GPU who? Meta to deploy Nvidia CPUs at large scale

Steven

GPU who? Meta to deploy Nvidia CPUs at large scale

Summary

Meta has already deployed Nvidia’s standalone Grace CPUs in CPU-only systems at scale and will expand its use of Nvidia silicon, including upcoming Vera CPUs and millions of GB300/Vera Rubin Superchips. The move spans general-purpose backend workloads and agentic AI tasks that don’t always require GPUs, while also tying into a much larger plan to deploy more Nvidia GPUs and networking as part of Meta’s hefty 2026 capex.

Key Points

  • Meta is among the first hyperscalers to deploy Nvidia Grace CPUs in CPU-only configurations at scale.
  • Nvidia says Grace delivers about 2x performance per watt on certain backend workloads compared with alternatives.
  • Meta will field millions of GB300 and Vera Rubin Superchips and begin using the Vera CPU family next year.
  • Grace specs: 72 Arm Neoverse V2 cores, up to 3.35GHz, up to 480GB LPDDR5x single-socket (960GB in Superchip), with very high memory bandwidth.
  • Vera CPUs raise core counts to 88 custom Arm cores, add SMT and confidential computing — the latter earmarked for private WhatsApp processing.
  • The move bucks a wider industry pivot toward bespoke Arm CPUs from cloud providers (eg Amazon Graviton, Google Axion).
  • Meta remains multi-vendor: it runs AMD Instinct GPUs and helped design AMD’s Helios rack; further AMD deployments are expected but not yet committed.
  • Deal scale is large — publicly hinted to be worth tens of billions to Nvidia as Meta pursues major infrastructure expansion.

Content summary

Until now Nvidia’s Grace CPUs mostly shipped as part of Superchips that pair Grace with Hopper/Blackwell GPUs. Meta, which has used Grace-Hopper Superchips in systems like its Andromeda recommender, is now running standalone Grace CPU systems for backend and some AI workloads that don’t need GPUs. Nvidia’s executives highlight energy-efficiency gains for these workloads. The forthcoming Vera CPU family adds higher core counts, SMT and confidential computing, which Meta plans to use for private processing in WhatsApp. Alongside CPUs, Meta will continue large-scale GPU and network deployments as part of its $115–$135bn 2026 capex plans.

Context and Relevance

This is important because a major hyperscaler publicly backing Nvidia’s datacenter CPUs changes competitive dynamics in server silicon. It validates Nvidia’s push beyond GPUs into Arm-based CPUs and into confidential-compute features, while signalling to Intel/AMD and cloud CPU incumbents that Nvidia designs can be production-grade at hyperscale. For infrastructure, AI ops and procurement teams, Meta’s adoption suggests new performance-per-watt and memory-bandwidth trade-offs to consider when architecting backend and inference systems.

Why should I read this

Short version: Meta backing Nvidia CPUs at scale is a big deal — it’s not just hype. If you care about where datacentre compute is headed, this tells you who’s winning which parts of the stack, why memory and power efficiency are driving choices, and how vendors are partnering to own more of the AI pipeline. We’ve skimmed the spec sheets and the marketing so you don’t have to — here’s the gist in one go.

Source

Source: https://www.theregister.com/2026/02/17/meta_nvidia_cpu/