Nvidia’s GTC push beyond GPUs
At GTC 2026 Nvidia signaled a push beyond GPUs with three new reference systems—the Groq LPX inference rack, Vera ETL256 CPU rack and an STX storage architecture—aiming to own inference, orchestration and storage for AI workloads (digitimes.com). That matters for cloud gaming and AI tooling because it shifts Nvidia toward offering full‑stack infrastructure, not just accelerators (digitimes.com).
NVIDIA’s Vera ETL256 rack packs up to 256 Vera CPUs in a single liquid‑cooled MGX reference design and, NVIDIA says, can sustain more than 22,500 concurrent CPU environments. (nvidia.com ) (nvidia.com) The Vera rack exposes up to 400 TB of LPDDR5X system memory and links CPUs to Rubin GPUs with NVLink‑C2C at about 1.8 TB/s of coherent bandwidth to speed CPU–GPU handoffs. (theregister.com ) (theregister.com) The Groq 3 LPX inference rack is a 256‑LPU, fully liquid‑cooled design NVIDIA describes as a rack‑scale low‑latency accelerator with on‑chip SRAM bandwidth measured in petabytes per second and a 35× inference‑throughput‑per‑megawatt claim versus Blackwell. (developer.nvidia.com ) (developer.nvidia.com) NVIDIA’s product pages and reporting show the LPX rack is built on the same MGX platform as Vera and will draw up to about 160 kW at peak in production configurations, matching the power envelope of Rubin‑class racks. (nvidia.com ) (nvidia.com) The BlueField‑4 STX storage reference architecture centers a storage‑optimized BlueField‑4 DPU plus ConnectX‑9 SuperNIC and Spectrum‑X networking, and NVIDIA claims STX delivers up to 5× token throughput, up to 4× better energy efficiency and roughly 2× faster data ingestion for long‑context inference. (investor.nvidia.com ) (investor.nvidia.com) NVIDIA says orchestration software dubbed Dynamo will classify requests and disaggregate serving—routing prefill and attention work to Rubin GPUs while sending latency‑sensitive FFN/MoE decode to LPUs—creating a coordinated pipeline across Rubin, Vera and LPX racks. (developer.nvidia.com ) (developer.nvidia.com) Major cloud and infra partners named by NVIDIA and event coverage — including CoreWeave, Lambda, Oracle Cloud Infrastructure, Dell, HPE, Lenovo, Supermicro, NetApp and others — are listed as early adopters or ecosystem builders for the Vera, LPX and STX reference systems. (nvidia.com ) (investor.nvidia.com)
