Nvidia pivots at GTC
Nvidia pushed deeper into AI inference at GTC while rolling in external LPU tech to hedge against custom‑ASIC rivals, and it also launched an open Agent Development Platform and toolkit for enterprise agents — a clear nudge toward agent orchestration and edge inference. The move signals compute is evolving into a platform play where software, orchestration and specialized inference hardware matter as much as raw GPUs. (digitimes.com) (x.com) (investorplace.com)
NVIDIA unveiled the Groq 3 Language Processing Unit (LPU)—the first chip coming from the assets it bought in the roughly $20 billion Groq arrangement—and said Samsung Foundry will manufacture the parts with first shipments slated for Q3 2026. (cnbc.com) NVIDIA described liquid‑cooled Groq 3 LPX racks that aggregate 256 LPUs with roughly 128 GB of on‑chip SRAM and very high scale‑up bandwidth (figures reported in vendor briefings as multi‑petabyte/sec class). (datacenterdynamics.com) The company claims a Vera Rubin + Groq 3 LPX pairing will deliver up to 35× higher inference throughput per megawatt compared with Blackwell‑only racks. (developer.nvidia.com) NVIDIA’s Agent Toolkit bundles OpenShell (an open‑source runtime), Nemotron open models, the AI‑Q agent blueprint and cuOpt skills, and NVIDIA says the AI‑Q hybrid design can cut query costs by more than 50%. (investor.nvidia.com) Vendors publicly named as early adopters include Adobe, Salesforce, SAP, ServiceNow, Siemens, CrowdStrike, Atlassian and others that NVIDIA listed during the keynote. (venturebeat.com) NVIDIA also announced NemoClaw, an enterprise OpenClaw stack that installs NVIDIA OpenShell plus Nemotron models in a single command, provides kernel‑level sandboxing and policy routing, and is available as open‑source on GitHub. (investor.nvidia.com) Two Democratic senators—Elizabeth Warren and Richard Blumenthal—sent a March 19, 2026 letter to Jensen Huang requesting details on the roughly $20 billion Groq deal and asking whether its structure was designed to avoid merger review. (warren.senate.gov) In his GTC keynote Jensen Huang said NVIDIA now has visibility to at least $1 trillion in cumulative AI infrastructure revenue through 2027—an update analysts noted is roughly double the company’s prior $500 billion order‑backlog signal. (bloomberg.com)
