Test report DSG-5009 · Rev B · tested November 17, 2026
AI Datacenter InfrastructureDevice under test
NVIDIA and Lambda put a 10,000-plus-GPU CPO cluster into service
NVIDIA and Lambda have deployed a co-packaged optics cluster with more than 10,000 GPUs. Quantum-X pairs 18 pluggable light sources with 144 800Gbps ports.
- Read
- 3 min
- Words
- 520
- Node
- 3nm
- Operator
- Elena Vasquez
Spec summary
- NVIDIA and Lambda disclosed Lambda’s production CPO cluster with more than 10,000 GPUs at SC26.
- NVIDIA’s Quantum-X CPO switch exposes 144 MPO/MTP ports at 800Gbps each.
- Quantum-X uses 18 pluggable chassis-front light sources for its package-mounted optics.
- Quantum-X carries NVIDIA’s first packaging generation; Spectrum-X will use an updated second-generation part.
- At OCP Summit 2026, another vendor’s keynote still described CPO as years from deployment.

By Elena Vasquez
More than 10,000 GPUs are running in Lambda’s production co-packaged optics (CPO) cluster, the first production deployment NVIDIA and Lambda disclosed at SC26.
The report was explicit: “it exists and is running today,” putting NVIDIA “at the forefront of deploying CPO.” The industry has spent decades trying to move CPO from the lab to production, with many participants expecting deployment to remain a few years away.
What hardware did NVIDIA deploy?
NVIDIA’s first version of the packaging appears in its Quantum-X switching product. Spectrum-X will use an updated, second-generation part.
NVIDIA presented co-packaged optics at Hot Chips 2025. The disclosed configuration includes:
- More than 10,000 GPUs in Lambda’s production cluster
- 18 pluggable light sources at the chassis front
- 144 MPO/MTP ports on the Quantum-X CPO switch
- 800Gbps per optical port
The 18 sources generate light on the route to optical modules mounted on the switch package. In a conventional front-panel design, those optical modules would sit outside the package and connect back to the switch chip.
Why does serviceability matter?
Keeping the sources pluggable allows technicians to service them. That arrangement addresses one of the key inhibitors to CPO in switches: integrating the light source into the package. Light sources tend to be the parts that fail in pluggable optics, making service access an important deployment advantage.
The disclosure also says the package-side connections are at least somewhat serviceable, which helps deployment. Together, the pluggable sources and those connections address serviceability at both the light-source and switch-package ends.
How does the design compare with pluggable optics?
Quantum-X uses 18 light sources to feed optical modules mounted on the switch package. A traditional switch would instead use either 144 800Gbps optical modules or 72 dual-800G modules.
That reduction leaves fewer light sources that can fail. It also makes the sources easier to cool than a larger array of individual pluggable modules would allow.
How does CPO change the power budget?
Moving optics from front-panel pluggables onto the switch package removes the need for numerous high-speed copper links to carry signalling over relatively long distances from the switch chip to the front cages. CPO can therefore cut power consumption significantly.
The recovered power can support more GPUs within the same cluster footprint. In a power-limited environment, more power allocated directly to GPUs directly affects the revenue generated per watt available to the cluster.
Neither the 18-source count nor the 144-port configuration quantifies that saving. Operators still need to determine how much of the recovered budget each platform can direct to GPUs.
What does production deployment settle?
At OCP Summit 2026, a few weeks before SC26, another vendor’s keynote still described CPO as years away from deployment. The Lambda installation shows that framing has already fallen behind an operating system.
The question has moved from feasibility to adoption. “We no longer have to ask if or when CPO will finally be deployed,” the report said. “The big question now is what the adoption curve will be.”
via ServeTheHome (Source)
More from Elena Vasquez
Show full bio
Senior reporter covering people, launches and day-to-day practice in hardware layer of artificial intelligence at Die Signal.
675 articles
Same lot · LOT-362D
- DSG-80187nmCisco and Nvidia Team Up on Fatter, Faster Pluggable Optics
- DSG-34013nmNVIDIA and Coherent Partner on Optics for Next-Gen Data Centers
- DSG-246528nmNvidia NVSwitch Challenges InfiniBand as Scale-Up AI Standard
- DSG-599828nmNvidia Aims to Supply Every Chip Inside AI Data Centers
- DSG-273828nmNvidia Expands CUDA-X Toolset for AI-Driven Semiconductor Manufacturing