Jensen Huang took the stage at Computex in Taipei — the world's largest PC trade show — and did what he does best: made everyone else feel like they're behind. Nvidia's big announcement this time wasn't another monstrous data center GPU. It was a PC chip. Specifically, the RTX Spark, a CPU Nvidia is calling a "superchip," rated at 1 petaflop of AI compute. And it's coming to a laptop near you this fall, courtesy of ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with Acer and Gigabyte queuing up behind them.

Let that sink in. Nvidia — the company that prints money selling GPUs to hyperscalers — is now gunning for the CPU market. On your desk. In your bag. Running your AI agents while you sip coffee.

What the RTX Spark Actually Is

Strip away the marketing language and here's what you're looking at: an ARM-based chip with enough integrated CPU, GPU, RAM, and CUDA stack to run local large language models without phoning home to a data center. Nvidia has baked in secure sandboxing — developed jointly with Microsoft — specifically designed to isolate AI agents like OpenClaw or Hermes Agent so they can operate without becoming a security nightmare. That last part matters more than most headlines will tell you. Running agents locally means your data doesn't leave the device. For enterprise customers especially, that's not a nice-to-have; it's a requirement.

The chip also targets Nvidia's traditional heartland — gamers — promising better image quality and AI-accelerated features across more than 1,000 games and applications. Over 100 software vendors have already pledged support, including Adobe, Blender, ComfyUI, Riot Games, and Xbox. That's not a trivial ecosystem to assemble.

Jensen's Actual Vision Here Is Much Bigger Than a Laptop

Huang isn't just trying to sell you a faster PC. He's selling a paradigm shift — one where you stop clicking through menus and start talking to your machine. "With RTX Spark and Microsoft Windows, you ask — and the PC does the work," he said. Frontier models, creative workflows, gaming. All local. All on a laptop.

This connects directly to what he told investors last month, when he declared he'd identified a brand new $200 billion market for Nvidia in CPUs — not just GPUs. He cited the Vera, Nvidia's high-end server CPU launched earlier this year, claiming $20 billion in sales already. Then he dropped this on the earnings call: "We'll have billions of agents, and those billions of agents will all use tools. And those tools are going to be like PCs, just like us humans using PCs today. We're going to need a lot more CPUs."

That's a coherent thesis. Whether it plays out is a different question.

Yes, Nvidia Has Tried This Before — and Crashed Spectacularly

Let's not pretend there's no history here. In 2013, Microsoft launched the Surface RT — an ARM-based Windows device powered by Nvidia silicon. It bombed so badly that Microsoft wrote off $900 million. Dell bailed. Partners ran for the exits. The whole ARM-on-Windows experiment became a cautionary tale taught in business school case studies about hubris and timing.

So what's different now? Quite a bit, actually. The Surface RT was underpowered and ran a crippled version of Windows that couldn't execute standard x86 applications. The RTX Spark is the opposite problem — it's genuinely powerful. Microsoft is branding its own RTX Spark machine the Surface Laptop Ultra and calling it "the most powerful Surface Laptop ever built." That's a very different positioning than the neutered tablet experiment of 2013.

Plus, the software story has completely changed. The AI application ecosystem didn't exist a decade ago. Today, running a local LLM isn't an academic exercise — it's something developers, researchers, and increasingly non-technical users actually want to do.

The Uncomfortable Questions Nobody's Answering Yet

Here's where I pump the brakes slightly. PC manufacturers have been conspicuously quiet on pricing. Given that Nvidia's own DGX Spark mini-computer — essentially what these Windows PCs appear to be based on — sells to developers for around $4,800, anyone expecting budget-friendly AI PCs should temper their enthusiasm.

The comparison that looms large: Apple's Mac Mini. It's become the go-to local AI inference machine for a huge chunk of the developer community precisely because it delivers serious performance at a price that doesn't require a budget approval process. If Nvidia's ecosystem partners price these RTX Spark machines at the high end of the market, they're not competing for the same buyers.

There's also the latency and battery life question — running a 1-petaflop AI workload on a laptop isn't free. Thermal throttling, power draw, and battery duration under real inference loads are the kinds of specs that matter for daily use but rarely show up in launch-day press releases. We'll need independent benchmarks, not curated demos.

The Bottom Line

Nvidia has assembled an impressive coalition of hardware partners, locked in a serious software ecosystem, and built a chip purpose-designed for the local AI agent use case. The secure sandboxing is smart engineering. The Microsoft co-development adds credibility. Jensen Huang's track record, especially recent quarters, makes it genuinely difficult to dismiss this as another doomed ARM-on-Windows adventure.

But "if it works" is doing a lot of heavy lifting in that sentence. The real test isn't the Computex demo. It's whether everyday users — not just developers — find enough value in local AI agents to pay a significant premium over their current machines. That answer arrives this fall, when the first RTX Spark laptops actually ship and real people start using them outside a controlled environment.

If Nvidia has genuinely cracked local AI agents that are fast, safe, and useful for normal humans? That's not hype — that's a real platform shift. If the experience is clunky, overpriced, or battery-hungry in practice? Well. We've seen this movie before.