Historically, the PC chip market has been Intel versus AMD, with Qualcomm making inroads on Windows on Arm devices since 2020. At Computex 2026 in Taipei on June 1, Nvidia CEO Jensen Huang added a third option.
Nvidia announced the RTX Spark, an Arm-based superchip — a single piece of silicon that combines a CPU (the general-purpose processor that runs your operating system and apps) and a GPU (the graphics processor, now also used heavily for AI computation) — aimed at laptops and compact desktops. The top-tier variant packs 6,144 CUDA cores, the same count as Nvidia's desktop RTX 5070, alongside up to 128 GB of LPDDR5X unified memory and a peak memory bandwidth of 600 GB/s. For context, that bandwidth matters because shuttling data between memory and processor is often the bottleneck when running large language models locally. The chip delivers up to 1 petaflop of AI compute, meaning it can theoretically execute one quadrillion AI math operations per second.
The RTX Spark uses Nvidia's Blackwell GPU architecture — the same generation powering its high-end data center cards — squeezed into a 45W to 80W power envelope, built on TSMC's 3nm manufacturing process. An 18-core variant with 5,120 CUDA cores also exists for thinner designs.
Who's Building RTX Spark PCs
More than 30 laptop models and 10 desktop systems are in development across Microsoft, Dell, HP, ASUS, Lenovo, and MSI. Microsoft's Surface Laptop Ultra is among the first confirmed devices, pairing the chip with 128 GB of RAM and a mini-LED display. First units are expected in fall 2026. Pricing hasn't been announced, though early indications suggest these will carry premium price tags.
Local AI as the Core Pitch
Huang's argument is that RTX Spark turns a laptop into a machine capable of running AI agents — software that takes multi-step instructions and carries them out autonomously, like researching a topic, reading files, and summarizing results — without sending data to a cloud server. "When it has an autonomous agent...you could talk to it...ask it to read files, go help you do some research," Huang said on stage.
That's a direct jab at the current state of on-device AI. Right now, most useful AI runs in the cloud (ChatGPT, Claude) or through weak on-device inference via the dedicated NPU chips in Qualcomm Snapdragon X and Apple M-series processors. Those NPUs are designed for narrow, low-power tasks. Nvidia's position is that running a genuinely capable model locally requires a full GPU, not an NPU add-on.
The comparison to Apple Silicon is intentional. The RTX Spark's 128 GB memory ceiling and 600 GB/s bandwidth are competitive with Apple's M3 Ultra, and Nvidia has outlined a three-generation roadmap: the current Spark, followed by Rubin (moving to LPDDR6 memory), and then Rosa Feynman. Qualcomm's Snapdragon X Elite maxes out at 64 GB and around 273 GB/s of bandwidth.
The catch: none of this is shipping yet, and Nvidia has said it will release independent performance benchmarks closer to launch. The spec sheet looks strong on paper. Whether it translates to real-world AI and productivity performance will depend on software optimization, driver maturity, and how well Windows on Arm handles the application compatibility issues that have dogged that platform for years. Fall 2026 can't come soon enough for a real answer.