Nvidia is looking at $200B CPU market with AI agent PCs from Microsoft, Dell, and HP

Ronni Holmvig Strøm · 2026-06-02

Nvidia opened Computex on Sunday by calling a CPU a superchip. The RTX Spark, a 1-petaflop part built to run AI agents locally, ships this fall inside Windows machines from Dell, HP, Lenovo, ASUS, MSI, and Microsoft's own Surface line, with Acer and Gigabyte to follow. The chip is the headline. The argument underneath it is the news. Last month, after another record quarter, Jensen Huang told investors he had found a fresh $200 billion market in selling CPUs rather than GPUs, and his reasoning was one sentence long. The world is about to fill with agents, and agents need computers too.

Compute Now Scales With Agents, Not People

The reframing is worth quoting in full. "We'll have billions of agents, and those billions of agents will all use tools," Huang said on the May earnings call, as reported by TechCrunch. "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."

For four decades the demand curve for personal computing tracked the number of people who wanted a machine. Huang is betting it now tracks the number of agents, and agents multiply faster than people do. A household with three humans might run thirty agents, each wanting its own sandbox and its own clock cycles. That is the math behind the $200 billion, and it is not speculative. The server CPU Nvidia shipped this year, Vera, has already moved $20 billion of it.

The product follows the math. "With RTX Spark and Microsoft Windows, you ask, and the PC does the work," Huang said at launch. The PC stops being a thing you operate and becomes a thing your agent operates while you describe what you want. That is a different machine than the one that has sat on desks since 1981, even if it boots the same Windows.

The Real News Is Where the Agent Lives

Strip away the benchmarks and the genuinely new claim is about location. For two years the default answer to "where does your agent run" has been someone else's data center. You rent the model by the token, and your context lives on a rack you will never see.

The RTX Spark machines give a better answer. They pair a secure sandbox, co-developed with Microsoft, with enough memory and CUDA to run capable models on the desk in front of you. A sandboxed agent on hardware you own is private by default and still working when the network goes dark. It is the difference between leasing every thought and owning the machine that thinks.

This is the direction the whole field has been pointing toward, and Nvidia is the first to put a consumer SKU behind it. The cloud will stay cheaper to start and easier to update. But the moment an agent runs all day on your behalf, owning it starts to matter more than renting it, the way people who run an agent on a Mac Mini today already understand. Nvidia is betting that instinct scales to the mass market. It is a good bet.

This Time the Inputs Are Inverted

Nvidia and Microsoft tried to put a Nvidia chip inside a Surface once before. In 2013 the ARM-based Surface RT cost Microsoft a $900 million write-off, and partners like Dell walked. The failure then was a weak chip with nothing to run on it.

Everything that broke in 2013 has flipped. The chip is more powerful, not less, and Microsoft is confident enough to call its version the Surface Laptop Ultra, "the most powerful Surface Laptop ever built." More than 100 Windows software makers have signed on, including Adobe, Blender, and Riot Games, and the real draw is not the app catalog at all. It is the agent.