What exactly are you buying when a laptop is sold You’re buying a bet that the work you do tomorrow will run better on your desk than in someone else’s data center. Microsoft just put a number on that bet: $2,599.99.
That’s the entry price for the Surface Laptop Ultra, the first Surface built around Nvidia’s RTX Spark chip. Preorders are open on Microsoft’s site in Platinum and Nightfall. The base config gets you a 5120-core GPU paired with an 18-core CPU. Step up to the 6144-core GPU and 20-core CPU and you’re at $3,699.99. Microsoft is pointing this squarely at the MacBook Pro, and framing the whole thing as its answer for the AI era.
What Microsoft is actually claiming
Nvidia’s pitch for RTX Spark is that it “delivers amazing creating, AI development, and gaming” in slim laptops and smaller desktops. That’s three workloads with three very different thermal and memory profiles, which tells you something about how this silicon is being positioned: one chip, sold to everyone who has ever complained their machine was too slow.
Microsoft’s own framing for the Ultra is about creative pros and makers who don’t want to pick between power and portability. Fair enough as marketing. The Laptop Ultra isn’t alone either, there’s a Surface RTX Spark Dev Box in the same family, and The Verge’s Tom Warren got hands-on with both at Build. A laptop plus a dev box built on the same chip is a more interesting signal than either device alone. It suggests Microsoft wants the local development story to be a product line, not a one-off halo device.
Three things I’d want answered before preordering
I review tools for a living, and the gap between a spec sheet and a Tuesday afternoon is where most of them fall apart. Here’s where my skepticism sits:
- Sustained performance. Core counts describe a ceiling, not a floor. A slim chassis running a local model for forty minutes is a thermal problem before it’s a compute problem. Nobody has published numbers I’d trust yet.
- Memory, and how much of it the GPU can actually see. For local AI work, usable memory is often the hard wall, not GPU cores. Microsoft hasn’t foregrounded this in what’s been announced, and that’s the spec I’d read first.
- Software maturity. New silicon means new drivers, new toolchain support, and a window where your framework of choice technically runs but runs badly. That window has a habit of lasting longer than launch coverage suggests.
Run the math on local versus rented
The honest question for anyone doing AI development isn’t whether the Ultra is fast. It’s whether owning compute beats renting it for your specific workload.
Local wins when you’re iterating constantly, working with data you can’t send anywhere, or tired of a meter running while you think. Those are real advantages and they’re underrated. Latency on your own machine is effectively zero, and there’s a creative difference between experimenting freely and watching a billing dashboard.
Rented wins when your workloads are bursty, when you need more memory than any laptop will ever carry, or when you need a different accelerator next quarter. At $3,699.99, the top Ultra represents a meaningful chunk of cloud credits. If your actual pattern is a heavy training run twice a month, you’re buying a very nice laptop and using it as a very expensive text editor between runs.
The group I’d actually recommend this to: people whose work is genuinely mixed. Video and 3D work during the day, model fine-tuning at night, maybe gaming after. A single machine that handles all three has real value, and that mix is exactly what Nvidia says Spark is built for.
What the MacBook Pro comparison tells us
Microsoft naming its target is useful. Apple’s argument has been that unified memory and efficiency let you do serious work unplugged. Microsoft’s counter is discrete-class Nvidia hardware in a portable, which means CUDA, which means most AI tooling works without translation layers or hunting for compatible builds. That’s not a small thing for developers. It’s probably the strongest practical argument in this device’s favor.
Whether it holds depends on battery life and heat, the two areas where the Apple comparison has historically gone badly for Windows machines.
My take
Interesting hardware, legitimate reason to exist, and a price that demands you be honest about your own workflow before you commit. I don’t preorder machines built on first-generation silicon, and I wouldn’t tell you to either. Wait for independent sustained-load testing and clear memory specs. If you’re the mixed-workload creative-plus-developer this was designed for, the case gets stronger fast. If you mostly need AI compute in bursts, your money probably goes further elsewhere.
🕒 Published: