At the SC26 supercomputing conference on November 16, 2026, AI cloud provider Lambda and NVIDIA announced the first large-scale commercial deployment of co-packaged optics (CPO) networking, connecting over 10,000 NVIDIA Blackwell Ultra GPUs in Quebec, Canada. The production rollout replaces front-panel pluggable transceivers with direct silicon photonics co-packaged alongside switch ASICs to slash operating temperatures, cut pod-scale latency, and boost interconnect reliability for distributed AI research clusters.
What happened at SC26
On November 16, 2026, at the SC26 supercomputing conference in Chicago, Lambda and NVIDIA detailed the production operation of NVIDIA Quantum-X Photonics switches connecting more than 10,000 NVIDIA Blackwell Ultra GPUs (GB300 NVL72 rack architecture) at Lambda’s Quebec data center campus. The deployment marks the world’s first verified commercial deployment of co-packaged optics (CPO) in an AI cloud environment.
Conventional supercomputing clusters rely on pluggable optical transceivers slotted into switch front panels, requiring high-speed electrical signals to traverse inches of copper circuit board traces. Co-packaged optics integrates optical engines directly onto the switch substrate beside the application-specific integrated circuit (ASIC). Shortening electrical paths to millimeters eliminates retimer and digital signal processor (DSP) chips, reducing signal degradation and power consumption.
Production testing across rack, scalable unit, and pod topologies revealed identical wire-speed throughput alongside lower latency: network latency dropped by 3.8% at the rack level and up to 10.1% at pod scale. Simultaneously, CPO registered a median optical operating temperature of 36°C, compared to 51°C for conventional pluggable transceivers. This 15°C thermal reduction underpins an up to 15x improvement in component reliability and extended mean time between interruptions (MTBI).
Key architectural specifications and performance benchmarks
The Quebec deployment utilizes NVIDIA Quantum-X InfiniBand Photonics Q3450-LD switches. Each switch delivers 144 ports of 800 Gb/s bandwidth (115.2 Tb/s non-blocking switching capacity). To resolve historical laser reliability concerns, the architecture employs 18 field-replaceable External Light Sources (ELS) that supply continuous-wave laser light to the co-packaged silicon modulators via polarization-maintaining optical fibers.
The following table details the operational differences reported between traditional pluggable optics and the production CPO implementation:
| Metric | Pluggable Optics (OSFP / QSFP-DD) | NVIDIA Quantum-X Photonics CPO |
|---|---|---|
| Optical Engine Placement | Front-panel cages (inches of PCB copper trace) | Co-packaged on ASIC substrate (<10 mm trace) |
| Laser Architecture | Internal DFB laser in every pluggable module | 18 blind-mate External Light Sources (ELS) |
| Operating Temperature | 51°C median under active training load | 36°C median (15°C thermal reduction) |
| Network Latency | Baseline transceiver serialization/retimer latency | 3.8% lower at rack scale; 10.1% lower at pod scale |
| Power Efficiency | Baseline (retimers burn 20% to 30% of switch power) | Up to 3.5x to 5x higher optical power efficiency |
| Component Reliability | Standard MTBF; high thermal stress on lasers | Up to 15x higher reliability; 4x fewer lasers |
| Switch Capacity | 64 to 72 ports of 800 Gb/s per chassis | 144 ports of 800 Gb/s (115.2 Tb/s aggregate) |
The installation leverages Lambda’s 142 MW Quebec campus, which operates on nearly 100% renewable hydroelectric power with a guaranteed Power Usage Effectiveness (PUE) of 1.2. The facility uses free outside air cooling for up to 80% of the year and consumes zero process water for server cooling, complementing the liquid-cooled GB300 NVL72 racks and CPO switch fabric.
Background and industry context
As artificial intelligence research has expanded to training distributed models across thousands of accelerators, interconnect latency and bandwidth have emerged as the primary scaling bottlenecks. Training a frontier AI model requires nodes to execute collective all-reduce operations synchronously across gradient barriers. Any link flap, packet drop, or latency jitter stalls the entire cluster, necessitating costly checkpoint restarts.
As data rates progressed to 200 Gb/s per lane (enabling 800G and 1.6T interfaces), high-frequency signal loss across PCB traces forced manufacturers to insert retimers that consume up to 30 watts per transceiver module. Industry groups including the Optical Internetworking Forum (OIF) and the Consortium for On-Board Optics (COBO) developed CPO specifications, while vendors such as Broadcom and Cisco demonstrated lab prototypes. However, production adoption was hindered by serviceability risks: a failed integrated laser traditionally required discarding the entire switch. Decoupling lasers into external, hot-swappable modules overcomes this operational barrier for cloud deployments.
Industry and leadership reaction
‘Networking has to keep pace with the demands of increasingly large AI clusters,’ said Dave Ward, President of Cloud Services at Lambda. ‘Bringing co-packaged optics into a deployment of this scale is an important step toward building the efficient and resilient infrastructure needed to support the next generation of AI.’
NVIDIA engineering leadership noted during the SC26 announcement that co-packaged silicon photonics flattens the energy curve for large AI factories by shifting power budgets from interconnect overhead back into tensor compute. Telecommunications and networking analysts covering the event at SDxCentral emphasized that the 15x reliability improvement is critical for enterprise training, where optical transceiver degradation represents a frequent cause of cluster downtime. Sustainability observers further noted that pairing 5x more efficient optics with a 1.2 PUE hydro-powered facility establishes a lower compute carbon footprint for energy-intensive scientific research.
What it means for researchers, universities and research offices
The commercial maturation of co-packaged optics presents several practical impacts for academic institutions, supercomputing centers, and research administration offices:
- Training Resiliency and Job Completion: Long-running foundation model runs in computational biology, materials science, and physics often encounter job interruptions caused by failing transceivers. A 15x reliability gain and cooler 36°C optical baselines minimize job failures and checkpoint rollbacks, directly supporting rigorous computational reproducibility.
- Cloud Allocation Efficiency: Pod-scale latency reductions of 10.1% accelerate collective communications, allowing university research teams to extract more training throughput per grant-funded cloud GPU hour.
- Facility Roadmaps and Infrastructure Planning: Campus IT leaders and research software engineers managing an institutional computational environment must evaluate CPO and liquid-cooled rack architectures (such as GB300 NVL72) when planning on-premises supercomputer refreshes versus commercial cloud allocations.
- Grant Compliance and Carbon Reporting: As funding bodies require detailed energy and environmental disclosures for compute-intensive grants, accessing CPO-equipped infrastructure operating at a certified 1.2 PUE simplifies carbon accounting and sustainability compliance.
What happens next
Lambda stated that cloud instances backed by the CPO-connected Blackwell Ultra cluster will be phased into production for enterprise and academic research workloads beginning in late 2026, with broader availability throughout the first quarter of 2027. Early access will focus on large-scale pre-training and multi-agent reasoning workloads.
On the technology roadmap, NVIDIA plans to extend co-packaged optics across its Quantum-X InfiniBand and Spectrum-X Ethernet lines. Concurrently, standards bodies including the IEEE and OIF are slated to ratify standardized 224G and 448G CPO interfaces in 2027, establishing multi-vendor interoperability for next-generation institutional supercomputing centers.
Sources
- Business Wire. (2026-11-16). Lambda Announces World’s First Large-Scale CPO Deployment of NVIDIA Photonics Switches. https://www.businesswire.com/news/home/20261116930878/en/Lambda-Announces-Worlds-First-Large-Scale-CPO-Deployment-of-NVIDIA-Photonics-Switches
- NVIDIA Corporation. (2026-11-16). Lambda Deploys NVIDIA Photonics Switches at Scale. https://blogs.nvidia.com/blog/lambda-photonics/
- SDxCentral. (2026-11-16). Lambda Touts 15x Higher Reliability Using NVIDIA’s CPO Switches for 10,000 Blackwell Ultra GPUs. https://www.sdxcentral.com/news/lambda-touts-15x-higher-reliability-using-nvidias-cpo-switches-for-10000-blackwell-ultra-gpus/
- Lambda Engineering. (2026-11-16). Unbox One of NVIDIA’s First Co-Packaged Optics Samples With Lambda. https://lambda.com/blog/unbox-one-of-nvidias-first-co-packaged-optics-samples-with-lambda
- NVIDIA Networking. (2026-11-16). Silicon Photonics and Quantum-X InfiniBand Architecture. https://www.nvidia.com/en-us/networking/products/silicon-photonics/







