A Comprehensive Overview: What Exactly Did Jensen Huang Say at the Taiwan GTC Conference?
Taylor Wilson
Jensen Huang declared at GTC Taipei that Nvidia is transforming from a GPU company into a full-stack AI infrastructure company, landing four initiatives at once — Vera Rubin mass production, a custom CPU, the DSX factory operating system, and an AI PC platform — all built on one thesis: tokens are revenue units, and compute is factory capacity.
"Tokens are revenue" — what does that business logic actually mean?
Huang's core argument: generative AI has entered the "can do real work" phase. Agentic AI — systems that observe, reason, plan, and call tools on their own — can execute tasks independently.
He cited GitHub data: AI-related code commits grew from roughly 300 million in 2023 to 500 million in 2025, with the first months of 2026 approaching a threefold increase.
This means → AI amplifies engineer output rather than replacing jobs; more code means more tokens consumed, and tokens become a priceable revenue unit.
In plain terms = compute is no longer a cost you stockpile. It works like a factory line that directly produces revenue — whoever has more compute can "ship" more AI services.
Vera Rubin in mass production — what does Nvidia's biggest-ever engineering project look like?
Vera Rubin has entered full mass production. Huang called it Nvidia's most ambitious engineering project in history, with roughly 40,000 engineers and the Taiwanese supply chain involved.
It is not a single GPU but a rack-scale — even multi-rack — system. Core components include the Rubin GPU, Vera CPU, ConnectX-9 networking, BlueField-4 security processor, Spectrum-X Ethernet switches, and a network architecture supporting 200 Gb CPO — co-packaged optics that place optical modules right next to the chip to shorten signal paths.
A manufacturing highlight: a cable-free midplane design cuts rack assembly from roughly 2 hours to about 5 minutes.
This reflects a shift from "building PCs at scale" to industrial assembly — a 24× speed-up that borrows its logic from automobile production lines.
A custom CPU enters the arena — whose lunch is Nvidia eating?
The Vera CPU carries 88 custom Olympus cores, delivers 1.2 TB/s memory bandwidth, and supports NVLink chip-to-chip interconnect, LPDDR5X memory, and PCIe Gen6.
Huang's performance claims: SQL database queries 3× faster, New York Stock Exchange real-time stream processing 6× faster, agentic sandbox performance 1.8× higher than x86.
This means → Nvidia is stepping directly into the data-center CPU profit pool, putting competitive pressure on Intel and AMD.
In plain terms = traditional CPUs were designed for human operators, but AI agents run in a nanosecond world that demands low latency and high bandwidth. Nvidia's logic: if a CPU must sit next to every GPU, build one optimized for AI yourself.
$500 billion for one AI factory — what problem does DSX solve?
Huang disclosed that a 1 GW-class AI factory now costs $50–60 billion, up from $20–30 billion, and could reach $80–100 billion.
DSX is the factory's digital twin — a 1:1 virtual replica — and operating system. Before ground is broken, it simulates power, cooling, networking, and rack layout inside Omniverse.
DSX MaxLPS recovers "stranded watts" — power allocated but not actually used — to boost GPU deployment density on the same power budget. DSX Flex lets the factory dynamically adjust consumption based on grid signals, turning it into a grid-balancing asset.
This means → when a single facility costs hundreds of billions, saving 1% of power saves hundreds of millions of dollars. DSX turns an AI factory from "plug in and run" into a finely managed capacity system.
Enterprise toolkit and open-source model — what is Nvidia doing on the software side?
Nvidia released an enterprise AI toolkit: the Open Shell security framework handles sandboxing, permissions, and privacy, alongside the Hermes framework and CUDA-X tools. Red Hat, Canonical, and Microsoft are among ecosystem partners.
A flagship use case: a chip-design super-agent built with Cadence compresses weeks of chip verification and debugging into hours — a speed-up of more than 40×.
Nvidia also launched Nemotron 3 Ultra, an open-source model using an SSM — state-space model, a sequence-processing architecture more compute-efficient than Transformer — combined with MoE, mixture of experts, which activates only a subset of parameters per query. Nvidia claims 5× faster inference, 30% lower total cost, and is releasing model weights, training data, and training scripts.
The AI PC arrives — what does RTX Spark mean?
Nvidia and Microsoft jointly launched the RTX Spark platform, powered by the N1X chip co-developed with MediaTek. It integrates a Blackwell-architecture GPU, a custom Grace CPU, and 128 GB of unified memory on TSMC's 3 nm process, with a claimed 1 PFLOP-class AI compute capability.
The new product line spans laptops, desktops, and the DGX Station — the latter packing up to 768 GB of memory, enough to run trillion-parameter models on a desk.
Huang defined the new PC as "a legacy OS plus a large language model" and predicted that in ten years, home AI supercomputers could be as common as home theaters.
This means → Nvidia is extending from cloud-scale compute all the way down to the desktop, aiming to make "running agents locally" a baseline PC function — a direct challenge to Qualcomm, Apple, and other edge-chip players.
Content is for reference only, not financial advice.