Huang: 'I think we've achieved AGI'
NVIDIA CEO Jensen Huang told press at GTC 2026, “I think we’ve achieved AGI,” arguing AI could soon run companies — a striking public claim from the keynote (ndtv.com). The remark came alongside GTC reveals like the Vera Rubin architecture and LPU plans that NVIDIA says push agentic AI forward (tomshardware.com).
NVIDIA’s Vera Rubin platform was unveiled on March 16, 2026 and the company said seven new chips are already in full production, describing a rack- and POD‑scale architecture that bundles Vera CPUs, Rubin GPUs, an NVLink‑6 switch, ConnectX‑9 SuperNICs, BlueField‑4 DPUs, Spectrum‑6 switches and an integrated Groq 3 LPU. (investor.nvidia.com) NVIDIA’s flagship NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs and, by the company’s published figures, delivers about 3.6 EFLOPS of NVFP4 inference and roughly 2.5 EFLOPS of training per rack along with tens of terabytes of HBM4 capacity. (videocardz.com) The Groq 3 Language Processing Unit in the Vera Rubin stack follows a roughly $20 billion asset/licensing agreement NVIDIA struck for Groq’s technology in late December 2025, and NVIDIA showcased Groq 3 as a production LPU co‑designed into the platform. (cnbc.com) That $20 billion Groq arrangement has drawn scrutiny from U.S. lawmakers: Senators Elizabeth Warren and Richard Blumenthal have asked questions about the transaction and whether it raises antitrust concerns. (msn.com) At GTC NVIDIA told investors it expects roughly $1 trillion in orders tied to Blackwell and Vera Rubin through 2027, and the company says more than 80 MGX ecosystem partners will support Vera Rubin deployments. (cnbc.com) OEMs and systems vendors including HPE and Pegatron announced Vera Rubin‑based systems at GTC 2026, with NVIDIA and partners indicating initial NVL72 samples and OEM systems will ship to tier‑one cloud and enterprise customers in late 2026 and broader production availability into early 2027. (businesswire.com)
