NVIDIA and Microsoft introduced NVIDIA RTX Spark, a superchip for a new class of Windows laptops and compact desktops built for on-device AI agents, at NVIDIA GTC Taipei on May 31, 2026 (NVIDIA; Microsoft). Per NVIDIA, RTX Spark pairs a Blackwell RTX GPU (6,144 CUDA cores, fifth-generation Tensor Cores) with a 20-core NVIDIA Grace Arm CPU co-designed with MediaTek, linked by NVLink-C2C, plus up to 128GB of unified memory and up to 1 petaflop of AI compute at FP4 precision. NVIDIA says the platform can run 120-billion-parameter models with up to 1 million tokens of context locally, and adds new Windows security primitives and the NVIDIA OpenShell runtime for private agents. Adobe is rearchitecting Photoshop and Premiere for it. ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI will ship RTX Spark laptops and desktops this fall, with Acer and GIGABYTE to follow (NVIDIA; WSJ; Tom's Hardware).
What happened
NVIDIA and Microsoft used NVIDIA GTC Taipei on May 31, 2026 to introduce NVIDIA RTX Spark, a superchip that anchors what the companies call the first Windows PCs purpose-built for on-device AI agents (NVIDIA; Microsoft). NVIDIA framed the launch as a reinvention of the personal computer, with CEO Jensen Huang saying "The PC is being reinvented" and positioning RTX Spark as "the new PC. The personal AI computer" (NVIDIA). WSJ reported NVIDIA billed the chip as "the most efficient PC chip ever built" (WSJ). RTX Spark laptops and compact desktops are due this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface and MSI, with Acer and GIGABYTE to follow; Tom's Hardware reported more than 30 laptops and 10 desktops are planned (NVIDIA; Tom's Hardware).
Inside the superchip
Per NVIDIA, RTX Spark pairs a Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores (FP4 precision) to a 20-core NVIDIA Grace Arm CPU over the NVLink-C2C interconnect, with up to 128GB of unified memory and up to 1 petaflop of AI compute (NVIDIA). NVIDIA said MediaTek co-designed the custom Arm CPU for power efficiency and connectivity (NVIDIA). Editorial analysis - technical context: the headline 1-petaflop figure is rated at FP4, a 4-bit format, so it reflects low-precision tensor throughput and is not directly comparable to the higher-precision FLOPS often cited elsewhere. The unified-memory design keeps model weights and long context next to the GPU, which is what makes large local models viable on a thin device.
Agents and the Windows partnership
NVIDIA and Microsoft said the platform can run 120-billion-parameter models with up to 1 million tokens of context locally, and pairs new Windows security primitives with the NVIDIA OpenShell runtime so agents run under user-defined policy, route queries to local models for privacy, and mask personal data sent to cloud models (NVIDIA; Microsoft). Microsoft CEO Satya Nadella said "Our goal is to deliver unmetered intelligence to every home and every desk with Windows," calling RTX Spark "a real breakthrough towards that vision" (NVIDIA). NVIDIA named open-source agent projects Hermes Agent, from Nous Research, and OpenClaw as early adopters of the new Windows stack (NVIDIA).
Creator and gaming stack
NVIDIA said RTX Spark carries its full CUDA, RTX, TensorRT, OptiX, DLSS and Reflex stack, targeting rendering of 90GB 3D scenes, 12K 4:2:2 video editing, 4K AI video, and AAA gaming at 1440p above 100 fps (NVIDIA). Adobe is rearchitecting Photoshop and Premiere for the platform for up to 2x faster AI and editing, and NVIDIA listed more than 100 software partners, including Blender, ComfyUI, Blackmagic Design and OTOY (NVIDIA).
Industry context
Independent coverage frames RTX Spark as a Windows-on-Arm platform, putting NVIDIA into direct competition with Intel and AMD x86 chips and Qualcomm's Arm-based Windows processors in the PC market (Tom's Hardware). Editorial analysis: heterogeneous Arm-CPU-plus-GPU designs with large unified memory track a broader industry shift toward running inference and long-context workloads on the client rather than the cloud; for practitioners that lowers the latency and cost of local agents, while concentrating more of the Windows AI stack on a single vendor's silicon. NVIDIA paired the consumer launch with NVIDIA DGX Station for Windows for enterprise deskside agents, signaling a top-to-bottom Windows push (NVIDIA).
What to watch
- •Independent silicon: third-party benchmarks of model latency, throughput and battery life on shipping units versus NVIDIA's FP4-based claims.
- •Software readiness: when NVIDIA and Microsoft ship developer toolkits, drivers, and OpenShell and security-primitive documentation that the local-agent claims depend on (more detail expected at Microsoft Build, June 2-3).
- •Pricing and availability: NVIDIA has not disclosed pricing, and the fall ship window and per-OEM configurations remain to be confirmed.
- •Ecosystem follow-through: Adobe, ComfyUI and llama.cpp optimizations, and whether OEM thermals sustain the advertised workloads.
Scoring Rationale #
NVIDIA and Microsoft launched RTX Spark, a Blackwell-plus-Grace-Arm Windows superchip with 128GB unified memory that NVIDIA says runs 120B-parameter models with 1M-token context locally, backed by every major OEM, Adobe and 100-plus software partners. For AI/ML practitioners it is a major platform for local agents and large on-device inference, and it marks NVIDIA's first real entry into the Windows PC chip market against Intel, AMD and Qualcomm. Scored just below the industry-shaking tier because the hardware does not ship until fall and the headline numbers are vendor FP4 claims with no independent benchmarks or pricing yet.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.