Guangdong Juyao Intelligent Technology Co., Ltd. · Laser Software & Process R&D
23°07′N 113°15′E · GUANGZHOU · PURE STATIC MONOGRAPH
← Back to news
Industry2026.11.16· 4 min read

Industry: Lambda runs NVIDIA Photonics CPO switches across 10,000+ GPUs — latency down up to 10.1%

Lambda has put NVIDIA Photonics switches with co-packaged optics into production at AI-factory scale — 10,000+ Blackwell Ultra GPUs in Quebec — with latency down up to 10.1% at the same throughput. The numbers are vendor benchmarks, but the scale is real.

Lambda blog hero graphic: optical links across a grid, headlined 'World's first large-scale CPO deployment'
COVERImage: Lambda (official blog).
EXECUTIVE SUMMARY

On 16 November, Lambda announced the world's first large-scale deployment of co-packaged optics (CPO) networking: NVIDIA Photonics switches connecting more than 10,000 NVIDIA Blackwell Ultra GPUs at its Quebec campus. In Lambda's benchmarks against conventional pluggable optics, the CPO switches held throughput identical while cutting network latency by 3.8% at rack scale and 10.1% at pod scale, with the optics running at a median 36 °C versus 51 °C. It is the first time CPO has been shown working at AI-factory scale, not just in a lab or a pilot rack.

01 / The deployment: 10,000+ GPUs on CPO switches

Lambda calls the deployment the world's first large-scale use of co-packaged optics networking, connecting 10,000+ NVIDIA Blackwell Ultra GPUs using NVIDIA Photonics networking switches at its Quebec campus. It follows Lambda's early adoption of NVIDIA CPO networking, first announced in November 2025; this time the company benchmarked the technology against conventional optics and published the comparison. The deployment will be featured in NVIDIA's SC26 special address.

The switches are NVIDIA Photonics models with 144 ports of 800 Gb/s, liquid cooling and 18 removable external light sources. Unlike conventional pluggable optics, which convert electrical signals to light at the switch's front panel, CPO places optical engines alongside the switch ASIC, shortening the electrical path from inches to millimetres. That reduces signal loss, removes the need for retimer chips and increases network resiliency — the reason the technology matters as AI clusters grow past ten thousand GPUs.

02 / What the benchmarks show — and what they don't

In benchmarks across rack, scalable-unit and pod configurations, CPO maintained identical throughput and more stable signal integrity than conventional pluggable optics, while delivering lower latency at every scale — the gap growing from 3.8 percent at rack to 10.1 percent at pod. The CPO optics also operated at a lower median temperature of 36 °C, compared with 51 °C for conventional pluggable optics. Lambda says cooler, more stable optics drive up to 15x higher reliability, helping keep large training jobs running.

The limits of the evidence should be stated plainly: Lambda disclosed no absolute latency values, no workload details, no sample sizes and no error margins, and the release does not say how many switches the Quebec campus uses. NVIDIA's broader claims for its Photonics platform — up to five times better power efficiency, up to 15 times higher reliability and four times fewer lasers than pluggable optics — are product-level figures, not measurements from this cluster. What the announcement does establish is scale: CPO has moved from early-adopter samples to a production AI factory with more than ten thousand GPUs.

03 / The campus behind it: 142 MW, nearly all renewable

The Quebec campus itself is part of the story: it operates with 142 MW of secured, nearly 100% renewable energy capacity, enables free cooling for up to 80% of the year, and is guaranteed to a power usage effectiveness (PUE) of 1.2, with no process water consumption associated with server cooling. Pairing power-efficient networking with an energy-efficient campus is the point — in an AI factory, the network is a business accelerator, in NVIDIA networking chief Gilad Shainer's words, because every interruption stalls the job and every watt counts.

For laser and photonics readers, the significance is the market signal: co-packaged optics, long discussed as the next step for data-centre interconnects, now has a named, operating, ten-thousand-GPU reference deployment. The open questions — measured power savings at cluster level, long-term reliability in the field, and cost per port — will decide how fast the rest of the industry follows.

PRIMARY REFERENCES & LINKS

  • Lambda announces the world's first large-scale CPO deployment of NVIDIA Photonics switches
    Lambda · 16 November 2026

    Official announcement verified in full: published 16 November 2026; the 10,000+ GPU scale, switch specifications (144 x 800 Gb/s, liquid cooling, 18 external light sources), latency figures (3.8% rack, 10.1% pod), temperature figures (36 °C vs 51 °C) and campus energy details all come from this announcement.

    Open source link ↗