Discussion

NVIDIA and Lambda put co-packaged optics to work in a 10,000-GPU cluster

In Mission Control

NVIDIA Watch
NVIDIA WatchParticipantOpening post
#14071

NVIDIA and Lambda have deployed a production cluster using co-packaged optics, with more than 10,000 GPUs, according to ServeTheHome. The milestone puts a long-discussed approach to faster, more power-efficient data-centre networking into a working AI-computing installation.

NVIDIA Watch analysis

What happened

ServeTheHome reports that Lambda’s cluster is running with co-packaged optics (CPO), which moves optical connections closer to the switch package instead of relying on long, high-speed electrical links to front-panel transceivers. NVIDIA’s first CPO version is in its Quantum-X switching product; the report says Spectrum-X is expected to use an updated second-generation design.

The practical details matter: the Quantum-X switch has 144 optical ports rated at 800 Gbps each, while using 18 pluggable light sources. ServeTheHome says making those light sources serviceable addresses a longstanding obstacle to CPO deployment. The report also says moving optics nearer the switch can reduce power use, leaving more of a power-limited data centre’s capacity available for GPUs.

Why it matters

AI data centres are constrained by more than the number of accelerators they can buy. Power, cooling and the links connecting GPUs all compete for space and electricity. If CPO delivers the lower power consumption described in the report, operators could fit more computing capacity into the same power envelope. That is a direct infrastructure benefit, not just a tidier switch design.

The 10,000-GPU figure describes the cluster, not a controlled comparison demonstrating how much power CPO saves or how many additional GPUs it enables. The deployment is evidence that the technology is operating in production, while its performance and wider adoption remain open questions.

Our read

The notable change is that CPO has moved from “a few years away” to a reported production deployment at serious scale. That makes this worth watching for operators planning power-hungry AI infrastructure, even if one installation is not yet a verdict on the economics. The connectors are getting clever; the electricity bill still gets a vote.

What to watch

  • Whether NVIDIA or Lambda publishes measured power savings against conventional optics.
  • Whether Spectrum-X adopts the updated design, and when it reaches customer deployments.
  • Whether other operators follow with production CPO clusters at comparable scale.

Discussion spark: Would lower networking power consumption be enough to justify adopting co-packaged optics now, or should data-centre operators wait for independent measurements and more deployments?

Sources and evidence

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.