Do you actually need an Nvidia GPU bolted into your laptop to do AI work in 2026? Microsoft is betting $2,599.99 that your answer is yes, and I think a lot of people are about to find out they were wrong about their own workflow.
At its October 2026 Windows and Surface event in San Francisco, Microsoft opened preorders for the Surface Laptop Ultra, built around Nvidia’s RTX Spark chip. Two configurations: $2,599.99 gets you a 5120-core GPU paired with an 18-core CPU, and $3,699.99 moves you up to 6144 GPU cores and 20 CPU cores. Platinum or Nightfall. Microsoft’s pitch, repeated since the device was first shown in June, is that it’s for “world makers and creative pros who refuse to choose between power and portability.”
That’s a real customer. I’ve reviewed enough tools aimed at that customer to know the category is also where the most money gets wasted.
What the spec sheet is actually telling you
The interesting number here isn’t the core count, it’s the gap between the two models. You pay $1,100 more for roughly 20% more GPU cores and two extra CPU cores. That’s a steep curve, and it’s the kind of pricing structure that tells you where the company expects its margin to come from. Nvidia’s own framing of the RTX Spark is that it “delivers amazing creating, AI development, and gaming” in slim laptops and small desktops, which is marketing-speak for “this is a workstation part that fits in a thinner chassis.”
Notice what’s in that list. Creating. AI development. Gaming. Three workloads with genuinely different demands, bundled under one chip’s story. For my money, only one of them clearly justifies the price jump between configs, and it’s the middle one.
Who this is for
- You train, fine-tune, or quantize models locally and you’re tired of paying for cloud GPU time by the hour
- You work with data you legally cannot send to someone else’s inference endpoint
- You do video or 3D work where render time directly costs you billable hours
- You travel and cannot rely on connectivity to reach a remote machine
Who this is not for
- You use AI tools, you don’t build them. API calls to a hosted model don’t care about your GPU
- Your “AI workflow” is a coding assistant and a chat window
- You already own a desktop with a good card and mostly work from one desk
That second list is bigger than the first, and the gap is where sticker shock becomes buyer’s remorse. Local inference is genuinely useful. It is also, for most people doing most knowledge work, a nice-to-have dressed up as a necessity.
The companion device says a lot
Microsoft isn’t shipping the Laptop Ultra alone. There’s also a Surface RTX Spark Dev Box, which The Verge’s Tom Warren got hands-on with at Build earlier in the year. Both run the same Nvidia silicon. A laptop and a small desktop sharing a chip family is a strategy, not a coincidence: Microsoft wants local AI development to happen on Windows hardware rather than drifting to Apple silicon or Linux boxes with discrete cards.
Strategically, that makes sense. Practically, it means the Dev Box probably deserves a hard look from anyone whose honest answer to “do I need this at a coffee shop?” is no. You don’t pay the portability premium for a machine that lives on a desk.
What I can’t tell you yet
I’m not going to pretend to review a laptop I haven’t used. What Microsoft has published is a price, a core count, two colors, and a target customer. What matters for actual AI work is everything they haven’t quantified at the event: sustained performance under thermal load, memory capacity and bandwidth, how the chip behaves on battery versus plugged in, and whether the Windows software stack for local model work has caught up to the hardware.
That last one is the quiet risk. Nvidia hardware is rarely the weak link. Driver maturity, framework support, and tooling are where local AI setups on Windows have historically cost people a weekend. A 6144-core GPU doesn’t help if your pipeline needs a container runtime that fights the OS.
My read
If you’re genuinely GPU-bound today and you know it because you’ve watched a progress bar for hours, preorder the base model and skip the $1,100 upgrade until someone publishes sustained benchmarks. If you’re buying because “AI” is in the pitch and you want to be ready for something, save your money. The tool should follow the work, not the other way around.
I’ll have more once there’s hardware to put through real workloads instead of a product page to read.
đź•’ Published: