On June 1, 2026, NVIDIA CEO Jensen Huang took the Computex 2026 keynote stage in Taipei and formally launched the RTX Spark superchip — a brand-new processor built specifically for Windows PCs. The chip fuses a Blackwell-architecture GPU with a Grace CPU, delivering up to 1 Petaflop of AI compute, and NVIDIA’s claim is maximal: this will “redefine the PC.” With Microsoft’s deep cooperation, the RTX Spark PC is positioned not as a tool but as the user’s “AI teammate.”
RTX Spark: NVIDIA’s First Consumer SoC
RTX Spark is NVIDIA’s first system-on-chip designed expressly for Windows laptops and mini PCs. One die integrates a Blackwell-architecture RTX GPU with 6,144 CUDA cores, fifth-generation Tensor Cores with FP4 precision support, and a 20-core Grace CPU, communicating over NVLink-C2C high-bandwidth interconnect.
- AI compute: up to 1 Petaflop (FP4 precision)
- Unified memory: up to 128GB at 800 GB/s bandwidth
- GPU cores: 6,144 CUDA cores, fifth-generation Tensor Core
- CPU cores: 20-core Grace CPU (custom-designed with MediaTek)
- Power efficiency: the most energy-efficient RTX chip yet, supporting all-day battery life
MediaTek’s participation in the custom CPU design contributed industry-leading efficiency, performance, and connectivity. The result in hardware form: RTX Spark laptops at 14mm thickness and roughly 1.36kg, with performance comparable to a desktop RTX 5070.
From Tool to Teammate: The Personal AI Agent Era
“For forty years, you opened applications, clicked, and typed. With RTX Spark and Windows, you just say what you need — and the PC does the work.” That is how Huang defined the new generation of AI PC on stage. The chip’s most consequential change is its personal AI agent capability: with 1 Petaflop of local AI compute and 128GB of unified memory, users can run a 120-billion-parameter large language model entirely on the device, with context windows up to one million tokens, no cloud required.
NVIDIA and Microsoft worked deeply together here, bringing new security primitives and the NVIDIA OpenShell runtime to Windows so that AI agents can run locally, safely, and privately. These agents execute tasks inside Windows applications, reason across multi-app workflows, generate images and video, write code plugins, and semantically search local files. Microsoft CEO Satya Nadella framed the ambition: “Our goal is to provide unlimited intelligence for every home and every desk. RTX Spark is the real breakthrough that makes this vision possible.”
Creation and Gaming: Full-Speed Across the Board
RTX Spark aims beyond AI workloads, accelerating creative and gaming scenarios alike:
- 3D rendering: extra-large scenes of 90GB and beyond (OptiX + DLSS)
- Video editing: native 12K 4:2:2 hardware encode/decode
- AI video generation: local 4K AI video
- Gaming: AAA titles at 1440p above 100 FPS, with ray tracing, DLSS, and Reflex
Adobe is re-architecting Photoshop and Premiere for RTX Spark with AI and graphics performance up 2x. NVIDIA also announced DLSS 4.5 Ray Reconstruction, built on a second-generation Transformer model, arriving in August and already integrated into Blender 5.3. More than 1,000 games and applications currently support RTX technologies, and over 100 Windows software vendors — Blackmagic Design, Blender, CapCut, ComfyUI, OTOY among them — are adopting the RTX Spark platform.
Why the SoC Move Matters More Than the Specs
The specification list is impressive, but the strategic meaning of this launch sits one level above the hardware. For NVIDIA, the PC has always been a market it supplied rather than designed: GPUs slotted into machines architected around Intel or AMD platforms. RTX Spark is the first time NVIDIA controls the whole silicon envelope of a consumer computer — CPU, GPU, AI accelerators, interconnect, and memory under one design authority. That changes what NVIDIA can promise OEM partners and what it can optimize: instead of tuning drivers for somebody else’s platform, the company now tunes an entire machine around AI workloads. The unified memory architecture is the enabling decision here — when a 120-billion-parameter model can address the full 128GB pool without copies between CPU memory and VRAM, the practical ceiling of what “runs locally” rises by a generation. It also explains the software push: security primitives, the OpenShell runtime, agent frameworks. Hardware that hosts autonomous agents on untrusted home networks needs a trust story, and NVIDIA is clearly building the platform layer to tell it.
For Windows itself, the launch marks the moment the “AI PC” label stops being a marketing tier on existing laptops. Previous AI PC branding rode on NPUs bolted beside traditional CPUs — enough for camera effects and transcription, not enough for real models. RTX Spark’s 1 Petaflop FP4 figure and million-token context window put actual workload-class AI on the battery-powered side of the desk, and the shift is visible in how Microsoft is talking: agents executing tasks inside applications, not assistants answering questions beside them. Whether developers build for this platform will decide whether “AI teammate” becomes reality or demo-ware, which is why the breadth of the launch-ecosystem list — creative suites, 3D tools, game engines, more than 100 committed Windows vendors — matters as much as any TFLOPS number.
Ecosystem Landing: On Shelves This Fall
Thin Windows laptops and compact desktops powered by RTX Spark arrive in fall 2026. The first partner brands: ASUS (which has already shown ProArt P16/P14 and a mini PC), Dell, HP, Lenovo, Microsoft Surface, MSI, with Acer and Gigabyte following. For consumers waiting on the AI PC transition, fall 2026 shapes up as the moment the category finally becomes purchasable rather than promised.
The Takeaway
Computex 2026 was a milestone for the AI PC. RTX Spark’s launch marks the PC’s evolution from “tool” to “AI teammate” — natural-language interaction, agents completing complex tasks locally, Windows-native agent support from Microsoft, full ecosystem adaptation from Adobe and partners, all inside a thin-and-light form factor. The era of the AI PC has formally arrived; the only remaining question is what the first generation of software written for it will look like.
