Moving the optics closer to the switch chip changed latency without changing throughput.
Lambda’s November 16 deployment report describes NVIDIA Photonics switches connecting more than 10,000 Blackwell Ultra GPUs in Quebec. Its benchmarks compared co-packaged optics (CPO) with conventional pluggable optics across rack, scalable-unit and pod configurations.
Lambda reports identical throughput, with latency reductions growing from 3.8% at rack scale to 10.1% at pod scale. Median optical temperature was 36°C versus 51°C.
The mechanism is physical. A pluggable module converts electrical signals to light at the front panel. CPO puts that conversion alongside the switching chip, shortening the electrical path and reducing signal loss. Fiber carries the signal onward.
Those are useful deployment measurements. The announcement also claims up to 15 times higher reliability, but doesn’t give the observation period or failure counts needed to evaluate that claim. Cooler optics alone don’t establish a measured fleet failure rate.
For a fabric review, I’d ask for latency distributions under the same traffic load, optical error counters, and the replacement procedure when an optical engine fails. That last question matters when the optics move beside the chip. Lambda identifies removable external light sources, but the announcement doesn’t describe optical-engine repair.


