> ## Content Index
> Fetch the complete content index at: https://www.notatechguy.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Lambda deploys NVIDIA Photonics across 10,000 GPUs
- URL: https://www.notatechguy.com/lambda-deploys-nvidia-photonics-across-10-000-gpus/
- Published: 2026-11-17T12:49:56.000Z
- Updated: 2026-11-17T12:49:56.000Z
- Description: Lambda's 10,000-GPU NVIDIA Photonics deployment cuts pod latency up to 10%, but every figure is vendor-tested with no independent verification.
- Author: Marcello Babbili
- Tags: Technology & AI, Nvidia, AI Chips, OpenAI

Lambda has connected more than 10,000 GPUs in NVIDIA GB300 NVL72 racks to NVIDIA Photonics co-packaged optics switches, reducing network latency by up to 10% in tests run by the vendor and its customer [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). Every performance figure in the announcement comes from NVIDIA and Lambda, with no independent benchmark, and several carry "up to" qualifiers that mark them as best-case results [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

**My read:** The temperature number is the one I find most credible, because it is a concrete median (36°C against 51°C) rather than a multiplier with an "up to" hedge. The 5x power and 15x reliability claims are vendor marketing figures with no methodology disclosed, and I would not repeat them as established fact. What matters here is that someone is actually running co-packaged optics at 10,000-GPU scale and reporting back, even if the reporting is self-interested. The latency drop, 4% at rack scale and 10% at pod scale, is modest but real if the test setup is honest, and it points to where CPO pays off: not inside a single rack, but across the pod where data has farther to travel.

### Co-packaged optics shrinks the electrical path and removes retimer chips

Conventional AI factory networks rely on pluggable optical transceivers mounted at the edge of each switch port. The electrical signal must travel inches across a circuit board from the switch chip to the transceiver, degrading along the way, and retimer chips are placed in that path to restore the weakened signal [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

Co-packaged optics places the laser and optical engine directly beside the switch chip, reducing the electrical path from inches to millimetres [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). That removes the need for retimer chips, lowering both power consumption and the component count in the signal path [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). NVIDIA's design uses 4x fewer lasers than conventional pluggable optics, reducing the number of parts that can fail [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

Each switch in Lambda's deployment has 144 ports running at 800Gb/s, with 18 removable external light sources supplying the optical engines [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). If a laser fails, a technician swaps the light-source module in the field. With traditional pluggable optics, a failed laser means replacing the entire module [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

Liquid cooling carries heat away from the switch chip and optical engines, and Lambda reports a median optics temperature of 36°C, compared with 51°C for conventional pluggable optics [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). This temperature figure is Lambda's own measurement, not independently verified [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

![Median optics temperature in Lambdas deployment](https://storage.ghost.io/c/6e/89/6e896869-22ef-4281-a213-b4c462c17cff/content/images/2026/11/chart_8c771dc70487af775f25.png)

### Latency drops modestly at pod scale, throughput holds steady

Lambda's internal tests, reported on NVIDIA's blog on November 16, found network latency dropped by about 4% at rack scale and by up to about 10% at pod scale [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). These results were produced by Lambda and published on NVIDIA's blog, with no third-party benchmark, no released test methodology, and no error bars [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). Across rack, scalable unit and pod configurations, throughput matched conventional pluggable optics [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). The tests also found more stable signal integrity with the CPO switches [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

NVIDIA is promoting its own photonics technology through a customer case study on its own blog, and no financial terms, contract value, or deployment timeline were disclosed [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com). The "up to 5x better power efficiency" and "up to 15x higher reliability" figures are vendor claims with no published methodology [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

Dave Ward, president of cloud services at Lambda, said networking has to keep pace with the demands of increasingly large AI clusters, and called the deployment an important step toward building the infrastructure needed to support the next generation of AI [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

### Cluster operators face new swap procedures and undisclosed pricing

Operators running large GPU clusters where network latency and component failure rates eat into utilisation should pay attention. A team running a 5,000-GPU training job that gains 4% lower network latency at rack scale is saving real compute hours, but only if the CPO switches are available at a price that beats the pluggable optics they replace. No pricing has been disclosed.

The field-replaceable light sources are the detail a data centre manager would check first. If a laser can be swapped in minutes rather than requiring a full module replacement, the mean time between interruptions goes up, and that is the metric that determines whether a training run finishes or restarts from a checkpoint.

A data centre operations lead managing a multi-rack GPU cluster would now track laser-module swap times as a new maintenance metric, since field-replaceable light sources shift the repair workflow from full-module replacement to component-level swaps [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

NVIDIA has not said when other cloud providers will deploy CPO at similar scale, and Lambda has not disclosed whether the 10,000 GPUs sit in a single data centre or span multiple regions [S¹](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com).

---

*Sources: [S1 — Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photoni](https://blogs.nvidia.com/blog/lambda-photonics/?ref=notatechguy.com) · [P2 — AI Factory Archives | NVIDIA Blog](https://blogs.nvidia.com/blog/tag/ai-factory/?ref=notatechguy.com) · [P3 — Open Source Projects, Technologies, and Organizations | NVIDIA Open S](https://opensource.nvidia.com/en-us/?ref=notatechguy.com) · [P4 — lambda (Lambda)](https://huggingface.co/lambda?ref=notatechguy.com)*

---

*Written from 4 sourced items, 4 of them primary.*

## More from Not A Tech Guy

- [Open-source AI infra book hits 7,869 GitHub stars](https://www.notatechguy.com/open-source-ai-infra-book-hits-7-869-github-stars/)
- [AI catastrophic forgetting fix retains 83% across five tasks](https://www.notatechguy.com/ai-catastrophic-forgetting-fix-retains-83-across-five-tasks/)
- [OpenAI doubles free ChatGPT for researchers to 200,000](https://www.notatechguy.com/openai-doubles-free-chatgpt-for-researchers-to-200-000/)