Lambda scales AI factories with NVIDIA Photonics CPO switches

Lambda is deploying NVIDIA Photonics co-packaged optics (CPO) switches to interconnect NVIDIA GB300 NVL72 racks containing over 10,000 GPUs as of 16 November 2026. (co-packaged optics (CPO): optics on the same chip package that shorten the electrical path from inches to millimeters.) The switch architecture provides up to 5x better power efficiency and 15x higher reliability than conventional pluggable optics. (pluggable optics: conventional optical components that are inserted into a port rather than being integrated onto the chip.) This transition is critical for cloud providers aiming to minimize network-induced stalls in massive training jobs.

Is this for you?

AI Infrastructure Engineers and Data Center Managers benefit from 5x better power efficiency and 15x higher reliability when deploying NVIDIA Photonics CPO switches.

  • If you are an AI Infrastructure Engineer, these switches offer a path to higher MTBI and lower latency for clusters with 10,000+ GPUs.
  • If you are a Data Center Facility Manager, these switches provide a 15°C reduction in optics temperature to improve power usage effectiveness.
  • If you are a Network Hardware Architect, you can see a reduction in retimer chips when moving to an 800Gb/s link.
  • Skip this if you are a home PC builder as this technology focuses on data center scale.
  • Skip this if you are a software developer looking for AI coding guidance.

What you can do now

  1. If you manage a data center, update the thermal management plan to account for the 36°C median optics temperature.
  2. If you are a network hardware architect, verify that NVIDIA Photonics complies with your required ethernet standard.
  3. If you are an infrastructure engineer, verify the fiber optic cable specifications required to connect NVIDIA Photonics switches.

Key figures

Power efficiency
Up to 5x better than pluggable optics
Reliability (MTBI)
Up to 15x higher than pluggable optics
Switch ports
144 ports at 800Gb/s
Optics temperature
36°C (CPO) vs 51°C (pluggable)
Latency reduction
4% at rack scale; 10% at pod scale

NVIDIA Photonics delivers 5x better power efficiency and 15x higher reliability

NVIDIA claims that NVIDIA Photonics addresses both power efficiency and resiliency by replacing pluggable transceivers with co-packaged optical engines on the switch package. Lambda reported a median optics temperature of 36°C for these switches, which is 15°C cooler than the 51°C recorded for conventional pluggable optics. This lower temperature leads to more stable signal integrity—the quality and reliability of the electrical/optical signal.

Latency drops by 4% at rack scale and 10% at pod scale

In Lambda’s tests, network latency—the time needed for moving data—dropped by about 4% at rack scale and by up to about 10% at pod scale. The network maintains the same throughput as conventional pluggable optics across all configurations. This confirms that co-packaging does not trade off speed for efficiency.

Co-packaged optics remove the need for retimer chips

Co-packaged optics (CPO)—optics on the same chip package—shorten the electrical path from inches to millimeters. This reduction in physical distance reduces the need for retimer chips, which are signal boosting chips used to clean and re-transmit signals. Removing these extra parts cuts both power draw and the complexity of the hardware path.

Removable light sources allow for easier field maintenance

NVIDIA Photonics uses 4x fewer lasers than conventional designs, which reduces the number of failure-prone components. Each switch includes 18 removable external light sources that supply light to the internal optical engines. Technicians can replace these individual light-source modules in the field, whereas traditional pluggable optics require replacing the entire module if a laser fails.

Who benefits from NVIDIA Photonics CPO switches

AI Infrastructure Engineers should prioritize these switches because they provide a path to higher MTBI—the mean time between interruptions—and lower latency for clusters with 10,000+ GPUs. Data Center Facility Managers should monitor the 15°C reduction in optics temperature to update their thermal management plans and improve power usage effectiveness. Network Hardware Architects can compare the bill of materials to see the reduction in retimer chips when moving to an 800Gb/s link.

Why home PC builders should skip this

Home PC builders should skip this technology as it focuses on data center scale rather than individual workstations; read our standard GPU and networking hardware reviews instead. Software developers looking for AI coding guidance should look at AI model training documentation or software framework guides instead of hardware interconnects.

A scenario for evaluating CPO: Mateo’s cluster

Mateo runs an infrastructure team for a startup building a 1,000-GPU cluster and needs to decide between standard pluggable optics and CPO. Mateo notes that his current pluggable optics run at 51°C. He sees that NVIDIA Photonics CPO switches run at 36°C and offer 15x higher reliability. He also notes that CPO reduces latency by 4-10% depending on scale. Mateo selects the NVIDIA Photonics CPO switches to ensure higher MTBI and better power efficiency for his cluster. If Mateo only needed a small 16-GPU cluster, the scale-dependent benefits (like the 10% pod-scale latency drop) would be less critical, and he might stick with pluggable optics for ease of initial setup.

Practical applications for NVIDIA Photonics

Use case Who it serves
Calculate power savings for an 800Gb/s interconnect Hardware procurement teams
Estimate MTBI for large-scale AI clusters Site reliability engineers
Plan thermal layout for GPU racks Hardware designers
Design a system that avoids retimer chips PCB designers
Reduce latency in pod-scale configurations Network architects

CPO replaces pluggable transceivers and retimer chips

NVIDIA Photonics replaces pluggable transceivers with integrated optical engines, offering 5x better power efficiency and 15x higher reliability. It also reduces the need for retimer chips in the signal path, which reduces the total part count and power draw. Finally, it reduces the need for full module replacements during failure, as technicians can swap out the 18 individual light sources instead.

What remains unverified about the technology

The source does not state the strongest argument against the claim, nor does it provide the specific conditions that would prove the claim wrong. Additionally, the source did not measure or disclose how NVIDIA Photonics impacts total training job duration. These points remain unknown at this time.

NVIDIA Photonics: what to check, compare and watch

Fiber optic cable
NVIDIA Photonics carries data over fiber between parts of the cluster. AI Infrastructure Engineer: verify the fiber optic cable specifications required to connect NVIDIA Photonics switches.
Pue
NVIDIA Photonics improves power efficiency, which helps lower the overall power usage effectiveness. Data Center Facility Manager: compare the projected pue with and without NVIDIA Photonics in the data center.
Thermal management plan
NVIDIA Photonics requires liquid cooling to carry heat away from the switch chip and optical engines. Data Center Facility Manager: update the thermal management plan to account for the 36°C median optics temperature of NVIDIA Photonics.
Ashrae thermal guidelines
NVIDIA Photonics requires specific cooling environments governed by industry standards like ashrae. Data Center Facility Manager: check if NVIDIA Photonics cooling requirements meet ashrae thermal guidelines.
Liquid cooling system
Liquid cooling is used to carry heat away from the switch chip in NVIDIA Photonics deployments. Data Center Facility Manager: configure the liquid cooling system to support the heat output of NVIDIA Photonics switches.
Ethernet standard
NVIDIA Photonics switches must adhere to standard networking protocols like ethernet. Network Hardware Architect: verify that NVIDIA Photonics complies with the required ethernet standard for the network.

NVIDIA Photonics: questions the original report leaves open

What is the strongest argument against it?
The original report did not say what the strongest argument against the claim is.
What would prove the claim wrong?
The original report did not say what would prove the claim wrong.
How does NVIDIA Photonics impact training job duration?
The original report did not say how NVIDIA Photonics impacts training job duration.

Reported by NVIDIA on 16 November 2026: Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics.

Related articles

ExtremeSpec is reader-supported: some links on this page are affiliate links, and as an Amazon Associate ExtremeSpec earns from qualifying purchases at no extra cost to you. Learn more.

Similar Posts