Physical AI News (Japan)

Physical AI, robotics and compute — reported from Japan.

日本語
· 日本語で読む

Lambda connects GB300 NVL72 racks with NVIDIA CPO switches; more than 10,000 GPUs

NVIDIA said Nov. 16, 2026, that Lambda is using NVIDIA Photonics CPO switches. The switches connect NVIDIA GB300 NVL72 racks equipped with more than 10,000 GPUs. NVIDIA cited improved network resilience and efficiency, but provided no performance figures.

Key facts

  • NVIDIA said Nov. 16, 2026, that Lambda is using NVIDIA Photonics CPO switches.
  • The switches connect NVIDIA GB300 NVL72 racks equipped with more than 10,000 GPUs.
  • NVIDIA described the effort as aimed at improving AI factory resilience and efficiency.

What happened

NVIDIA announced Nov. 16, 2026, that cloud provider Lambda is using NVIDIA Photonics co-packaged optics (CPO) switches. The switches connect NVIDIA GB300 NVL72 racks equipped with more than 10,000 GPUs. NVIDIA described the deployment as an effort to improve network resilience and efficiency.

Why it matters

Lambda's configuration uses CPO switches to connect GB300 NVL72 racks equipped with more than 10,000 GPUs. NVIDIA says the goal is to improve resilience and efficiency, making this an example of how network design is also a consideration for operators building and running large-scale GPU environments.

Analysis

The announcement focuses not on GPU computing performance but on the network connecting GB300 NVL72 racks containing more than 10,000 GPUs. NVIDIA links the use of CPO switches to AI factory resilience and efficiency, but the announcement provides no comparative figures for latency, bandwidth, power consumption or behavior during failures. What can be confirmed for now is the adoption and its stated aims, not the extent of any measured performance improvement. The computing infrastructure can be broken down into components: GPU racks provide processing resources, CPO switches connect the racks, and Lambda's cloud environment is the means of access. At a scale of more than 10,000 GPUs, the network linking the systems, as well as the number of compute systems, becomes a factor in the overall service's availability and efficiency. However, the announcement does not specify the number of switches deployed or the network configuration. It also does not explain how fault tolerance will be ensured or whether "efficiency" refers to power, cost or communications performance. Articles published by this newspaper on the same day covered NVIDIA's new GPU for science and engineering, Rubin HPC and expansion of its CUDA-X libraries. These developments concern computing hardware and software, respectively, and there is no information linking them directly to Lambda's adoption of CPO switches. Taken together, the news illustrates the need to treat inter-rack networking as a distinct design element in AI infrastructure, rather than evaluating it solely by its compute resources. The broader industry implication is that expanding GPU supply alone does not explain all the requirements for deploying large-scale computing infrastructure. Cloud providers must include the choice of connectivity and network availability in their design considerations, in addition to procuring and operating GPU racks. At the same time, this announcement alone does not show under what conditions CPO is advantageous compared with conventional approaches or how it might change the requirements for deployment and maintenance. Points to watch include whether details are disclosed on the actual switch configuration, performance and power measurement conditions, operations during failures, and the racks covered.

Details

According to an NVIDIA blog post, Lambda is using NVIDIA Photonics CPO switches for the network that links multiple GPUs and connects GB300 NVL72 racks. The blog describes the GPU count as "more than 10,000." The text provided does not specify the number of switches, the configuration, the deployment date or quantitative performance figures for resilience or efficiency.

Sources

Translated from the Japanese original: /news/lambda-nvidia-no-cpo-suitchi-de-gb300-nvl72-rakku-wo-setsuzo-742a0a. The Japanese version is authoritative if they differ.