Nvidia Is Turning the PC Into an AI Machine. Nobody Asked.
At Computex, Nvidia unveiled RTX Spark — an Arm+Blackwell superchip that runs 120B-parameter AI models locally while playing games at 1440p. Jensen Huang isn't selling GPUs anymore.
Jensen Huang stood at Computex and announced that your next PC will run a 120-billion-parameter AI model locally while simultaneously playing AAA games at 1440p and 100fps. He called it RTX Spark. The chip is real, the specs are real, and the fact that nobody actually requested this does not appear anywhere in the presentation.
The hardware case for RTX Spark is genuinely formidable: up to 128GB of LPDDR5X unified memory, 300 GB/s bandwidth, 20 Arm CPU cores paired with a Blackwell GPU carrying 6,144 CUDA cores. For years, Apple’s M-series chips owned the “serious local AI work” conversation by default — unified memory architecture where the CPU and GPU share the same memory pool meant Apple Silicon could run models that Windows machines needed a dedicated server to touch. Nvidia just ended that argument. The Arm CPU plus Blackwell GPU superchip in a single package is the Windows Apple Silicon moment that didn’t arrive when Microsoft tried it. It arrived because Nvidia decided the local AI story was too valuable to leave to Cupertino.
Jensen’s pitch is “personal AI agent platform.” Your PC becomes a local brain that handles complex tasks before they ever hit the cloud. The pitch lands differently depending on what kind of person you are.
Microsoft is rearchitecting Windows for RTX Spark. Adobe is rebuilding Photoshop and Premiere Pro around it. Dell, HP, Lenovo, ASUS, MSI, and the Microsoft Surface Ultra are all launching partner devices this fall. That’s a complete platform announcement before the hardware ships. Jensen’s pitch is “personal AI agent platform”: your PC becomes a local brain that handles complex tasks before they ever hit the cloud, with a 1-million-token context window running inference on-device. The pitch lands differently depending on what kind of person you are. If you’re already running local LLMs, already thinking about your machine as an inference box, this hardware is legitimately exciting. If you’re a normal person who uses a laptop to work and watch things, an AI agent watching you play games is not a feature you went looking for. The Surface Ultra will ship at a price that puts it in enterprise IT budgets, not gaming aisles. That gap between the keynote audience and the consumer audience is where most of the Spark skepticism lives.
The roadmap goes three generations deep: RTX Spark on Blackwell, then Rubin on LPDDR6, then Rosa Feynman after that. A vertical integration play extending years out. Nvidia is not shipping a new GPU here. It’s betting that the future of PC computing looks more like the iPhone than like a Newegg cart: closed hardware-software systems optimized for a specific use case, with the platform company controlling the full stack. The DIY market, the upgrade cycle, the discrete GPU culture that Nvidia’s consumer business was built on — all of that is secondary to this play. Jensen is betting against it, or at least betting that a large enough segment of the market will trade upgrade flexibility for performance density.
The 300 GB/s bandwidth is overkill for 99% of consumer use cases, and the “agentic AI” framing is a sophisticated way to say your device will run more inference on more of your data. The Surface Ultra will land at a price point that routes it through corporate procurement, not Best Buy. These specs will trickle down across generations, as they always do, but the “personal AI agent platform” pitch at launch is an enterprise announcement cosplaying as a consumer product. The person who buys a $3,000 Surface Ultra in fall 2026 is not the same person who built a gaming PC last year. And Google is already making the same AI-as-OS argument on Android, so the competition for “where does your AI live” is not a two-player game.
RTX Spark ships this fall across Dell, HP, Lenovo, ASUS, MSI, and the Surface Ultra. Whether it changes the PC market depends less on what the hardware can do and more on whether the software stack actually builds the agent layer in ways that people notice in daily use. The specs are real. The chip architecture is sound. Microsoft and Adobe are committed. The gap Jensen needs to fill is the one between “your PC is technically capable of being a personal AI agent” and “you actually want a personal AI agent on your PC.” Apple’s M-series chips are formidable hardware that most people use for email and spreadsheets. Capability doesn’t create demand. If RTX Spark changes how people think about what a PC is for, it’ll be because someone built a specific application — not a platform pitch — that made them reach for the Spark machine instead of anything else. That application doesn’t exist yet. Nvi Sources: Tom’s Hardware · Nvidia Newsroom · Storage Review
