The development
What happened
Lambda announced a co-packaged-optics network connecting more than 10,000 NVIDIA Blackwell Ultra GPUs at its Quebec campus on 16 November. The AI cloud operator says its tests found lower network latency and cooler optics than conventional pluggable designs, without increasing measured throughput.
Key facts
Co-packaged optics places optical engines beside the switch chip rather than relying on conversion at the front panel. Shorter electrical paths remove the need for retimer chips; the optics are on the switch, not inside the AI model.
The deployment uses NVIDIA Photonics switches with 144 ports running at 800 Gb/s, liquid cooling and 18 removable external light sources. Lambda identifies the connected hardware as Blackwell Ultra GPUs at its Quebec campus.
Lambda reports network-latency reductions of 3.8% at rack scale and 10.1% at pod scale in its own tests, while throughput stayed the same as conventional pluggable optics. These figures are not whole training-job speedups.
The operator reports median optics temperatures of 36°C versus 51°C for the comparison design. This is first-party testing; TideReach has not independently measured the hardware, long-term failure rates or completed AI workloads.
The context
Why it matters
TideReach assessment: large AI jobs depend on communication among GPUs, so the network can affect useful compute time as well as raw capacity. The deployment is a concrete infrastructure report, but lower link latency or cooler optics should not be read as the same percentage improvement in model training, agent response times or operating costs. Those outcomes need workload-specific evidence.