The Surface RTX Spark Dev Box is Microsoft’s answer to the Mac Studio for AI developers. Announced at Build 2026, it packs NVIDIA’s RTX Spark superchip (128GB unified memory, 1 petaflop AI compute) into a compact aluminium chassis — preloaded with every developer tool you need to start running AI models locally out of the box.
This is not a consumer PC. It is a dedicated AI development workstation designed for one purpose: running and building with AI models locally.
Specs
| Chip | NVIDIA RTX Spark (N1X) |
| CPU | 20-core NVIDIA Grace (ARM) |
| GPU | Blackwell (6,144 CUDA cores) |
| Memory | 128GB unified (CPU + GPU shared) |
| AI compute | 1 petaflop |
| Thermal | 100W sustained (aluminium chassis as heatsink) |
| OS | Windows 11 Pro |
| Form factor | Compact desktop (mini PC) |
| Availability | Fall 2026 |
What’s preinstalled
The Dev Box ships ready-to-code:
- Windows 11 Pro
- Visual Studio Code
- GitHub Copilot (with MAI-Code-1-Flash)
- WSL2 with GPU passthrough (full Linux + CUDA on Windows)
- CUDA toolkit
- Python
- Git
- Node.js
- PowerShell 7
No setup. Unbox, power on, run ollama pull qwen3.7:27b, start coding with AI. This is the value proposition: zero-configuration AI development.
What models run on it
128GB unified memory supports models up to ~120B parameters. For the full breakdown, see Best LLMs for RTX Spark. Highlights:
| Model | Memory needed | Speed (est.) |
|---|---|---|
| Qwen 3.6/3.7 27B | 16GB (Q4) | 40-60 t/s |
| Llama 4 Scout | 60GB (Q4) | 20-35 t/s |
| Mistral Medium 3.5 | 24GB (Q4) | 30-50 t/s |
| Step 3.7 Flash | 100GB (Q4) | 15-30 t/s |
| 120B dense model | 70GB (Q4) | 10-18 t/s |
Plus Aion 1.0 models for on-device agent workflows.
Surface RTX Spark Dev Box vs Mac Studio
| Surface RTX Spark Dev Box | Mac Studio M4 Ultra 128GB | |
|---|---|---|
| Memory | 128GB unified | 128GB unified |
| GPU | Blackwell (1 PFLOP) | Apple GPU (~27 TFLOPS) |
| CUDA | ✅ (native + WSL2) | ❌ |
| OS | Windows 11 Pro | macOS |
| Preloaded dev stack | ✅ (VS Code, Copilot, CUDA, Python) | ❌ (manual setup) |
| WSL2 Linux | ✅ (GPU passthrough) | ❌ |
| Form factor | Mini PC | Mini PC |
| Thermal | 100W sustained (aluminium) | Silent (fan-based) |
| Price (est.) | TBD (~$3,000-5,000?) | $3,999 |
| Available | Fall 2026 | Now |
The Dev Box’s main advantages: CUDA support (critical for many AI tools), preloaded dev stack (zero setup), and WSL2 GPU passthrough (run Linux AI tools natively). Mac Studio’s advantages: available now, silent operation, proven ecosystem.
For the full comparison, see RTX Spark vs Mac Studio.
vs Building your own
A custom AI workstation with 128GB requires:
- Multi-GPU setup (2× RTX 5090 = only 64GB total) — doesn’t match unified 128GB
- Or expensive server hardware (A100 80GB = $15K+)
- Manual software configuration
- No warranty/support from one vendor
The Dev Box solves this with a single integrated package. 128GB unified memory in a form factor you can’t replicate with discrete GPUs.
vs NVIDIA DGX Spark
| Surface RTX Spark Dev Box | DGX Spark | |
|---|---|---|
| OS | Windows | Linux |
| Target | Developers (code + AI) | ML researchers (always-on) |
| Form factor | Mini PC (desktop) | Deskside workstation |
| GPU tier | Consumer Blackwell | Data-center Blackwell |
| Dev stack | ✅ Preloaded | Manual |
| 24/7 operation | Not primary use case | ✅ Designed for it |
Choose the Dev Box if you want Windows + development tools. Choose DGX Spark if you need an always-on Linux inference server.
Who should buy this
✅ Buy if you:
- Develop AI-powered applications and need local model testing
- Currently rent cloud GPUs for AI development ($100+/mo)
- Need CUDA support that Mac Studio cannot provide
- Want zero-configuration AI development (unbox → code)
- Value Windows + WSL2 for the best of both worlds
❌ Skip if you:
- Already have a Mac Studio 128GB (similar capability, different ecosystem)
- Only need small models (14-27B) that run on any RTX 4090
- Primarily use cloud APIs and don’t run models locally
- Need it before fall 2026 (not available yet)
FAQ
When does it ship?
Fall 2026, alongside consumer RTX Spark laptops. No exact date or pricing.
How much will it cost?
Not announced. Estimate: $3,000-5,000 based on specs (128GB unified memory + Blackwell GPU + Surface branding). Likely comparable to Mac Studio pricing.
Can I upgrade the memory?
Unlikely. Unified memory architectures (like Apple Silicon and now RTX Spark) are soldered. 128GB is what you get. Choose wisely.
Does it run Linux natively?
Windows 11 Pro with WSL2 + full GPU passthrough. This gives you a Linux environment with CUDA access. Not bare-metal Linux — for that, get DGX Spark.
Is it silent?
Aluminium chassis as heatsink with 100W sustained thermal design. Likely quieter than a traditional GPU workstation but not silent like a Mac Studio. Wait for reviews.
Can it replace a cloud GPU subscription?
For inference (running models): yes, for models ≤120B. For training large models: no. See RTX Spark vs Cloud GPUs for the break-even analysis.