Skip to main content
Back

Microsoft's Surface RTX Spark Dev Box Runs 120B AI Models Without Cloud

Microsoft's new compact desktop packs NVIDIA's Blackwell GPU and 128GB of unified memory to run 120B-parameter models locally without touching the cloud

  • What happened: Microsoft announced the Surface RTX Spark Dev Box at Build 2026, a compact AI developer workstation powered by NVIDIA's new RTX Spark chip.
  • Key specs: 1 petaflop of AI compute, 128GB unified memory, 100W thermal envelope, runs 120B-parameter models locally.
  • The chip: NVIDIA RTX Spark combines a 20-core Grace ARM CPU, Blackwell GPU (6,144 CUDA cores), and unified LPDDR5X memory on a single TSMC 3nm package.
  • Out-of-box stack: Ships with VS Code, GitHub Copilot, WSL2 with full CUDA support, and PowerShell 7 pre-installed on Windows 11 Pro.
  • Competition: Directly rivals the NVIDIA DGX Spark ($4,699, Linux) and Apple Mac Studio; CUDA support is the key edge over Apple Silicon for ML workflows.
  • Availability & price: Coming later in 2026 via microsoft.com; no price announced, estimated $3,000-5,000 range; still pending FCC approval.

Microsoft just pulled the wraps off the Surface RTX Spark Dev Box at Build 2026 in San Francisco, and it is one of the most interesting pieces of developer hardware in years. The pitch is simple: a compact desktop that can run frontier-scale AI models entirely on your desk, no cloud API calls required. This is not a consumer PC with an AI sticker slapped on the box. It is a purpose-built local inference and fine-tuning machine.

The chip that makes it possible

Everything starts with NVIDIA's RTX Spark, the company's first-ever system-on-a-chip designed for Windows PCs. RTX Spark fuses a CPU, GPU, and memory controller onto a single package. Unlike traditional designs where the processor and graphics card operate separately with their own memory pools, RTX Spark uses a unified memory architecture similar to Apple Silicon, meaning the CPU and GPU share the same pool of fast LPDDR5X RAM, eliminating the bottleneck that occurs when data has to shuttle between separate chips.

The chip packs a 20-core Grace CPU built alongside MediaTek, a Blackwell RTX GPU with 6,144 CUDA cores, an NPU, and up to 128GB of unified LPDDR5X memory, all on a single TSMC 3nm package. Blackwell is the same GPU architecture powering NVIDIA's RTX 50 series desktop cards. This is reportedly equivalent to an RTX 5070 in terms of CUDA core count. The difference is that no consumer RTX card ships with anywhere near 128GB of VRAM -- and that memory headroom is the whole point.

Three numbers that define the machine

Three numbers define the device: a 100W sustained thermal envelope, up to 128GB of unified memory, and approximately 1 petaflop of AI compute from NVIDIA's RTX Spark chip. To put the memory figure in context: most consumer GPUs top out at 24-32GB of VRAM. Running a 70B-parameter model in full precision typically requires around 140GB. With quantization (compressing model weights to use less memory), 128GB of unified memory is enough to run much larger models comfortably.

  • 1 petaflop of AI compute (measured in FP4 with sparse matrices -- a compressed number format optimized for inference speed)
  • 128GB unified memory shared between CPU and GPU, no VRAM limit
  • 120 billion parameter models can run locally, per NVIDIA
  • 100W thermal envelope -- more sustained headroom than any laptop form factor allows

Zero setup, day one coding

The Dev Box ships with a developer-optimized Windows 11 Pro experience, preconfigured for AI development, minimizing setup time so you can get to a code-ready state faster. The Windows 11 experience is developer-first with Visual Studio Code, GitHub Copilot in Windows Terminal, WSL, and PowerShell 7, all preconfigured and ready to go.

Windows Subsystem for Linux 2 (WSL 2) with support for NVIDIA's CUDA technology will also come out of the box. CUDA is the programming framework that the entire ML ecosystem -- PyTorch, TensorFlow, vLLM, and virtually every inference library -- is built on. This is a key differentiator from Apple Silicon, which requires developers to use Apple's Metal framework instead. For PyTorch and most ML inference toolchains, the CUDA ecosystem is significantly more mature than Metal/MLX. Builders running CUDA-dependent workflows like vLLM or ExLlamaV2 will find the Dev Box a better match than Mac Studio.

The real story: Microsoft's bet on local-first AI

Microsoft used Build 2026 in San Francisco to push AI development away from a cloud-only model, announcing the Surface RTX Spark Dev Box and previewing Project Solara as hardware and platform moves for running agents locally, across PCs, desktops, wearables, and cloud-connected devices. The phrase Microsoft kept repeating at Build was "unmetered intelligence" -- a direct shot at the token-based billing model of cloud AI APIs. The idea is that local AI compute reduces token costs, lowers cloud dependency, and allows developers to iterate faster.

This is Microsoft's clearest admission yet that the next phase of Windows AI cannot live entirely in Azure. If agents are going to be persistent, contextual, and useful outside canned demos, Microsoft needs intelligence closer to the user -- and much closer to the machine. For developers who have been paying per-token to prototype agentic pipelines, the economics of a local machine become compelling fast.

Who wins, who loses

The clearest winner is NVIDIA. The RTX Spark is their first real foothold in the Windows PC silicon market, a space long dominated by Intel, AMD, and Qualcomm. NVIDIA named the following partners who will be developing RTX Spark-powered machines: ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with more to follow from Acer and Gigabyte. 30 laptops and 10 mini desktops have been confirmed. The Surface Dev Box is Microsoft's flagship reference design for this entire ecosystem.

Qualcomm is the most obvious loser. The Snapdragon X series has been the dominant Windows-on-ARM chip for the past two years, but it was never built for sustained AI workloads at this scale. RTX Spark is NVIDIA's direct answer to the unified memory revolution and a direct challenge to Qualcomm's Snapdragon X series that has been dominating the Windows on ARM conversation. Cloud providers also face pressure at the margin -- not for large training runs, but for the iterative, experimental workloads that developers currently route through APIs.

How it stacks up against the competition

DeviceMemoryAI ComputeEcosystemEst. Price
Surface RTX Spark Dev Box128GB unified~1 PFLOPCUDA / WindowsTBD (est. $3-5K)
NVIDIA DGX Spark128GB unified~1 PFLOPCUDA / Linux$4,699
Apple Mac Studio M4 Ultraup to 192GBLower TFLOPSMetal / macOS$4,400+

NVIDIA's DGX Spark is the most direct competitor and is available today. The DGX Spark runs Ubuntu natively and has a similar CUDA-first developer posture. The Surface Dev Box is the Windows-native answer to the same problem. The Mac Studio M3 Ultra offers more raw unified memory at 192GB but has no CUDA support. For teams already in the Windows ecosystem or those who need CUDA without managing a Linux box, the Dev Box fills a real gap.

What is now possible

The practical unlock here is running models like Llama 3 70B, Qwen 72B, or Mistral Large locally at reasonable speeds, with no API costs, no data leaving your machine, and no rate limits. NVIDIA states that the RTX Spark can deliver up to one petaflop of AI compute with 128GB of unified memory and can run models with up to 120 billion parameters locally. That means a developer can run a full agentic pipeline -- retrieval, reasoning, code generation -- entirely offline. For enterprises with data privacy requirements, this is significant.

The Surface RTX Spark Dev Box is a Windows 11 secured-core PC that works with BitLocker, Microsoft Defender, Entra ID, and Intune to help keep sensitive workloads secure. That enterprise security stack, combined with local inference, is a combination that cloud-only setups cannot match for regulated industries.

The open questions

The caveats are real. No pricing has been disclosed and no independent benchmarks exist. Leaked estimates put RTX Spark-based systems starting above $2,899, with the Surface Dev Box likely landing in the $3,000-5,000 range given its specs and branding -- but that is speculation until Microsoft makes it official. The Surface RTX Spark Dev Box is still awaiting FCC approval, so some caution is warranted.

Microsoft hasn't announced a price for the Surface RTX Spark Dev Box, but says it will be available this year at microsoft.com in the United States. While you can't make a pre-order, you can sign up to learn more on the device's web page. If you need a local AI workstation right now, the DGX Spark is the only comparable option shipping today. But if you are building on Windows and want the full CUDA stack without the Linux overhead, the Dev Box is worth watching closely when it ships later this year.