Jensen Huang (黄仁勋) crowned “Token King” as Nvidia unveils Rubin platform, Vera CPU and Groq LPU at GTC

Big projection, bigger questions

Nvidia CEO Jensen Huang (黄仁勋) used his GTC keynote to make a bold claim: he projected Nvidia’s AI-chip revenue would reach at least $1 trillion by 2027 — a statement that instantly spawned the “Token King” meme on social media. The scale of the claim matters. IDC’s more conservative industry forecasts put the entire semiconductor market near $890 billion in 2026 under optimistic growth assumptions, while Nvidia’s fiscal 2025 revenue was about $215.9 billion. Can one company capture orders of magnitude more AI demand in two years? Huang framed the answer around tokens, inference, and new architecture.

Key announcements: Rubin, Vera, LPU and NemoClaw

Huang unveiled the Rubin family — a multi-rack “Vera Rubin” platform with HBM4 memory (bandwidth reportedly up to 22 TB/s), claims of 5× inference performance over Blackwell, much lower per-token costs, and a Rubin Ultra high-end iteration planned for 2027. He also said the Vera CPU has entered mass production and will be delivered in the second half of the year; it was reportedly slated to appear in clouds from Alibaba (阿里巴巴), ByteDance (字节跳动), Meta, Jiagu Cloud (甲骨云) and CoreWeave. Another curious reveal was Space‑1 Vera Rubin, described as a module for space data centers but without many details. Huang highlighted industry shifts such as full liquid cooling and the first CPO (co-packaged optics) Spectrum‑6 SPX switch — changes that touch PCB, CCL and other supply-chain nodes.

Betting on inference and the token economy

Huang made an explicit strategic pivot: inference now dominates the narrative. He mentioned inference nearly four times more than training in his talk, and introduced Groq LPUs — “token accelerators” — that Nvidia will integrate into Vera Rubin racks to speed low-latency token generation while GPUs handle high-throughput prefill and attention. Huang said Groq LP30 is in production with Samsung as foundry and expected to ship later this year; future LPX parts will fold into Nvidia’s NVFP4 compute domain. On the software side Nvidia launched NemoClaw, a hardened, enterprise-focused stack for deploying OpenClaw-style agents with built-in privacy and policy engines, plus a Nemotron model family and a “Nemotron Alliance” of startups — evidence Nvidia is trying to own not just silicon but the agent software layer.

Market and geopolitical context

These product bets come against a complex backdrop: rising ASIC competition (Google TPU and others), changing thermal and board designs, and geopolitical constraints such as US export controls that shape where cutting‑edge systems can be sold. It has been reported that Nvidia’s manufacturing and supply relationships — including Samsung as a foundry partner — will be critical to meeting Huang’s aggressive timeline. Investors and enterprise buyers will watch whether Nvidia can deliver the promised efficiency and scale without ceding ground on cost or regional access. The question remains: can the company convert a token-centric roadmap into the trillion-dollar reality Huang sketched, or will specialization and geopolitical friction rebalance the field before then?