GTC pushed inference focus
GTC 2026 emphasized inference and agentic computing as the next battleground, with NVIDIA positioning deeper into inference to blunt custom‑chip rivals. The messaging frames inference readiness and latency as strategic priorities for production apps. (hyperight.com) (digitimes.com)
NVIDIA positioned Vera Rubin as a six‑chip, tightly co‑designed platform that the company says delivers up to a 10× reduction in inference token cost compared with its prior Blackwell generation. (investor.nvidia.com) The Rubin NVL72 rack is specified around 50 PFLOPS per Rubin GPU and a multi‑rack throughput that NVIDIA lists in the exaFLOPS range (NVIDIA and third‑party breakdowns put rack figures in the low EFLOPS). (hashrateindex.com) GTC introduced the Groq 3 Language Processing Unit (LPU) and liquid‑cooled LPX racks designed for inference: NVIDIA described 256‑LPU LPX racks with roughly 128 GB of aggregate on‑chip SRAM and 640 TB/s of scale‑up bandwidth to prioritize ultra‑low‑latency token generation. (datacenterdynamics.com) NVIDIA’s entry of Groq technology followed a roughly $20 billion December 24, 2025 asset/licensing agreement that moved Groq’s LPU IP and leadership into NVIDIA’s inference roadmap ahead of GTC. (cnbc.com) CEO Jensen Huang used the keynote to frame the commercial shift to inference and agentic workloads, forecasting about $1 trillion in AI chip orders through 2027 as demand moves from model training to production serving. (cnbc.com) NVIDIA announced Dynamo 1.0, an open‑source “inference operating system” for AI factories that the company says is being adopted by major cloud providers and that NVIDIA claims boosts inference throughput and efficiency in production deployments. (nvidianews.nvidia.com) On the agentic‑software front NVIDIA launched the NemoClaw stack for the OpenClaw ecosystem — a single‑command installer that deploys Nemotron models plus the new OpenShell runtime to add sandboxing, privacy, and enterprise guardrails (NVIDIA also published the NemoClaw repo). (investor.nvidia.com)
