At GTC Taipei 2026, NVIDIA rolled out more than a dozen major announcements.
The lineup was broad: Vera, a data center CPU built for AI agents; RTX Spark, a platform for personal AI PCs; DGX Station for Windows, a desktop AI supercomputer for enterprises; new robotics foundation models; autonomous driving platforms; and a broader AI factory stack.
This was not just a product launch.
It felt more like Jensen Huang laying out NVIDIA’s roadmap for the next stage of AI.
The first AI boom put NVIDIA at the center of AI compute. This new roadmap points to a bigger ambition: NVIDIA does not just want to sell GPUs into the AI cycle. It wants to become the infrastructure layer underpinning the next generation of AI applications.
1. Vera: A CPU Built for the Agent Era
AI demand is moving from pure training workloads toward inference and agentic systems. Vera is NVIDIA’s answer to that shift.
For years, “AI compute” basically meant GPUs. But AI agents change the workload.
An agent is not just generating a paragraph of text. It may need to break down a task, call tools, write code, query databases, read files, run workflows, check results, and coordinate with other agents.
GPUs still matter. But these agentic workflows also create much heavier demand for CPUs. Task scheduling, data processing, system orchestration, security isolation, and workflow management all become critical.
That is why NVIDIA’s move deeper into CPUs matters.
Vera is not NVIDIA’s first data center CPU. But it is the first one designed specifically for AI agent workloads. NVIDIA calls it “the CPU for agents,” targeting AI agents, reinforcement learning, data processing, and inference-heavy pipelines.
2. RTX Spark: Bringing AI Back to the Personal Computer
NVIDIA is not only going after the data center. It is also pushing AI onto the personal computer.
RTX Spark is NVIDIA’s new platform for AI PCs. Built for Windows PCs and personal AI agents, it delivers 1 petaflop of AI performance and is designed to run AI agents, large models, creative apps, and developer workloads locally.
The first wave of products is expected from Dell, Lenovo, HP, ASUS, MSI, Microsoft Surface, and others.
Today, most consumer AI still lives in the cloud. You open ChatGPT, Claude, or Gemini on your laptop, but the real computation happens in a remote data center.
Traditional PCs were built around apps. To write, you open Word. To analyze data, you open Excel. To edit video, you open Premiere. To code, you open an IDE.
But if personal AI agents become truly useful, the PC may shift from an app-first to a task-first interface.
Instead of opening software one by one, you tell the computer what you want done. The agent then moves across apps, files, and web pages to complete the task.
That is the future RTX Spark is pointing to.
AI will no longer be just a chatbot in a browser tab. It could run locally, understand your files, apps, and workflows, and help execute much more complex tasks.
3. DGX Station for Windows: Local AI Supercomputing for Enterprises
NVIDIA also introduced DGX Station for Windows.
Think of it as an enterprise-grade AI workstation, or a desktop AI supercomputer for serious local workloads. NVIDIA says it can build, run, and connect always-on AI agents locally, and support frontier models with up to 1 trillion parameters on-device.
RTX Spark is the entry point for personal AI PCs. DGX Station for Windows is aimed at enterprises, developers, researchers, and advanced productivity use cases.
It solves a different problem: not every AI workload belongs in the cloud.
Enterprises care about data security, privacy, latency, cost, and customization. In finance, healthcare, manufacturing, R&D, design, engineering simulation, and software development, many datasets cannot simply be uploaded to an external cloud platform.
If companies can run stronger models and agents locally, they get another option: more private, more controlled, and potentially more efficient.
4. Cosmos 3: Giving AI a Model of the Physical World
One of the more important signals from this GTC was NVIDIA’s growing focus on robotics and Physical AI.
NVIDIA introduced Cosmos 3, an open-world foundation model for Physical AI. It uses a mixture-of-transformers architecture to combine visual reasoning, world generation, and action prediction in one system.
The goal is easy to describe but hard to achieve: help AI understand and simulate the physical world.
Large language models are mainly built for text, code, images, and knowledge. They can answer questions, write articles, analyze problems, and generate code.
Robots and self-driving cars face a different kind of problem.
Objects move. Roads change. Space has constraints. Actions have consequences. Mistakes can create real safety risks.
That makes Physical AI much harder.
It does not just need to see. It needs to understand the environment.
It does not just need to generate an answer. It needs to predict what happens after an action.
It does not just output content on a screen. It has to act in the real world.
Cosmos 3 is NVIDIA’s attempt to build a more general foundation layer for robotics, autonomous driving, industrial vision, digital twins, and other physical-world AI applications.
Together with Isaac, Omniverse, Jetson, and NVIDIA’s broader robotics stack, the company is trying to become the default toolchain for robotics companies.
NVIDIA may not build every robot itself. But it wants robot makers to use its chips, models, simulation platforms, and developer tools to train and deploy them.
This should sound familiar.
CUDA played a similar role in the AI training era. It helped NVIDIA turn hardware into an ecosystem.
If robotics eventually becomes a large-scale industrial wave, the key bottlenecks will not just be robot hardware. They will also include compute platforms, simulation systems, model training, synthetic data generation, and deployment tools.
NVIDIA is trying to control those choke points early.
5. Autonomous Driving: Robotaxis Are Physical AI Too
NVIDIA also expanded the ecosystem around DRIVE Hyperion, positioning it as a global platform for robotaxis.
DRIVE Hyperion is an L4-ready autonomous driving platform built on NVIDIA Halos, the company’s full-stack safety system.
NVIDIA also introduced Alpamayo 2 Super, a 32-billion-parameter open reasoning VLA model — vision-language-action — designed for robotaxi reasoning, planning, and action.
It also launched AlpaGym, a closed-loop reinforcement learning framework for training autonomous driving models in simulation.
A car is basically a robot on wheels.
It needs to perceive the world around it, understand roads and traffic participants, predict what may happen next, make driving decisions, and execute those decisions safely in the real world.
So autonomous driving, robotics, and industrial AI may look like separate markets. But underneath, they share the same core problems: perception, reasoning, simulation, action, feedback, and safety.
NVIDIA wants one Physical AI platform to cover all of them.
All Roads Lead to the AI Factory
Put these announcements together, and they point to one idea: the AI Factory.
NVIDIA is no longer just a GPU company or a chip supplier. It is trying to become the operating system for AI infrastructure.
In the next stage, AI will not be limited to training large models in data centers.
It will move into enterprises, PCs, cars, robots, factories, and eventually everyday life.
The Market May Still Be Looking at NVIDIA Too Narrowly
Right now, the market still tends to value NVIDIA through the lens of GPUs, data centers, and large-model capex.
But GTC Taipei 2026 showed a much broader roadmap.
AI agents, AI PCs, local enterprise AI workstations, robotics, autonomous driving, Physical AI, and AI factories — any one of these areas could become a major industry wave.
If even a few of them scale, they could expand NVIDIA’s addressable market and deepen its moat.
The market already understands NVIDIA as the biggest winner of the current AI compute cycle.
The bigger question is whether NVIDIA can also become the default infrastructure company for the next AI era.
$英伟达(NVDA)$ $特斯拉(TSLA)$ $Meta Platforms, Inc.(META)$ $谷歌A(GOOGL)$ $苹果(AAPL)$ $微软(MSFT)$ $亚马逊(AMZN)$ $阿里巴巴-W(09988)$ $小米集团-W(01810)$ $腾讯控股(00700)$ $MINIMAX-WP(00100)$
Comments