Lambda's NVIDIA CPO Cluster: 10,000+ GPUs, Pod Latency 10.1% Lower
Lambda published its first benchmarks for co-packaged optics across 10,000+ Blackwell Ultra GPUs in Quebec; NVIDIA says CPO cuts optics to as low as 9 W per interface against often 30 W.
4 min read

An early NVIDIA Quantum-X Photonics CPO switch sample in a Lambda rack, from Lambda's June 2026 blog post. It is not a photo of the Quebec installation.
Image: Lambda
Source · Press image from the Lambda blog, used for news reporting.
Lambda said on Nov. 16, 2026 that co-packaged optics (CPO) switches connecting more than 10,000 NVIDIA Blackwell Ultra GPUs at its Quebec campus cut network latency by up to 10.1 percent against conventional pluggable optics, in benchmark results it published in a Business Wire release. Lambda calls the cluster the world's first large-scale CPO deployment.
The scale itself is not new. In March 2026 Lambda said it was deploying NVIDIA Quantum-X800 InfiniBand Photonics CPO in an AI factory with more than 10,000 Blackwell Ultra GPUs, building on an early-adoption announcement in November 2025. That release named no site, and the November release does not say the two are the same cluster. What is new is the location and the first measured results at that scale.
Lambda said it tested the switches against pluggable optics at rack, scalable-unit and pod scale. Throughput was identical, it said, while latency was lower at every scale, with the gap growing from 3.8 percent at rack level to 10.1 percent at pod level. The company did not say what the latency baseline was. The CPO optics ran at a median 36°C against 51°C for pluggables, which Lambda ties to "up to 15x higher reliability", a figure NVIDIA's Gilad Shainer repeated in the release.
The release describes the switches as NVIDIA Photonics models with 144 ports of 800Gb/s, liquid cooling and 18 removable external light sources. It does not give a switch count or a measured power saving for the Quebec cluster, so the power case has to be built from figures NVIDIA and Lambda published earlier.
How co-packaged optics cut network power
A pluggable transceiver turns electrical signals into light at the switch's front panel, so the signal first crosses the circuit board. NVIDIA says that path loses up to 22 dB and often costs 30 W per interface. Co-packaged optics put the optical engine beside the switch chip, which NVIDIA says cuts the loss to about 4 dB and the draw to as low as 9 W. These are vendor figures, and NVIDIA's headline for the product line is 3.5x better power efficiency than traditional methods. NVIDIA does not say what the 3.5x covers; its 30 W and 9 W per-interface figures are a 3.3x ratio on optics alone, by our arithmetic.
At switch level the gain is smaller. In a June 1, 2026 post on an early NVIDIA Quantum-X Photonics Q3450-LD switch, Lambda put a standard switch at 7.0 kW and the CPO switch at 3.95 kW, a saving of 3.05 kW per switch, as Wccftech also reported. By Voltgrid's arithmetic that is a 44 percent cut for the whole switch, or about 21 W per port across 144 ports, in line with the roughly 21 W gap between NVIDIA's 30 W and 9 W per-interface figures, though neither company ties those figures to the same port.
Lambda puts the CPO switch at 3.95 kW against 7.0 kW for a standard switch, a vendor figure from an early sample rather than a Quebec measurement.
Graphic: Voltgrid News
Lambda's sizing table for GB300 NVL72 clusters shows what that could mean at this scale. For a 10,368-GPU cluster, the row nearest the Quebec figure, it lists 216 CPO switches and 658 kW of network power freed, equal to 470 extra GPUs at the 1,400 W Blackwell Ultra TDP (thermal design power) Lambda used. Against the 14.5 MW those GPUs draw by TDP alone (our math), the saving is about 4.5 percent. Lambda cautions that CPO does not automatically add deployable GPUs; rack layout and cooling decide how many freed watts become compute. The table is a sizing illustration, not a Quebec measurement, and the release does not say the Quebec cluster uses GB300 NVL72 racks.
The campus and the switch
The Quebec campus has 142 MW of secured, nearly 100 percent renewable capacity, free cooling for up to 80 percent of the year and a guaranteed PUE of 1.2, Lambda said, with no process water used for server cooling. PUE is total facility power divided by IT power.
The 144-port, 800Gb/s specification matches NVIDIA's Quantum-X Photonics InfiniBand switch, the model Lambda's March 2026 release calls Quantum-X800 InfiniBand Photonics and its June 2026 post calls the Q3450-LD. NVIDIA's Spectrum-X Photonics switches come with 128 or 512 ports, not 144. The Lambda blog lists the Q3450-LD as a 4U unit with 115.2 Tb/s of switching capacity, a 48V DC busbar power input and four liquid-cooling quick disconnects, the same infrastructure model as GB300 NVL72 racks. The November release itself names no model and does not say Quebec uses the Q3450-LD.
This is the optics side of the same power squeeze Voltgrid has tracked on the pluggable path, where Cisco and NVIDIA are drafting USFP optics to carry 1.6 Tbps on four 400G lanes. Pluggables are not going away: NVIDIA said its photonics switches will sit alongside pluggable transceivers from suppliers including Coherent, Eoptolink, Fabrinet and Innolight.
What is not confirmed yet
Lambda's benchmark data, the latency baseline and the reliability claim come from Lambda and NVIDIA, and no independent test has been published. Lambda has not said how many switches the Quebec cluster uses, what power it measured against pluggables, or whether Quebec is the site of the deployment it announced in March 2026. The release says the deployment will be shown in NVIDIA's SC26 special address, where more detail may follow. "World's first" is Lambda's own claim.
Sources
- 01Lambda Announces World’s First Large-Scale CPO Deployment of NVIDIA Photonics SwitchesBusiness Wirebusinesswire.com
- 02NVIDIA's First Co-Packaged Optics Switch Lands at Lambda, Cutting 3kW Per Rack and Freeing Power for 3,137 Extra GPUsWccftechwccftech.com
- 03Scaling AI Factories with Co-Packaged Optics for Better Power EfficiencyNVIDIA Technical Blogdeveloper.nvidia.com
- 04NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches to Scale AI Factories to Millions of GPUsNVIDIA Newsroomnvidianews.nvidia.com
- Published
- Sources
- 4 cited
