6,144. That’s the GPU core count on the top-end Surface Laptop Ultra, the Nvidia-powered machine Microsoft debuted in San Francisco on October 7, 2026. For $3,699.99 you get those cores paired with a 20-core CPU. Step down to $2,599.99 and you’re looking at 5,120 GPU cores and an 18-core CPU. Both run on the Arm-based Nvidia RTX Spark processor, and both are up for preorder on Microsoft’s site right now in Platinum and Nightfall.
I build bots for a living. My reaction to spec sheets is usually a shrug, because the thing that actually determines whether I ship on time is rarely raw silicon. But these numbers made me stop and reconsider a decision I’ve made by default for about three years: renting compute instead of owning it.
The rental habit nobody questions
If you’ve built anything involving language models, you know the pattern. You prototype against an API. You hit rate limits. You bump your tier. You discover that your conversational agent, the one that felt clever in testing, costs real money every time someone asks it a dumb question. And because your laptop can’t run anything useful locally, the API stays in the loop forever, including during development, including during the fifty-seventh time you tweak a system prompt and re-run the same test suite.
That loop is where local compute earns its keep. Not in production necessarily, but in the grind of building. Every iteration that doesn’t leave your machine is an iteration that costs nothing, leaks nothing, and doesn’t fail because someone else’s data center had a bad afternoon.
What Microsoft is actually aiming at
The positioning here is explicit: Microsoft is going after Apple’s MacBook Pro on AI work, video workloads, and gaming. Nvidia’s pitch for the RTX Spark chip is that it “delivers amazing creating, AI development, and gaming” in slim laptops and smaller desktops. Note the middle item in that list. AI development, called out specifically, on a laptop.
That framing matters for anyone in our corner of the world. The MacBook Pro became the default machine for a lot of local model work almost by accident, on the back of unified memory and a chip architecture that happened to suit the workload. Microsoft looked at that and decided not to cede the category. Satya Nadella and Jensen Huang sharing a stage is a fairly direct statement about who’s supposed to buy this thing.
Where I’m skeptical
Arm on Windows has a history, and it isn’t a kind one. The RTX Spark is Arm-based, which means the question every bot builder should ask before preordering is boring and unglamorous: does my toolchain run? Not “is there a port in progress.” Does it run, today, without me spending a weekend compiling things.
Python itself is fine. The pain lives downstream in native extensions, in CUDA assumptions baked into libraries, in that one dependency three layers deep that ships x86 wheels only. Container workflows add another layer of questions. None of this is unsolvable, and none of it is reflected in a core count.
The price is the other thing. $2,599.99 is a serious entry point, and $3,699.99 buys a lot of API credits. If you’re building bots that mostly orchestrate hosted models, stitching together retrieval and tool calls and routing logic, the extra cores buy you very little. That work is network-bound, not compute-bound. A cheap machine and a good connection does it.
Who should actually care
The case gets strong in a few specific situations:
- You’re fine-tuning smaller models and the feedback loop on hosted training is killing your momentum.
- You’re working with data that legally cannot leave your machine, which rules out the API-first path entirely.
- You’re building multimodal pipelines where video processing and model inference sit in the same workflow, and shuttling large files to a remote GPU is the slow part.
- You iterate constantly on prompts and agent logic, and your per-experiment cost has quietly become a line item you’d rather not explain.
If none of those describe you, this is a very nice laptop that will mostly run your editor.
The practical move
Preorders are open, which means nobody outside of Microsoft and Nvidia has run a real workload on this hardware yet. Everything above is reasoning from specifications and stated intent, not benchmarks. I’d wait for someone to try an actual local inference stack on an RTX Spark machine and report back honestly about what broke.
What I find more interesting than the laptop is the direction it implies. Two companies with enormous incentives to keep AI work in their data centers just put serious effort into a machine designed to do that work locally. They’re betting that a meaningful number of developers want the option. Based on my own cloud bill, that’s not a bad bet.
🕒 Published: