At its GTC keynote in San Jose, NVIDIA’s chief executive Jensen Huang laid out a decisive pivot in the company’s strategy and, implicitly, the architecture of the modern data centre. The message was simple and consequential: the AI market is moving from train‑heavy workloads to an era dominated by inference — real‑time “thinking” and token generation — and the economics of that shift will shape which vendors and cloud providers prosper.
Huang quantified the opportunity in heady terms. He said NVIDIA’s next two platform generations — Blackwell and Vera Rubin, followed by an even denser Rubin Ultra built on a Kyber vertical‑integration rack — could unlock at least $1 trillion of cumulative revenue by the end of 2027. To service that demand the company is unbundling inference into two bespoke stages: a prefill stage for tokenization handled by the Vera Rubin family, and a low‑latency decode stage outsourced to a newly integrated Groq LPU (a data‑flow, SRAM‑heavy processor optimized for deterministic, token‑generation workloads).
The engineering rationale for splitting jobs across different silicon is straightforward. High throughput and huge memory footprints favour NVIDIA’s GPUs and NVLink fabric, while low‑latency, predictable token decoding benefits from Groq’s static scheduling and on‑chip SRAM. Huang said this tandem — combined with system‑level co‑design including NVLink 72/144 fabrics, liquid cooling, and shared software layers — produces order‑of‑magnitude improvements in tokens‑per‑watt and cost per token, the crucial metric for commercial inference.
Beyond chips and racks, Huang pitched a software and economic stack to match. OpenClaw, an open‑source agent framework that he likened to Linux or HTML, is positioned as the new operating system for intelligent agents. NVIDIA is shipping NemoClaw, an enterprise‑grade reference design that wraps OpenClaw with security, governance and “policy engines” so agents can be used safely inside corporate networks. Huang argued that every company needs an “OpenClaw strategy” and predicted SaaS would evolve into “GaaS” — agentic software as a service — with engineers assigned annual token budgets.
The announcement was heavy with partnerships and product timing. Huang said Groq’s LP3 LPU would ship in the second half of this year, Vera Rubin systems with their multi‑chip racks would begin rolling out in volume in the second half of 2026, and a Kyber‑based Rubin Ultra was staged for 2027. He also previewed long‑term successors (codenamed Feynman) and a CPU line (Vera CPU) tailored for single‑threaded, tool‑driven agent tasks. Separately, NVIDIA disclosed plans with aerospace partners to prototype a Vera Rubin Space‑1 module, signalling an ambition to export high‑density inference to orbit.
Huang framed these product moves inside a broader ecosystem push. NVIDIA is integrating cuDF/cuVS for accelerated structured and vector data, expanding open models (Nemotron, Cosmos, GROOT, BioNeMo, Earth‑2) and deepening cloud partnerships with AWS, Microsoft, Google, Oracle and specialized cloud players like CoreWeave. He stressed that the company’s strategy is vertically integrated at the platform level while horizontally open: an end‑to‑end stack of silicon, systems and domain libraries meant to lock in long useful life for installed GPUs and stretch the company’s software flywheel.
For enterprises and cloud operators the implications are immediate. If Huang’s numbers hold, the business case for modernising data centres into “token factories” will spur massive capex for land, power and NVLink‑enabled systems. The pressure to reduce per‑token cost will favour suppliers with scale, software depth and supply‑chain breadth. At the same time Huang publicly tied open models and agent frameworks to sovereign and industry‑specific AI, pitching both open‑source and proprietary levers to enable localized AI without ceding control of sensitive data.
Huang’s address was both technical roadmap and manifesto: GPUs will not be abandoned, but they will be joined by specialized accelerators and new system designs that favor latency‑sensitive inference. For policy makers and procurement officers the upshot is a new calculus of industrial strategy, energy planning and vendor dependency: the world’s token factories will be capital‑intensive and concentrated, with implications for competition, national resilience and regulation.