Vera Rubin + Rubin CPU reveal
NVIDIA unveiled the Vera Rubin platform — pairing a new Vera CPU with Rubin GPUs to push integrated ‘AI factories’ and next‑gen inference workflows — a clear signal GTC is shifting from pure training to full system stacks (finance.yahoo.com) (siliconangle.com). Analysts say this makes it easier for enterprises to build end‑to‑end AI products rather than bolt models onto legacy infra (finance.yahoo.com).
NVIDIA announced the Vera Rubin platform at GTC on March 16, 2026 and said seven new chips are already in full production to scale so‑called “AI factories.” (nvidianews.nvidia.com) The public product stack named in the announcement includes NVL72 Rubin GPU racks, Vera CPU racks, Groq 3 LPX inference accelerator racks, BlueField‑4 STX storage racks and Spectrum‑6 SPX Ethernet racks. (investor.nvidia.com) Published spec breakdowns list the Rubin GPU at roughly 50 PFLOPS (NVFP4), with ~288 GB HBM4 and ~22 TB/s memory bandwidth per accelerator, and an NVL72 rack design quoted at about 3.6 EFLOPS aggregate. (hashrateindex.com) NVIDIA’s Vera CPU is described as an 88‑core “Olympus” design with support for very large LPDDR5X memory pools (reports cite up to 1.5 TB per CPU and ~1.2 TB/s memory bandwidth), and NVIDIA positions Vera as the host CPU for reinforcement‑learning and agentic workloads. (tomshardware.com) OEM partners showed immediate system designs: HPE demoed a GX240 blade built on Vera that packs 16 Vera CPUs per blade with options that scale to 640 Vera CPUs in a chassis, and Compal demonstrated HGX Rubin NVL8 integrated solutions at GTC. (sdxcentral.com) NVIDIA and industry coverage say the chips are in full production and system SKUs will ship or be available to partners in the second half of 2026, while some DGX Rubin NVL8 configurations will still use Intel Xeon 6 host CPUs in the near term. (investor.nvidia.com)
