Lifestyle
Lambda links 10,000+ GB300 GPUs with NVIDIA Photonics CPO switches; NVIDIA claims up to 5x better power efficiency
NVIDIA says cloud provider Lambda uses NVIDIA Photonics co-packaged optics switches to connect GB300 NVL72 racks across more than 10,000 GPUs. Latency and temperature figures come from company testing cited in NVIDIA's blog.
News date: About 4 min read
Contents
In a blog post dated November 16, 2026, NVIDIA said cloud provider Lambda is using NVIDIA Photonics co-packaged optics (CPO) switches in a GB300 NVL72 deployment spanning more than 10,000 GPUs. GB300 NVL72 is NVIDIA's rack-scale GPU system. CPO means placing the parts that turn electrical signals into light in the same package as the switch chip. The post does not state a specific rollout stage or availability scope. The content reflects NVIDIA's and the company's own statements and in-house test results.
What changes with this deployment? How CPO differs from traditional pluggable optics
The traditional approach plugs separate optical transceiver modules into switch ports, known as "pluggable optics." According to NVIDIA, NVIDIA Photonics replaces these pluggable transceivers with co-packaged optical engines on the switch package. The optical engine sits in the same package as the chip and converts electrical signals to light, which travels over fiber between parts of the cluster. NVIDIA says co-packaging shortens the electrical path from inches to millimeters. That reduces signal loss and removes the need for retimer chips (relay chips that regenerate electrical signals over long distances), which cuts power draw and the number of components in the signal path.
| Aspect | NVIDIA Photonics CPO | Traditional pluggable optics |
|---|---|---|
| Power efficiency | Up to 5x (per NVIDIA) | Baseline |
| Reliability | Up to 15x (per NVIDIA) | Baseline |
| Number of lasers | Reduced to 1/4 (per NVIDIA) | Baseline |
| Median optics temperature (company-reported) | 36°C | 51°C |
| If a laser fails | Light source module replaceable in the field | Entire module must be replaced |
| Throughput | Same as pluggable | Baseline |
What does it mean in practice for AI training and inference?
NVIDIA notes that in large training jobs, a single broken link can stall thousands of GPUs. In agentic inference, where many AI agents run tasks, one network interruption can cascade into delays across many agents. NVIDIA says the network is central to AI factory productivity, and every watt spent on networking is a watt not available for compute.
In company testing, network latency fell about 4% at rack scale and up to about 10% at pod scale. A rack is a single-cabinet GPU system, while scalable units and pods are progressively larger groups made up of multiple racks. Across rack, scalable unit and pod configurations, throughput matched traditional pluggable optics. Dave Ward, president of cloud services at Lambda, said: "Networking has to keep pace with the demands of increasingly large AI clusters."
What are Lambda's configuration specs and limits?
According to NVIDIA, the switches Lambda uses each have 144 ports at 800Gb/s per port, and liquid cooling carries heat away from the switch chip and optical engines. Each switch has 18 removable external light sources that supply the optical engines, and technicians can replace these light source modules in the field. The post also mentions that company testing found more stable signal integrity.
Frequently asked questions
Is this technology generally available yet?
NVIDIA's blog only describes Lambda's deployment. It does not state general availability, plans or regional scope, so it does not tell you whether the technology is generally available.
Will CPO switches make the network faster?
Per NVIDIA, throughput is the same as pluggable optics. In company testing, latency fell about 4% at rack scale and up to about 10% at pod scale.
What happens if a laser fails?
NVIDIA says technicians can replace the light source module in the field. With traditional pluggable optics, the whole module has to be replaced.
Browse the latest news in this topic
Found this guide helpful? Share it with someone who needs it
Lifestyle
AMD Says Lux Supercomputer Is Open to Scientific Users at Oak Ridge, Ranked No. 14 on TOP500 and Built on MI355X GPUs
AMD announced that the AMD-powered, HPE-built Lux supercomputer is now open to scientific users at the US Department of Energy's Oak Ridge National Laboratory and is the first dedicated AI Factory for the Genesis Mission. AMD encourages interested users to contact OLCF.
Lifestyle
Google report: attackers use AI to find vulnerabilities from software updates and build exploits; Google introduces AI Threat Defense
Google Threat Intelligence Group (GTIG) said on November 16, 2026 that its AI Threat Tracker report found attackers using autonomous AI frameworks to analyze software updates, identify patched flaws and build exploits. Google also introduced AI Threat Defense, combining Gemini, Wiz, CodeMender and Mandiant, but gave no availability stage or pricing.
Lifestyle
NVIDIA Unveils Rubin HPC GPU: Over 50,000 CUDA Cores per Chip, Expected in 2028
At SC26, NVIDIA announced Rubin HPC, a new member of the Vera Rubin platform aimed at scientific, engineering and analytics simulation. Each GPU has over 50,000 CUDA cores, with availability expected in 2028.
Lifestyle
Google Cloud Publishes MLPerf Training v6.1 Results: 8,192 Ironwood TPUs Hit Target in 1.70 Minutes, Plus the First Agentic Reinforcement Learning Test
Google Cloud says it submitted 10 results to the MLPerf Training v6.1 AI benchmark using Ironwood TPUs and NVIDIA Grace Blackwell GPUs, covering pretraining and the first agentic reinforcement learning benchmark. This is a results announcement, not a product launch; all figures are Google's own.
Sources
Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.