How Nvidia learned to embrace the light in its quest for scale

Steven

How Nvidia learned to embrace the light in its quest for scale

Summary

Nvidia is shifting from copper-only fabrics to photonic interconnects to scale GPU clusters well beyond single-rack limits. At GTC it outlined multi-rack optical scale-up plans (Vera Rubin NVL576, Rosa Feynman NVL1152) and has committed billions to optics suppliers and partners to secure lasers, co-packaged optics (CPO) and optical I/O for future NVLink fabrics.

Key Points

  • Nvidia intends to use photonic interconnects to expand compute domains from single racks to multi-rack systems holding many hundreds — eventually more than a thousand — GPUs.
  • Copper remains ideal for short, intra-rack links but hits reach, thermal and power limits when scaling further.
  • Pluggable optics were too power-hungry at the scale required; CPO dramatically reduces power and module count, making optical NVLink practical.
  • Feynman-generation systems (shipping mid–late 2028) may offer copper or co-packaged optical NVLink, possibly with CPO in switches and/or GPUs.
  • Nvidia has invested heavily (Coherent, Lumentum, Marvell, etc.) to secure lasers, CPO chiplets and optical supply chains needed for widescale deployment.

Content Summary

Nvidia realised that copper backplanes — used in systems such as the NVL72 “Oberon” — cannot scale indefinitely because electrical signals degrade over short distances and packing more GPUs into a single rack creates thermal and power strain. Early optical options (pluggable modules) would have required vast additional power when multiplied across hundreds of GPUs. Advances in near-packaged and co-packaged optics changed the calculus, cutting power and complexity and enabling optical NVLink fabrics.

The company plans a hybrid approach: copper for the first layer inside racks and photonics (pluggables, NPO or CPO) for spine or multi-rack links. For the Feynman generation Nvidia is exploring multiple packaging/topology choices — from adding CPO to switch trays to integrating optics into GPU packages — to reduce tiers and latency. To make this happen Nvidia has invested billions in optics manufacturers and tied up partnerships (notably with Marvell) to secure components and develop optical I/O technologies.

Context and Relevance

This marks a major infrastructure shift in AI datacentre design. Optical NVLink and CPO affect rack topology, power and cooling, procurement and vendor roadmaps. Hyperscalers, cloud AI providers and hardware suppliers should monitor how CPO matures and how Nvidia’s topology decisions influence latency, cost and deployment models over the next few years.

Why should I read this?

Short and blunt: Nvidia just bet hard on photonics and is buying the supply chain to back it up. If you care about how big AI clusters will be built, where power and latency trade-offs land, or who’s going to control optical supply, this is the move that changes the game — and we’ve saved you the time by pulling the essentials out.

Source

Source: https://go.theregister.com/feed/www.theregister.com/2026/04/05/nvidia_optical_scale_up/