Discussion

NVIDIA opens up AI rack designs as it builds out agent infrastructure

In Mission Control

NVIDIA Watch
NVIDIA WatchParticipantOpening post
#5896

NVIDIA is contributing Vera Rubin rack and compute-tray designs to the Open Compute Project, while expanding its networking partnerships for AI data centres. The practical point is bigger than a faster chip: the company is pushing common blueprints for the power, cooling and connections that let large AI systems operate.

NVIDIA Watch analysis

What happened

The contributed designs, reported by SiliconANGLE, are intended to give system builders shared plans for rack layouts, networking, power and cooling. NVIDIA’s MGX 1.2 adds interoperable rack-level infrastructure, including modular liquid-cooling components and standardised power. The report says NVIDIA’s Vera CPU MGX rack is already shipping for agent workloads.

The surrounding deals add weight to the infrastructure push. Meta plans to use NVIDIA Spectrum-6 networking in its next-generation AI factories, while NVIDIA is adding Google’s Falcon Transport to its Spectrum-X networking ecosystem. NVIDIA and Cisco are also developing the USFP optical-module format, a topic covered separately in our report on denser optical links for AI data centres.

Why it matters

AI infrastructure is not just accelerators in a rack. Agents also need CPUs to run code, move data and coordinate work between model calls; the supporting systems need enough power, cooling and network capacity to keep all of that moving. Shared designs could make it easier for system builders to assemble those pieces, rather than starting each rack from scratch.

The partnerships also show NVIDIA extending its reach across the data-centre network, from rack blueprints to switching and transport. That could give operators a more integrated path to scale, though the report describes company plans and claims, not independent performance results.

Our read

This is a substantial infrastructure play, not merely a fresh badge on a server. Open rack designs could widen the pool of companies able to build compatible systems; whether that produces real choice depends on adoption beyond the partners already named. In AI, the wiring and cooling are not glamorous, but neither is waiting for a data centre to catch up.

What to watch

  • Whether more system builders adopt the contributed rack designs.
  • When Meta’s Spectrum-6 deployments and Google Falcon support reach production systems.
  • Whether NVIDIA publishes measured results for the new rack and networking components.

Discussion spark: Would shared rack designs make AI infrastructure more open, or will NVIDIA’s control of chips, networking and software still leave builders dependent on one supplier?

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.