Nvidia goes full‑stack
At GTC 2026 Nvidia pushed beyond GPUs with three new systems — Groq LPX inference racks, Vera ETL256 CPU racks and an STX storage reference architecture — signaling a move to own the full AI stack. CEO Jensen Huang also said 40% of revenue now comes from non‑cloud customers, underscoring a strategic shift into enterprise, edge and government buys. (digitimes.com 247wallst.com
NVIDIA’s Vera CPU is an 88‑core, Arm‑based data‑center processor NVIDIA says delivers twice the energy efficiency and 50% higher rack‑scale performance than traditional CPUs. (investor.nvidia.com) NVIDIA’s own rack blueprint consolidates up to 256 liquid‑cooled Vera CPUs in a single rack and the company says that configuration can sustain more than 22,500 concurrent CPU environments. (investor.nvidia.com) The Groq 3 LPX rack uses 256 Groq LPUs and, according to NVIDIA’s technical blog, exposes 40 PB/s of on‑chip SRAM bandwidth and 640 TB/s of high‑radix chip‑to‑chip bandwidth across the rack to cut per‑token jitter. (developer.nvidia.com) NVIDIA and independent coverage say Groq LPX can deliver up to 35× higher inference throughput per megawatt for trillion‑parameter decode workloads versus prior Blackwell systems, with LPUs operating as per‑token accelerators inside the Vera/Rubin stack. (storagereview.com) The BlueField‑4 STX storage reference architecture inserts a high‑performance “context memory” layer (CMX) between GPUs and traditional storage, and NVIDIA claims STX can yield up to 5× token throughput, 4× energy efficiency and 2× faster data ingestion for long‑context inference. (investor.nvidia.com) NVIDIA says early STX adopters include CoreWeave, Lambda, Mistral AI, Nebius, Oracle Cloud Infrastructure and Vultr, and multiple storage vendors such as Dell, NetApp, HPE and Supermicro are building systems around the STX blueprint. (investor.nvidia.com) The company’s December agreement to license Groq’s inference technology and hire senior Groq staff was valued at roughly $20 billion in media reports, a structure CNBC and others described as an asset/IP/talent deal rather than a traditional buyout. (cnbc.com) That $20 billion move drew scrutiny from lawmakers over whether the deal sidesteps merger review, and industry reporting notes NVIDIA has positioned the Groq assets as a licensed, non‑exclusive technology that will be folded into Rubin‑era systems. (markets.financialcontent.com) NVIDIA executives have framed these components as revenue levers: the company reiterated multi‑year targets that include a roughly $1 trillion AI‑chip-and‑systems opportunity by 2027, and channel leaders say Groq LPX and Vera rack products will be routed through partners “over time” to reach enterprise, edge and sovereign deployments. ( )
