Jensen Huang (黄仁勋) Keeps Silicon Valley Up Again — Vera Rubin, a Vera CPU and “$1 trillion” of Demand

The big reveal

NVIDIA (英伟达) CEO Jensen Huang (黄仁勋) used a late-night GTC keynote in San Jose to lay down what he called a generational leap: the Vera Rubin platform — seven new chips, five rack designs and a “giant” AI supercomputer — plus a new Vera CPU and a raft of systems and software that NVIDIA says are tuned for the next wave of agentic (autonomous) AI. Huang claimed the stack can slash token-generation cost by an order of magnitude, train Mixtures-of-Experts models with only a quarter of the GPUs previously needed, and that NVIDIA now “sees at least $1 trillion” of demand into 2027. Those are company assertions; they come from NVIDIA’s stage metrics and road map.

The technology

Vera Rubin is billed as an integrated NVL72 system tying Rubin GPUs, Vera CPUs, NVLink‑6 switches, BlueField‑4 DPUs and ConnectX‑9 SuperNICs into a liquid‑cooled, tightly coupled fabric. NVIDIA described 45°C warm‑water cooling, NVLink‑C2C coherent CPU‑GPU links with 1.8 TB/s, and claims such design choices yield dramatic gains in per‑watt inference throughput. Huang also introduced a Vera CPU — an Arm v9.2‑A Olympus‑core part with 88 cores/144 threads and what NVIDIA calls “spatial multithreading” — and showed an LPU derived from Groq technology to accelerate low‑latency decoding; it has been reported that NVIDIA has acquired the team behind Groq and integrated that LPU into the platform. NVIDIA framed the stack as optimized for “agentic AI” — models that can plan, act and use tools — and said partners ranging from Alibaba (阿里巴巴) and Lenovo (联想) to Meta, Oracle Cloud and major OEMs will deploy the systems.

Business implications and geopolitical context

Beyond chips, Huang sketched new commercial models — “tokens” as a graded commodity for AI factories and an open‑source initiative called NemoClaw that he urged every company to adopt — and promised an annual cadence of new architectures. If NVIDIA’s $1 trillion figure holds even remotely true, it accelerates demand for cutting‑edge packaging, fabs and interconnects. That raises geopolitical frictions: advanced AI silicon, optical co‑packaging and space‑qualified modules (NVIDIA also showed a Space‑1 Vera Rubin orbit module) sit squarely in areas already subject to export controls and scrutiny between the U.S., Taiwan (台积电/TSMC) and China. How quickly hyperscalers and governments deploy these systems — and under what export rules — will matter as much as the raw performance numbers. Who gets to buy the newest AI factories? That question may keep policymakers and CTOs up at night as much as Jensen Huang keeps Silicon Valley awake.