Nvidia launches RTX Spark super chip, says it will power PCs designed for AI agents

Nvidia has launched its new ARM-based chip – the RTX Spark – for Windows PCs during the company's GTC Keynote at Computex. But that's not all. The company also launched new AI infrastructure, open models, and more. Here is everything you need to know.

In Short

• Nvidia launches RTX Spark chip for Windows PCs
• RTX Spark promises maximum agentic AI performance
• Nvidia also released new AI computing platforms, open models

Nvidia has announced a new era for AI PCs. During the company’s GTC Keynote at Computex in Taipei, CEO Jensen Huang announced a new ARM-based chip – the RTX Spark – And focused on agentic AI for Windows PCs.

During the keynote, Jensen Huang claimed that AI was “now a profit generator,” and everything was now going to be based on agents. AI agents basically help you automate different tasks, without having to manually do the entire process yourself. That is, the agent can do your routine tasks on its own.

And now, Jensen Huang stated that Microsoft and Nvidia were reinventing the PC for the first time in 40 years thanks to agentic AI. The company revealed the Nvidia RTX Spark, which forms the foundation of the new agentic platform on Windows, with up to 128GB of LPDDR5X RAM, 6,144 GPU cores with 20 CPU cores. It is similar to the GB10 chip that Nvidia launched in the DGX Spark last year.

Though Nvidia says that there will be less powerful versions of the RTX Spark, going as low as 16GB of RAM.

RTX Spark comes with everything the world has created

Jensen Huang claimed that the new chip “runs everything the world has ever created, plus it runs agents.”

Nvidia says that with RTX Spark, you can run agents locally via its OpenShell runtime on Windows PCs. That is, you can even have your own AI agent doing things for you on a PC with this chip.

For creators, Nvidia says that RTX Spark can render 90GB-plus 3D scenes, edit 12K videos or generate AI videos in 4K. Nvidia also said Adobe was tailoring Photoshop and Premiere for RTX Spark from the ground up, with Firefly-powered Generative Fill in Photoshop and Generative Extend in Premiere among the accelerated tools.

According to Nvidia, the platform can deliver up to two times faster AI, editing, colouring and effects across creative workflows

Nvidia states RTX Spark brings together technologies including CUDA, RTX, DLSS, TensorRT, OptiX, Reflex and G-SYNC in slim Windows laptops with all-day battery life and in small desktop PCs.

When will Nvidia RTX Spark laptops launch?

On hardware, the RTX Spark laptops are confirmed to come in 14-inch to 16-inch sizes, be as slim as 14 millimetres and as light as three pounds (1.3kg). Nvidia says that you can expert the devices to feature aluminium chassis and tandem OLED displays with G-SYNC.

Nvidia states that RTX Spark-powered slim Windows laptops and PCs will be available this fall from Asus, Dell, HP, Lenovo, Microsoft and MSI.

Microsoft has unveiled the Surface Laptop Ultra, which comes powered by the RTX Spark, which will come with up to 128GB of LPDDR5X RAM, and a 15-inch mini-LED display with 2,000 nits of peak brightness. The Laptop Ultra will launch before the end of this year.

The launch marks Nvidia’s entry into a new round of PC chip development after its initial attempt with the Microsoft Surface RT in 2012. It is believed that the launch of RTX Spark comes as Qualcomm’s exclusive license for Windows 11 ARM chips likely ended.

Nvidia announces new Vera CPUs

But PCs were not the only focus for Nvidia during the keynote. The company also doubled down on agentic AI infrastructure. Vera Rubin was one of the highlights of the keynote. Vera Rubin is the first multi-rack podscale supercomputer built for the agentic AI.

According to Huang, the economics of CPUs for agents is different from that of humans. “Every moment that an agent is waiting keeps it from going to the next step,” he said. To tackle this issue, Nvidia designed new Vera CPUs with a new architecture. Jensen Huang added, “This is a CPU built for agents.”

Nvidia listed OpenAI, Anthropic, and SpaceX as early adopters of Vera CPUs.

Vera CPUs are part of Vera Rubin. Jensen Huang termed Vera Rubin as the “most ambitious endeavour in the Nvidia history,” and focuses on every part of the agentic process. Huang claimed that everyone at Nvidia had worked on Vera Rubin, and called it a “miracle.”

That is, Vera Rubin serves as a full stack AI infrastructure that can be used by companies for the entire AI workflow. This is powered by Nvidia’s DSX systems that can allow for more GPUs being deployed in datacentres with less energy and water consumption – something that has been a concern as more datacentres are set up globally.

DSX refers to Nvidia’s ecosystem for AI infrastructure, the way RTX refers to its GPUs.
Huang claimed that AI infrastructure was becoming increasingly important today, going so far as saying, “Compute is revenue now. Compute is profit. The absence of compute is now loss.” He added, “The more you buy, the more you make.”

Huang confirmed that Vera Rubin was now in full production, and twice as large as Grace Blackwell computing platform. Additionally, he stated, “What used to take two hours to assemble one Grace Blackwell stack now only takes 5 minutes.”

In a video clip shown during the keynote, Nvidia confirmed that Microsoft and Dell already had working Vera Rubin stacks.

New Nvidia AI toolkit for developers

The company also unveiled a new agent toolkit for enterprise AI. This toolkit comes with four key things – Models and agents like Claude Code, Codex, Harnass like Openshell that keeps everything private and secure within your enterprise, tools from Nvidia’s CUDA-X library, and the company's OpenShell runtime operating system.

And the company also unveiled its Nemotron 3 Ultra open AI model. As per the company, you get the data that was used to train the model. Nvidia listed companies including Sarvam and Perplexity as its partners in open model development.

The company claims that Nemotron 3 Ultra is five times faster and 30 per cent cheaper than even the most cost efficient AI models. Apart from Nemotron 3 Ultra, Nvidia also announced Cosmos 3 – an open multimodal platform for physical AI. and Isaac Groot platform for humanoid platforms.

The company also unveiled Alpamayo for its Hyperion self-driving platform. Alpamayo is the world’s first reasoning autonomous vehicle model that can talk out loud as it drives.

- Ends