The first production-scale test of co-packaged optics in an AI factory landed this week. Lambda announced what it calls the world's first large-scale CPO deployment, connecting more than 10,000 NVIDIA Blackwell Ultra GPUs with NVIDIA Photonics switches at its Quebec campus, and NVIDIA's companion post frames the same project around resiliency and watts S2. Read alongside Vera Rubin's debut MLPerf results S1, the announcements sketch the network's changing job: compute keeps getting more efficient, so the switch's task is to stop wasting it.

Throughput parity was the point

Lambda benchmarked its CPO switches against conventional pluggable optics across rack, scalable unit and pod configurations. Throughput held identical at every scale, with lower latency — a gap that grew from 3.8 percent at rack scale to 10.1 percent at pod — plus more stable signal integrity and a median optics temperature of 36°C versus 51°C for pluggables S3. The switches are 144-port, 800Gb/s parts with liquid cooling and 18 removable external light sources each, and the host campus is specced at 142 MW of secured, nearly 100% renewable capacity, free cooling up to 80% of the year, and a guaranteed PUE of 1.2 S3.

The parity is the story. If CPO cost bandwidth, operators would wait a generation. Matching pluggables while running cooler removes that objection and moves the buying decision onto different ground: power and interruptions.

The reliability claim, and what is actually measured

NVIDIA claims co-packaged optics deliver up to 5x better power efficiency and up to 15x higher reliability than pluggable transceivers, while using 4x fewer lasers; mounting the optical engines millimeters from the switch ASIC shortens the electrical path from inches, drops the retimers, and lets technicians swap a field-replaceable light-source module instead of a whole transceiver. Its framing of the stakes: in a large training job, one interrupted link can stall thousands of GPUs S2.

What has actually been published is narrower. Lambda's benchmarks cover throughput, latency, signal integrity and optics temperature — not field uptime or a measured MTBI figure. The 5x and 15x multiples are vendor claims: the rollout proves CPO runs at scale without a bandwidth penalty, not that it has a demonstrated interruption record. That sets a concrete ask for the next buyer — demand an MTBI number from your own pilot before treating reliability multiples as fact. The operational shifts are real either way: spares planning moves from transceivers to light sources, and watts saved on optics flow back to compute.

Compute keeps raising the bar

The context is Vera Rubin's first MLPerf round. NVIDIA reports up to 2x per-GPU performance over GB300 NVL72 in MLPerf Training v6.1, about 97% scaling efficiency from 256 to 1,024 GPUs on DeepSeek-V3, and — in a provisional result still awaiting MLCommons verification — more than 10x throughput per watt versus GB300 on agentic inference in its first MLPerf Endpoints submission S1.

That trajectory sharpens the network's role. Every gain in compute efficiency raises the cost of an idle link, because the machine being stalled is worth more per watt every generation. Vendors will keep quoting port counts and terabits, but the spec that will separate operators at 10,000 GPUs and beyond is mean time between interruptions — and CPO's first large-scale reference is built around that claim. The field data that confirms or breaks it is the next result worth watching.

Sources and editorial method

This AI-assisted report synthesizes the sources below. Automated source checking is not independent human verification; follow the original reporting for context and updates.

Common questions

What is co-packaged optics in AI data center networking?

CPO mounts optical engines on the switch package next to the ASIC instead of at the front panel, shortening the electrical path from inches to millimeters. That cuts power, removes retimer chips and reduces laser count. Lambda now runs NVIDIA Photonics CPO switches across a 10,000+ GPU deployment in Quebec.

Does CPO make AI cluster networks faster and more reliable?

Lambda's benchmarks matched pluggable optics on throughput and showed lower latency — 3.8 percent at rack scale, up to 10.1 percent at pod — with optics at 36°C versus 51°C. The reliability claims, up to 5x power efficiency and 15x higher reliability, are NVIDIA vendor figures; no field MTBI data has been published yet.