NVIDIA and Microsoft have jointly unveiled RTX Spark, a new category of Arm-based Windows 11 PCs purpose-built for local AI workloads. Announced on May 31, 2026, the RTX Spark platform pairs a 20-core NVIDIA Grace CPU with a next-generation Blackwell RTX GPU, unified memory configurations of up to 128 GB, and an AI compute capability reaching 1 petaflop. This is not just another laptop or desktop — it is the first hardware platform to bring full NVIDIA CUDA acceleration to Windows on Arm, with an explicit focus on running large language models, AI agents, and other compute-intensive workloads entirely on-device.
For years, the promise of Windows on Arm has been defined by efficiency and connectivity, spearheaded by Qualcomm’s Snapdragon processors. But the arrival of RTX Spark signals a shift in strategy: Arm-based Windows machines are no longer just about all-day battery life and always-connected PCs. Together, NVIDIA and Microsoft are positioning the platform for the AI era, where local compute capability determines the utility of the device. The RTX Spark represents the confluence of three critical technologies: NVIDIA’s data center-grade Grace CPU design, the latest Blackwell GPU architecture with dedicated AI accelerators, and a deeply integrated software stack that fully leverages CUDA on Arm.
The heart of RTX Spark is NVIDIA’s Grace CPU, a 20-core Arm Neoverse-based processor originally designed for hyperscale AI workloads. In RTX Spark, it delivers not just high single-threaded performance but also the memory bandwidth and power efficiency needed to feed a large GPU over a high-speed NVLink-C2C interconnect. This chiplet design bonds the CPU and GPU with 900 GB/s of coherent bandwidth, effectively creating a single, massive compute complex. The unified memory architecture — available in 64 GB and 128 GB configurations — allocates a shared pool of LPDDR5X memory directly accessible by both CPU and GPU. For AI practitioners, this removes the traditional bottleneck of PCIe transfers and allows models of up to 70 billion parameters to run locally with minimal quantization.
The Blackwell RTX GPU inside RTX Spark includes the latest fourth-generation Tensor Cores and a dedicated Transformer Engine, which specializes in accelerating inference and fine-tuning of transformer-based neural networks. Combined with 128 GB of unified memory, developers can load entire models into VRAM-like space without fragmentation or copy overheads. This is a game-changer for local agentic AI systems that need to keep multiple models in memory simultaneously — for example, a vision model, a speech recognition model, and a large language model all cooperating in real time. NVIDIA claims the peak AI throughput is 1 petaflop of INT8 tensor operations, a figure that surpasses many previous-generation cloud GPU instances and puts desktop AI on par with high-end data center accelerators from just a few years ago.
From a software perspective, the most significant break from the past is native CUDA support. CUDA — NVIDIA’s parallel computing platform and programming model — has historically been tightly coupled to x86 hosts. With the move to Grace-based Arm systems in the data center, NVIDIA wrenched CUDA open for AArch64. The RTX Spark extends that to Windows. Developers can now install CUDA Toolkit 12.8 or later directly onto a Windows 11 on Arm installation and compile or run existing CUDA C/C++ applications without modification. NVIDIA’s proprietary Arm-compatible driver stack, built in collaboration with Microsoft, maintains full feature parity with x86 GPUs, including support for CUDA streams, graphs, and MPS (Multi-Process Service).
The operating system layer has been optimized as well. Microsoft has been steadily improving Windows 11 on Arm for years, delivering ARM64EC emulation to bridge legacy x64 binaries and a growing catalog of native Arm applications. With the RTX Spark announcement, Microsoft confirmed that all Windows AI subsystem components — from the ONNX Runtime to DirectML to the Windows Copilot Runtime — now ship with native ARM64 binaries tailored for the Blackwell GPU. Even the Windows Subsystem for Linux (WSL) gains full access to the CUDA stack, meaning that Linux-based AI workflows can run unmodified in a WSL2 Arm virtual machine while tapping into the full power of the GPU. The combination of Windows’ familiar productivity environment with the raw AI capacity of a supercomputer under the desk is a new proposition that neither Apple’s Macs nor traditional x86 workstations can match today.
Apple has demonstrated the potential of unified memory architectures with its M-series chips, where the CPU and GPU share a high-bandwidth memory pool. However, Apple’s AI framework — Core ML with the ANE (Apple Neural Engine) — lacks the programmability and ecosystem breadth of CUDA. Developers targeting AI workloads on macOS must either refactor code into Apple’s Metal API or resort to cloud-based solutions. RTX Spark delivers the same unified-memory advantage while retaining the full NVIDIA AI stack: CUDA, cuDNN, TensorRT, Triton Inference Server, and all the major deep learning frameworks. The result is a developer experience that requires zero code changes when moving from a cloud p4d.24xlarge to a local RTX Spark box.
Qualcomm’s Snapdragon X Elite and X Plus, the incumbents in the Windows on Arm space, have already pushed the boundaries of efficient computing with integrated AI engines hitting up to 45 TOPS. Yet they rely on a heterogeneous compute model with smaller, shared memory pools and less powerful integrated graphics. While they excel at ultrabook form factors and the kinds of lightweight AI features found in Windows Studio Effects, they cannot host a large language model for agentic workloads. RTX Spark is designed first and foremost as a developer workstation or a high-end desktop replacement, with ample thermal headroom and a focus on raw throughput. It is not a direct competitor to thin-and-light Qualcomm devices but rather a complementary tier that establishes a new performance ceiling for Arm-based Windows systems.
The practical use cases for RTX Spark span far beyond simple chat bots. With enough unified memory and compute to run frontier models like Llama 4 70B at full precision, developers can build complex AI agents that reason, plan, and execute tasks over extended periods — all locally, with no data leaving the device. This addresses growing enterprise concerns about data privacy, latency, and recurring cloud costs. A financial analyst could run a confidential research agent that reads thousands of documents, cites sources, and generates reports without ever pinging an external API. A game developer could use an AI-assisted design agent to generate textures, write dialogue, and even test gameplay logic in a fully self-contained environment. Creative professionals gain the ability to chain multiple AI models for video editing, 3D rendering, and real-time audio transcription simultaneously, all on a single device.
NVIDIA also announced RTX Spark support for its NIM (NVIDIA Inference Microservices) platform. NIM packages pre-optimized AI models into containers that developers can deploy locally. Combined with Microsoft’s Azure AI Studio for Windows, users can orchestrate hybrid workflows where fine-tuning is done locally on RTX Spark and inference is scaled to the cloud when needed. The same CUDA code and model artifacts move seamlessly between environments, giving organizations a flexible path that starts with local development before expanding to production deployments.
The hardware form factor of RTX Spark remains somewhat opaque. NVIDIA shared early renders of a compact desktop enclosure reminiscent of the Mac Studio, complete with quiet cooling and a range of I/O: Thunderbolt 5, USB4, 10Gb Ethernet, and display outputs capable of driving four 6K monitors. The device is expected to consume approximately 200–300 watts under peak load, which places it in a similar power envelope to a high-end gaming PC or a mid-range workstation. Microsoft is working with its usual stable of OEM partners — Dell, HP, Lenovo, and Samsung — to produce their own branded RTX Spark systems beginning later this year. The Surface team is also said to be developing a premium all-in-one desktop powered by the platform, though no launch timeline was confirmed.
Pricing remains the elephant in the room. With 128 GB of LPDDR5X and a top-tier Blackwell GPU, RTX Spark will almost certainly cost several thousand dollars. NVIDIA avoided announcing a specific number, but based on data center Grace-Blackwell pricing and the cost of equivalently equipped workstation GPUs like the RTX 6000 Ada Generation, a fully loaded Spark could easily land between $8,000 and $12,000. That positions it as a professional tool, not a consumer gadget. However, for businesses that currently spend tens of thousands per month on cloud GPU instances, a one-time hardware purchase that lives on an employee’s desk quickly becomes an economic no-brainer. Microsoft hinted at special enterprise subscription models that bundle Windows 11 Enterprise, Azure credits for hybrid bursting, and priority support, making the total cost of ownership more predictable.
Critically, the Arm architecture does not come without baggage. While the ARM64EC emulation layer is impressive, a non-trivial fraction of legacy Windows software — particularly drivers for specialized peripherals and some deep-rooted system utilities — still ships only as x64 binaries and may exhibit performance penalties or outright incompatibility. NVIDIA and Microsoft emphasized that all core developer tools, including Visual Studio 2026, .NET 10, Python 3.13, and all major AI frameworks, are now native ARM64 first-class citizens. But enterprises with deep dependencies on specific legacy applications will need to audit compatibility before adopting RTX Spark as a standard desktop. Microsoft is expanding its App Assure program to help with this transition, offering free engineering resources to address any blocking issues for strategic accounts.
The announcement also included the standard “coming soon” placeholder for developer kits. NVIDIA confirmed that select ISVs and research institutions will receive early access systems in July 2026, with general availability of the first OEM models targeted for October 2026, just in time for the holiday build season. A developer-focused preview of the NVLink-C2C interconnect API will ship alongside to allow third-party vendors to design custom software that fully exploits the CPU-GPU coherency. Microsoft separately announced that Windows 11 version 26H2, expected in September, would include a new “AI Optimize” power plan specifically tuned for RTX Spark, which dynamically adjusts memory frequency and core clocks based on the active AI pipeline.
The broader implications for the Windows ecosystem are substantial. For over a decade, x86 has been the uncontested king of the Windows PC. Arm began as a niche for fanless laptops, but with RTX Spark, it now challenges x86 on its home turf: raw performance. NVIDIA’s entry also means that Microsoft has a credible answer to Apple’s hardware-software integration, one that leverages the world’s largest GPU ecosystem rather than building a walled garden from scratch. It’s a bold bet that the future of the PC is not just an AI-infused thin client, but a self-contained AI powerhouse capable of running the most demanding models without a cloud connection. That bet might just reshape what we expect from a Windows computer — and RTX Spark is the first piece of hardware designed from the ground up to live up to the promise.