Surface RTX Spark Dev Box Costs $5,999: Is Microsoft's Local-AI Desktop Worth It?

Update, October 8, 2026: Laptop Ultra configuration pricing is now reported (24GB base from $2,599.99, 128GB up to about $5,899.99): Which Surface Laptop Ultra configuration to buy for local AI.

Update, October 8, 2026: What runs on this hardware: MAI-Code-1.1 Flash runs locally on Windows.

Microsoft has now put real prices on its Nvidia RTX Spark hardware. The Surface RTX Spark Dev Box is open for preorder from $5,999.99 in the US and ships in November, while the Surface Laptop Ultra starts at $2,599 with availability from October 16, according to The Gadgeteer's launch coverage. Both machines carry up to 128GB of unified memory, which is the number that matters for running large local models.

The hardware was first teased at Build in early June. At that time Engadget reported that Microsoft had not listed any pricing. Today's launch fills in the missing price.

TL;DR: prices, specs and dates

Sources: Microsoft's product page and The Gadgeteer for the Laptop Ultra details; verify final configurations before ordering.

What exactly is the Surface RTX Spark Dev Box?

Think of it as a small, quiet-ish desktop built for sustained AI work. Microsoft describes it as "the powerhouse for local AI and agent development" with "a petaflop for your desktop": Nvidia's RTX GPU alongside 128GB of unified memory, a single pool shared between CPU and GPU. Microsoft says the intended jobs include long-running training, agentic pipelines and local fine-tuning.

Several design choices stand out from the product page:

• A 100W thermal envelope inside an aluminum body that "doubles as a cooling system," aimed at keeping performance consistent through long runs rather than bursting and throttling.
• 1,000 air vents in a grid chassis, which Microsoft calls a nod to the machine's 1,000 teraflops. That is a design flourish, not a spec.
• Ports: two USB-C, USB-A, HDMI, Ethernet and a headphone jack.
• Ready-to-code Windows: VS Code, GitHub Copilot CLI, WSL and PowerShell 7 are installed and configured, with Windows settings tuned for coding.

The Gadgeteer's summary of the software setup also mentions Git, GitHub CLI, Python and Node. The reasonable takeaway is that Microsoft wants a developer to unbox, log in and be running a local model through WSL within an hour.

How does $5,999 compare with other local-AI boxes?

This is where the launch gets interesting. In its June report Engadget described the Dev Box as Microsoft's answer to Nvidia's DGX Spark and AMD's Ryzen AI Halo PC, and at that time cited $3,999 for both. That figure is stale: as we covered in DGX Spark 64GB: what $4,999 actually buys, Nvidia raised the Founders price to $4,699 in February and is adding an OEM 64GB SKU at $4,999 on October 23.

We have not verified AMD's current pricing, so it is omitted. On these numbers the Surface carries roughly a $1,300 premium over the Founders box. Microsoft's pitch is the Windows developer experience, first-party Surface support and a single vendor for hardware and OS.

If you are comfortable with Linux, that pitch may not move you. If your team lives in Windows, standardizes on Copilot and WSL, and wants a machine IT can manage, it might. For more on the Linux and Windows tradeoffs for local inference, see our coverage of the Perplexity Portable Computer with an RTX GPU and the earlier jamesob RTX 6000 Pro local-LLM build.

Can 128GB really run 120B+ models?

Microsoft says the Dev Box targets local models with more than 120 billion parameters. Simple arithmetic shows why 128GB is the threshold: a 120B-parameter model at 8-bit precision needs roughly 120GB for weights alone, leaving almost nothing for the KV cache, the operating system or your editor. At 4-bit quantization the weights drop to about 60GB, which leaves room for long contexts and a few other processes.

Two caveats from the sources are worth keeping in mind:

1. Unified is not the same as GPU-addressable. The Gadgeteer notes that for the Laptop Ultra "the amount the GPU can address is lower than the system's total and depends on the configuration and workload." Expect the same logic on the desktop.
2. Petaflop claims depend on precision. Headline AI-compute figures are usually quoted at low-precision formats. Real token-per-second numbers on a dense 120B model are bounded by memory bandwidth, not peak flops. Wait for independent benchmarks before assuming interactive speeds.

That is a pattern we flagged around the original chip announcement; our recap of Nvidia's Computex 2026 keynote and the N1X Arm laptop chip piece explain the Grace-plus-Blackwell design behind RTX Spark.

What about the Surface Laptop Ultra?

At $2,599 the Laptop Ultra is the cheaper way into the same platform. The Gadgeteer lists a 15-inch PixelSense Ultra mini-LED touchscreen with up to 2,000 nits of peak HDR brightness (measured over a 10 percent window, so sustained brightness is lower), up to 20 CPU cores, 6,144 GPU cores and 128GB of unified memory. It supports up to three 4K external displays, has magnetic USB-C charging, a full-size SD card reader and a removable storage drive.

The tradeoffs are plain. The touchscreen supports fingers only; Surface Pen and Slim Pen are unsupported. And, as The Gadgeteer cautions, buyers should choose the memory and processor configuration for the applications they will run rather than assume that the starting price includes the maximum specs. In other words, the $2,599 laptop is probably not the 128GB model. We do not have configuration-level pricing, so check Microsoft's store.

Windows is changing too

The same Microsoft event announced new Windows features: new ways to run AI agents, to combine local models with cloud services, and to carry out tasks directly from Search. We have not independently confirmed the details beyond The Gadgeteer's summary, so treat them as announced rather than shipped. The strategic point is that Microsoft is pairing hardware with an OS story: local inference on the box, agents orchestrated by Windows, and cloud models as a fallback.

If you are building agents that talk to tools, the Model Context Protocol guide explains the plumbing most of these Windows agent features are likely to rely on, though Microsoft has not spelled that out in the sources we reviewed.

Should you buy, wait or skip?

Buy now if you have a funded Windows-based team, need a managed on-prem box for sensitive data, and the 128GB pool covers your target model with headroom. Preorder windows for first-generation hardware often come with the longest waits and the least feedback, but you will have a November delivery estimate.

Wait if you can tolerate a few weeks. Independent reviewers will have the Laptop Ultra from October 16, which gives the first real performance data for the same chip and memory architecture. Tokens-per-second on a 70B and a 120B model will tell you whether the Dev Box justifies its premium.

Skip if you only run models up to about 30B parameters. A conventional GPU workstation or a high-memory Mac will serve you at lower cost. Apple's side of this story is covered in our posts on the M6 Mac mini and M5 Ultra Mac Studio and on-device LLMs on the M6 Mac mini.

Or consider the software side first. Nvidia's RTX Spark platform is also the target for local agent tooling. Our piece on Nvidia and Hermes Desktop on RTX Spark shows what that stack looks like before you spend $6,000.

Questions we cannot answer yet

• Memory bandwidth. Neither source quotes it, and it determines generation speed for large dense models.
• Real-world sustained power. A 100W envelope sounds modest next to discrete-GPU desktops; sustained throughput depends on clocks.
• Availability outside the US. The Dev Box preorder is US-only through Microsoft. No international date was given.
• Upgradeability. The laptop has a removable drive; the Dev Box's service story is not described.
• Return policy and support terms on a $6,000 preorder.

We will update this post when independent benchmarks land, starting with the Laptop Ultra reviews after October 16.

Related reading

• Nvidia and Hermes Desktop on RTX Spark
• Nvidia N1X Arm laptop chip at Computex 2026
• Nvidia Computex 2026 and Nemotron 3 Ultra recap
• Perplexity Portable Computer with an RTX GPU
• Running state-of-the-art LLMs locally on an RTX 6000 Pro
• Apple M6 Mac mini and M5 Ultra Mac Studio for AI
• What is MCP? Model Context Protocol guide
• Official: Microsoft Surface RTX Spark Dev Box

Prices, availability and specifications are as reported on October 7, 2026 and may change before shipping. Follow @explainx_ai for updates.