After selling graphics cards for 30 years, NVIDIA is bringing new AI chips to reinvent the "personal computer".

"Fifteen to twenty years ago, we had an idea called the mobile phone. Today, we have a better idea, and its name is the PC."

The one who said this was Jensen Huang. On June 1st in Taipei, standing beside a new chip called RTX Spark, he elevated the PC to the same level as, or even higher than, the mobile phone.

Jensen Huang first reviewed the 40 - year development history of Windows PCs | Image source: NVIDIA

You know, for the past thirty years, NVIDIA has always sold you that graphics card - you insert it into a computer built by someone else. This time, it's the other way around. It wants to sell you the entire computer itself.

What's even more unusual is that Jensen Huang didn't come alone. Acer, Asus, Dell, Gigabyte, HP, Lenovo, Microsoft, and MSI - eight manufacturers that usually compete fiercely on the shelves all stood behind the same chip. A company with such a high market value that it doesn't need to make laptops at all actually made a laptop chip and dragged half of the PC industry into it.

This chip, codenamed N1/N1X during the R & D phase, was leaked for more than half a year. But bigger than a single chip is the ambition Jensen Huang showed - what he wants to reinvent is the "personal computer" itself.

From "Accessory" to the "Brain"

For a long time, the "brain" of a PC was the CPU from Intel or AMD. NVIDIA was the muscle - you added it to the motherboard for gaming or rendering. It was powerful, but it was always the one being inserted.

NVIDIA has planned a complete product lineup for the AI PC era | Image source: NVIDIA GTC live stream

RTX Spark is different. It's a complete SoC with the CPU, GPU, and memory soldered together. A Windows laptop will be built around it.

This time, what NVIDIA wants to do is not just stronger muscle but the brain of the entire computer.

It doesn't have x86 authorization, so this path was originally a dead end. So NVIDIA turned to Arm - a 20 - core Grace CPU jointly developed with MediaTek, combined with its own Blackwell GPU. This is the first time it has put its name in the most core position of a mainstream Windows laptop, instead of just sticking a dedicated graphics card label in the corner of the case.

The chip specifications are jointly defined by NVIDIA and Microsoft | Image source: NVIDIA GTC live stream

The significance of this lies in the fact that the laptop market has an annual shipment of about 150 million units. A company that has become the world's top - valued company with data - center GPUs specifically entered this market. It won't be just a side business for a test. What it wants is the brain itself.

As for why it's now - the appetite of local AI for computing power and memory has grown, and Microsoft has opened the door of the Windows on Arm Copilot+ ecosystem to new players like NVIDIA and MediaTek. The door opened, and NVIDIA walked in.

Putting a Petaflop into a Laptop

The specifications of RTX Spark are very impressive - a 20 - core Grace CPU, a Blackwell GPU with 6144 CUDA cores (with performance roughly equivalent to that of a desktop RTX 5070 laptop - grade), up to 128GB of unified memory, 70 billion transistors, and TSMC's 3nm process. But looking at these numbers alone doesn't mean much.

What really made me sit up straight during the live stream was the simultaneous appearance of two other words on a "laptop".

An AI performance of one petaflop (equivalent to 1000 TFLOPS) and 128GB of unified memory - these two things didn't belong to laptops in the past; they belonged to workstations or even computer rooms.

Currently, half of the PC industry has joined the RTX lineup | Image source: NVIDIA GTC live stream

NVIDIA's official usage scenarios are straightforward: locally editing 12K videos, rendering large - scale 3D scenes, and running large AI agents locally. In other words, those tasks that previously had to be sent to the cloud or required carrying a hot and heavy gaming laptop can now theoretically be done on a portable device.

And the most crucial and hardest - to - copy thing for competitors is CUDA.

People doing AI development understand its significance. In recent years, almost all training and inference frameworks were first developed for CUDA and then adapted to other platforms. For someone who wants to run models locally, there were only a few options in the past: carrying a gaming laptop with a noisy fan, buying a Mac with a large unified memory, or using cloud - based GPUs. Each option had its drawbacks.

There is no doubt about the value of the CUDA developer ecosystem | Image source: NVIDIA GTC live stream

What RTX Spark does is to integrate CUDA and 128GB - level unified memory into a thin and light device. For those who need to run large models locally, there is finally a decent alternative to the Mac on their desks - and in terms of memory and the CUDA ecosystem, it may even be a better fit.

I have to be honest. This product is not for everyone. Someone who only uses a computer to browse the web, write documents, and attend meetings probably won't understand what a petaflop and 128GB mean - just like most people don't need a workstation that can render movies. But for developers, content creators, and those who use local large models as production tools, this is the most suitable piece of silicon in the past few years.

The powerful performance configuration also determines that it is currently more suitable for developers rather than ordinary users | Image source: NVIDIA GTC live stream

NVIDIA hasn't announced the price yet. The manufacturers that will use RTX Spark in their laptops also haven't been announced. But judging from its usual style, it will probably be expensive.

NVIDIA Has Never Just Wanted to Sell a Chip

If you think this press conference was just about "NVIDIA releasing a laptop chip", you're underestimating its ambition.

RTX Spark is just the entry point. At the same keynote, NVIDIA also introduced Claw - a home - use AI agent box that can run your agents 24/7 and connect to your home devices, acting as an always - online personal assistant. There's also the DGX Station specifically designed for Windows: 768GB of memory, 20 petaflops, a developer supercomputer that can run trillion - parameter models and sit on your desk. And there's the continuously evolving Nemotron model.

Products for the home - use Claw ecosystem | Image source: NVIDIA GTC live stream

The "better idea than the mobile phone" that Jensen Huang mentioned at the beginning translates into this set of products. In the future he envisions, the PC will no longer be just a computer. It will be like a dishwasher or a home theater, an indispensable and always - online AI device at home, running various personal agents and assistants for you.

What NVIDIA is envisioning is not just a faster computer but an "always - online home AI".

This product line is long. From the current Blackwell to the upcoming Rubin and then to Feynman - NVIDIA has laid out the roadmaps for desktops, laptops, and workstations all the way to 2030. This is not a one - time deal but a product line that will last for several years.

Not only laptops but also a full - range of products from desktops to workstations | Image source: NVIDIA GTC live stream

Looking back at the list of the eight manufacturers standing behind NVIDIA at the beginning, the meaning becomes clear. When Acer to Microsoft are willing to bet their flagship models, it shows that the entire industry believes that local, always - online, agent - driven personal computing is where the PC is headed. NVIDIA didn't just release a chip. It, together with Microsoft and the entire OEM front, announced a direction and invested in silicon, models, and devices.

Next, It's All About Software

The most exciting part - the hardware - is already on display. The chip is powerful, the vision is grand, and the industry has come together.

What remains is something that has always been slower than making chips.

To make daily software and workflows truly utilize this piece of silicon. A laptop that can run large models needs something worth running locally; a "home AI" needs someone to handle its daily tasks. This is not something NVIDIA can achieve in a single press conference. It takes time, the support of Microsoft, and the participation of countless developers.

There are also some specific things to watch in the short term: RTX Spark will be launched in the fall, the price hasn't been announced, the full - fledged models will be unveiled by different manufacturers one by one, and the progress of Windows application adaptation also needs to be monitored. These will determine whether it will end up being a powerful tool in the hands of a few or the next computer for the majority.

Actually, RTX Spark and Claw are just one aspect of this press conference.

On the same day, NVIDIA also took two steps forward in its agent ambitions:

Firstly, it officially mass - produced Vera - a CPU specifically designed for AI agents. The official claims that it is 1.8 times faster than x86. OpenAI, Anthropic, ByteDance, and the New York Stock Exchange are already on its customer list.

Secondly, it announced the open - sourcing of the 550 - billion - parameter Nemotron 3 Ultra, specifically designed for agents that need to run autonomously for a long time. It is already running on the systems of CrowdStrike and Palantir.

Your RTX Spark on the desk, the Claw in the living room, the Vera and Nemotron in the computer room - Putting these things together, you'll understand that what NVIDIA wants is never just a single chip but an entire stack from the terminal to the cloud, specifically designed for agents.

NVIDIA Vera product specifications | Image source: NVIDIA GTC live stream

But among this entire stack, the one that really crosses a threshold is the laptop.

For the first time in thirty years, NVIDIA is no longer the card to be inserted into the computer but the computer itself. As for whether it can really reinvent the "personal computer", the answer will come this fall.

This article is from the WeChat official account "GeekPark" (ID: geekpark), author: Zhang Yongyi, editor: Jing Yu. Republished by 36Kr with permission.