At Computex in Taipei on June 1, NVIDIA unveiled a chip that puts the company in a business it has never been in: making the central processor of a personal computer. The chip, called RTX Spark, is an Arm-based system-on-a-chip that combines a Grace CPU with a Blackwell GPU on a single die. NVIDIA is pitching it at what chief executive Jensen Huang has identified as a $200 billion addressable market for AI-optimized CPUs, and it is coming to market with the backing of Microsoft, Dell, HP and five other major computer makers.
The technical specs are aimed squarely at the claim that AI agents will need serious local compute. RTX Spark delivers roughly one petaflop of FP4-precision AI performance, about 1,000 trillion operations per second, and supports up to 128 gigabytes of unified LPDDR5X memory shared between the CPU and GPU. That combination lets a laptop run large language models of up to 120 billion parameters locally, with a context window of a million tokens, without sending data to a cloud. The chip was co-developed with MediaTek, which contributed the CPU architecture, power management and memory controller.
The design answers a specific complaint about the current generation of AI laptops. Today’s machines pair an x86 processor with a discrete GPU over a PCIe connection, with two separate memory pools, so every AI workload shuttles data across the bus. RTX Spark unifies CPU, GPU and memory in one package, the same architectural logic that made Apple’s M-series chips the benchmark for on-device AI. NVIDIA says the result fits in a 14-millimeter chassis weighing about three pounds.
The lineup behind it is unusually broad for a first-generation platform. Microsoft is building the Surface Laptop Ultra on the chip. Dell, HP, ASUS, Lenovo and MSI confirmed RTX Spark-based devices for the fall, with Acer and Gigabyte following. For NVIDIA, that is the difference between a reference design and a market entry: eight manufacturers shipping on one architecture gives software developers a reason to optimize for it.
The market math is what drew NVIDIA in. Huang told analysts on the company’s fiscal first-quarter earnings call in May that agentic AI workloads, which require the CPU to plan, reason and coordinate as models execute tasks, had opened a $200 billion opportunity in processors. The company’s server-side CPU, Vera, is already on track to generate about $20 billion in revenue this fiscal year, according to analyst projections cited by management, and NVIDIA sees the PC as the next front.
The competitive stakes are high. Intel and AMD together control more than 95 percent of PC processor volume, a duopoly that has held for three decades. Previous Arm-based challengers have struggled: Qualcomm’s Snapdragon X captured less than 1 percent of the market a year after its 2024 launch, according to Mercury Research data cited in press coverage. NVIDIA’s pitch is different, built around AI performance rather than battery life, but the compatibility questions that slowed Qualcomm apply here too, and software support for a new architecture is built one developer at a time.
The strategic logic for NVIDIA is straightforward. Its data-center business is enormous but cyclical, and the company has spent two years preparing for the moment when the AI buildout matures. The PC represents a second front: a market where every unit sold carries an NVIDIA chip, where AI agents create demand for local inference, and where the company can extend its franchise beyond the server rack. Huang’s framing on the earnings call, that agents change where computing happens, is the intellectual foundation for the whole effort.
The shift from assistant to agent is central to the pitch. Microsoft, Apple and Google have all shipped AI PCs built around assistants that answer questions. An agent is a different thing: it runs for hours, calls tools, operates other software and works on the user’s behalf, which means it needs continuous local compute and low-latency access to context. NVIDIA argues that architecture favors unified-memory systems like RTX Spark, where the model never leaves the machine.
NVIDIA’s own ecosystem adds pressure on rivals. The company’s software stack, from CUDA to its robotics and agent frameworks, is already the standard in data centers, and extending it to the PC gives developers a familiar path. The company also has something Intel and AMD cannot easily copy: a GPU business that lets it co-design the AI accelerator and the CPU as one product, rather than two companies trying to coordinate.
The first RTX Spark machines arrive in the fall, and the reviews will be merciless. Early verdicts will turn on application compatibility, battery life and whether local agents actually work better than cloud ones. But the direction of travel is already clear. The PC is becoming an AI device, and NVIDIA, for the first time, wants to own the brain inside it.
