Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics
AI-assisted digest, checked against the original source before publishing. How we work
Lambda is wiring more than 10,000 GB300 GPUs with NVIDIA's co-packaged optics switches, reporting lower latency and cooler optics than pluggable transceivers.
Lambda is deploying NVIDIA Photonics co-packaged optics (CPO) switches to connect its NVIDIA GB300 NVL72 racks in an AI factory spanning more than 10,000 GPUs. The cloud provider is positioning the optics, rather than compute, as the lever for network power draw, resiliency and mean time between interruptions at cluster scale.
In Lambda's tests with the switches, network latency dropped by about 4% at rack scale and by up to about 10% at pod scale, while throughput held steady against conventional pluggable optics across rack, scalable unit and pod configurations. "Networking has to keep pace with the demands of increasingly large AI clusters," said Dave Ward, president of cloud services at Lambda, calling the co-packaged optics deployment an important step toward infrastructure for the next generation of AI.
The hardware differs from conventional designs by placing optical engines on the switch package alongside the chip, bringing electrical-to-optical conversion closer to the switch die. NVIDIA cites up to 5x better power efficiency and up to 15x higher reliability than pluggable optics, plus 4x fewer lasers and therefore fewer components that can fail. Each switch carries 144 ports running at 800Gb/s, with liquid cooling pulling heat from the switch chip and optical engines; Lambda reports a median optics temperature of 36°C against 51°C for pluggable optics, with more stable signal integrity in its tests.
The stakes sit in what a single interruption costs: in a large training job, one broken link can stall thousands of GPUs, and in agentic inference a disruption can cascade latency across many agents. Every watt spent networking is a watt unavailable for compute. Practically, the design changes servicing too — each switch has 18 removable external light sources that technicians can swap in the field, whereas a failed laser in pluggable optics typically means replacing the whole module.
What to watch: whether the reported latency and thermal gains hold as clusters grow beyond 10,000 GPUs, how field replacement of light-source modules behaves in production SLAs, and whether co-packaged optics becomes the default for GB300-class deployments rather than a specialist choice. Lambda's throughput parity with pluggables will also be tested as pods scale up.
Key facts
- • Lambda cluster spans 10,000+ NVIDIA GB300 NVL72 GPUs.
- • Latency fell ~4% rack scale, up to ~10% pod scale.
- • Each Photonics switch: 144 ports at 800Gb/s, liquid-cooled.
- • Median optics temperature 36°C vs 51°C pluggables.
- • CPO claims 5x power efficiency, 15x reliability, 4x fewer lasers.
Why it matters
Network power and downtime become first-order constraints on cluster capacity, so builders sizing AI factories should budget for optics thermals and failure domains alongside GPU count. The field-replaceable light-source design shifts day-two operations toward module swaps instead of transceiver replacement, and CPO's adoption path likely depends on whether these latency and reliability gains repl
Spotted an error? Send a correction.
Get the day's AI news in one email
The Daily Epoch · 60-second read · unsubscribe anytime