Lambda said on November 16 that it has put co-packaged optics networking into production at a scale the industry has talked about for years but rarely shown. The company, which calls itself the Superintelligence Cloud, announced what it describes as the world's first large-scale deployment of co-packaged optics, or CPO, connecting more than 10,000 NVIDIA Blackwell Ultra GPUs at its Quebec campus using NVIDIA Photonics switches. NVIDIA, in its own account of the deployment, said the switches connect GB300 NVL72 racks holding those GPUs.
The change is in where the optics sit. In a conventional switch, pluggable transceivers convert electrical signals to light at the front panel. In a CPO switch, the optical engines sit alongside the switch ASIC on the same package, shortening the electrical path from inches to millimeters. Lambda and NVIDIA say that reduces signal loss, removes the need for retimer chips, and cuts the number of parts that can fail in the signal path.
Each NVIDIA Photonics switch in the deployment has 144 ports running at 800 gigabits per second, liquid cooling, and 18 removable external light sources that supply light to the optical engines. Because those light sources can be replaced in the field, a failed laser does not require replacing an entire transceiver module, a point both companies stress for very large training jobs, where one interrupted link can stall thousands of GPUs.
Lambda said it benchmarked CPO against conventional pluggable optics at rack, scalable unit, and pod configurations. Throughput was identical in every configuration, it said, and signal integrity was more stable, while latency was lower at every scale. The gap grew from 3.8 percent at rack scale to 10.1 percent at pod scale. The CPO optics also ran cooler, at a median temperature of 36 degrees Celsius, compared with 51 degrees for the pluggable optics. NVIDIA says its Photonics design delivers up to five times better power efficiency and up to 15 times higher reliability than conventional pluggable optics, and uses four times fewer lasers. Those are vendor figures for the platform, not independent measurements of the Quebec cluster.
"Networking has to keep pace with the demands of increasingly large AI clusters," said Dave Ward, president of cloud services at Lambda. "Bringing co-packaged optics into a deployment of this scale is an important step toward building the efficient and resilient infrastructure needed to support the next generation of AI." Gilad Shainer, senior vice president of networking at NVIDIA, said that in an AI factory every interruption stalls the job and every watt counts, and that Photonics is meant to give customers more productive GPU hours from their investment.
Lambda paired the networking announcement with the campus around it. It said the Quebec site has 142 megawatts of secured, nearly 100 percent renewable energy capacity, can use free cooling for up to 80 percent of the year, is guaranteed to a power usage effectiveness of 1.2, and uses no process water for server cooling. The deployment, first flagged when Lambda announced early adoption of NVIDIA CPO networking in November 2025, will also be featured in NVIDIA's SC26 special address.
For enterprise and cloud operators, the significance is less the headline number than the direction it points. As clusters grow past ten thousand accelerators, the network, not the GPU, increasingly decides how much of that expensive compute is actually working at any moment. Lambda's results, if they hold in production, suggest CPO is ready to move from roadmap slides into that role.