6,144. That’s the GPU core count on the top configuration of Microsoft’s new Surface Laptop Ultra, paired with a 20-core CPU and a $3,699.99 price tag. Step down to the $2,599.99 model and you get 5,120 GPU cores with an 18-core CPU. Both run Nvidia’s RTX Spark chip, both are up for preorder on Microsoft’s site, and both come in Platinum or Nightfall.
Microsoft is pitching this thing at the MacBook Pro, which tells you who it thinks the buyer is: creative pros, developers, people who have been grudgingly paying the Apple tax because the hardware held up. Nvidia’s own framing is that RTX Spark “delivers amazing creating, AI development, and gaming” in slim laptops and small desktops. Fine. I build bots for a living, so let me translate what that actually means for people like us.
Local inference stops being a hobby project
Most of the bot developers I talk to have settled into the same split workflow. You prototype against a hosted API, you iterate on prompts in a browser tab, and when you need to test a local model you either spin up a rented GPU instance or you accept that your laptop is going to sound like a hair dryer for the next forty minutes.
A discrete Nvidia GPU with this kind of core count changes the shape of that workflow. CUDA is still the path of least resistance for anything model-related, and that matters more than raw benchmark numbers. Every quantization tool, every fine-tuning script, every inference server you’d reach for on a Tuesday afternoon assumes CUDA exists. When your dev machine speaks the same language as your deployment target, you stop writing two versions of everything.
For bot architecture specifically, the useful part is the feedback loop. Testing a retrieval pipeline against a locally hosted model means you can run the embedding step, the vector search, and the generation all on one machine with no network hop and no per-token meter running. You iterate faster because nothing is waiting on an API queue. You also get to test the failure modes that only show up when you’re running the model yourself, like what happens to your conversation state when generation stalls mid-stream.
What Microsoft did not tell us
Core counts make for good slides and poor engineering decisions. The numbers that actually determine whether you can run a mid-sized model comfortably are the ones Microsoft hasn’t put front and center. How much memory does the GPU get, and is it unified or dedicated? What’s the memory bandwidth? What happens to sustained clock speeds after twenty minutes of load inside a chassis thin enough to be called slim?
That last one is not a nitpick. Thermal behavior is the difference between a laptop that fine-tunes a small model overnight and one that throttles into uselessness halfway through. Nvidia says RTX Spark is built for slim laptops and smaller desktops, which suggests the power envelope was a design constraint rather than an afterthought. I’d still want to see a long-running job before I made a purchase decision.
The Dev Box is the more interesting device
Microsoft announced two RTX Spark machines. The Verge’s Tom Warren got hands-on with both at Build, and the second one is the Surface RTX Spark Dev Box. A desktop form factor with the same silicon family removes the thermal ceiling that limits any laptop, and for anyone whose bots live on a server rather than a demo screen, a small desktop that sits under your desk and takes jobs over the network is a more natural fit than a $3,699 portable.
My guess is a lot of teams end up wanting both, which is probably the point. The laptop for building and demoing, the box for the work that runs while you sleep.
Is it worth the money
Depends entirely on what you build. Some honest framing:
- If your bots are thin wrappers around a hosted API, this is an expensive laptop that happens to play games well. Your bottleneck is prompt design, not compute.
- If you’re fine-tuning, running open-weight models, or building anything with local embeddings and retrieval, owning the hardware starts to make sense. Predictable cost, no network dependency, no data leaving your machine.
- If you’re deciding between this and a MacBook Pro, the CUDA ecosystem is the real argument. Apple silicon runs local models fine, but you’ll spend time on tooling translation that the Nvidia path skips entirely.
The $1,100 gap between configurations is a lot for 1,024 extra GPU cores and two CPU cores. Without memory specs, I’d be cautious about assuming the top model is the better buy. Memory capacity is usually what stops you from loading a model, not core count.
What I like about this announcement is less the hardware and more what it signals. Local AI development is being treated as a mainstream workload worth designing a flagship laptop around. For those of us who have been duct-taping dev environments together for years, that’s a welcome change.
🕒 Published: