 percent lower at pod scale than with conventional pluggable optics, while maintaining the same throughput. Latency measures delay; throughput measures how much data the network carries. During a shared computation, communication delays can leave processors waiting for one another. Lambda benchmarks
An interruption can leave expensive processors waiting for the rest of a training job. NVIDIA describes the deployment as an effort to improve both efficiency and the time between network interruptions. The practical prize is more productive computing from equipment already installed. NVIDIA account
“Networking has to keep pace with the demands of increasingly large AI clusters,” said Dave Ward, Lambda’s president of cloud services. The announcement builds on the company’s early-adoption commitment in November 2025. Lambda statement
Moving the conversion closer
Optical fiber carries information as light, while the switching chip processes electrical signals. In a conventional arrangement, those signals travel across a circuit board to optical modules plugged into the front of the switch. At very high speeds, the electrical journey consumes power and complicates the task of preserving a clean signal. NVIDIA technical explanation
Co-packaged optics puts the light-conversion hardware beside the switching chip, shortening that electrical route. The architecture reduces the need for signal-conditioning components along the way. Fiber then carries the data between parts of the cluster. It changes how the processors communicate rather than making the GPUs themselves optical computers. Switch architecture
Lambda’s deployment uses liquid-cooled switches with 144 ports, each operating at 800 gigabits per second. NVIDIA says each switch has 18 removable external light sources. These supply light to the optical engines and allow technicians to replace a light-source module in the field. Deployment hardware
Cooling and maintenance are part of the design
The external lasers serve a different role from the optical engines beside the chip. NVIDIA’s technical account describes keeping light generation in a separately controlled thermal environment to limit wavelength drift and premature aging. Modular replacement also makes a failed light source easier to service without disturbing the surrounding switch infrastructure. External laser design
Lambda reports a median optics temperature of 36°C, compared with 51°C for its conventional pluggable-optics comparison. Its published results cover networking performance and component temperatures; whole-campus electricity savings and measured end-to-end training speedups remain unquantified in the announcement. Reported results
The engineering work extends into manufacturing. NVIDIA describes coordinating photonic circuits, electronics, fiber alignment and thermal control with partners including TSMC. The technology’s move into a large cluster brings those packaging decisions into day-to-day operations, where serviceability and consistent signal quality matter alongside headline bandwidth. Manufacturing background
Primary sources
Sources · 4
- https://lambda.com/blog/worlds-first-large-scale-deployment-of-nvidia-photonics-switches-with-cpo
- https://blogs.nvidia.com/blog/lambda-photonics/
- https://developer.nvidia.com/blog/scaling-ai-factories-with-co-packaged-optics-for-better-power-efficiency/
- https://developer.nvidia.com/blog/how-industry-collaboration-fosters-nvidia-co-packaged-optics/