Microsoft’s Surface Laptop Ultra started shipping on October 16, and the first wave of in-depth reviews has landed. This is the first laptop built around Nvidia’s RTX Spark platform — an Arm chip with a full Blackwell GPU, up to 128GB of unified memory, and a promise to run serious AI models locally instead of in the cloud. Prices run from $2,599.99 to $5,899.99.
Reviewers have now had it in the lab for days, and the picture they paint is genuinely exciting — and genuinely alarming in places. Here is what the reviews actually found, and who this machine is really for.
What the Surface Laptop Ultra actually is
The heart of this laptop is the Nvidia RTX Spark N1X, which is the consumer version of the GB10 chip that’s been sitting inside Nvidia’s DGX Spark mini-PCs for about a year. It’s a system-on-a-chip: an Nvidia Grace Arm CPU (co-designed with MediaTek) fused with a Blackwell RTX GPU, sharing one pool of LPDDR5X unified memory that the CPU and GPU both draw from.
Two configurations of the chip exist. The entry version pairs an 18-core CPU (9 performance + 9 efficiency cores) with a 5,120-core Blackwell GPU and up to 64GB of memory. The full-fat version steps up to a 20-core CPU and a 6,144-core GPU with up to 128GB of unified memory. The rest of the package: a 15-inch mini-LED touchscreen (3270×2180, 120Hz, 2,000-nit peak HDR), Wi-Fi 7 (with Wi-Fi 8 routers already shipping), a 4.5-pound aluminum chassis, generous ports including HDMI and a full-size SD slot, and user-removable storage.
That 128GB figure is the headline because no other Windows laptop offers it: a typical gaming laptop’s GPU tops out around 24GB of VRAM, and an AI model has to fit entirely in that. On the Ultra, a large model can draw from the entire shared pool — Microsoft says the platform can run models with more than 120 billion parameters locally.
One caveat: Microsoft’s “up to 1 petaflop of AI compute” is theoretical FP4 with 2:4 sparsity — a mode almost no real workload uses. A fairer figure is ~500 teraflops. Still a lot of laptop, but keep the petaflop in frame.
What reviewers praise: genuine AI muscle
StorageReview’s deep lab evaluation (Oct 17) — testing a 128GB configuration against a Razer Blade 16, a Dell Pro Precision 5 16S, and others — is the most revealing test so far:
- Local AI speed is the real deal. Running gpt-oss:20b and Qwen3.6-35B-A3B locally, the Ultra generated at 58 to 89 tokens per second with up to 96GB of memory allocated to the GPU. For context, the Blade 16’s RTX 5090 was 2.3x faster — but only on models small enough to fit in its 24GB of VRAM. Anything bigger simply doesn’t run on the Razer.
- The fastest multi-core CPU in its test group. It scored 6,963 in Cinebench 2026 multi-core — 34 percent ahead of the Blade 16 and 47 percent ahead of the Dell Precision. An LLVM compile finished in 206 seconds versus 327 for the next-fastest machine.
- Surprising wins outside AI. One review found the Surface the fastest in 4K-to-1080p Handbrake transcoding, beating a MacBook Pro M5 Max by a second. Gaming, long the Achilles’ heel of Windows on Arm, took a real step forward: God of War: Ragnarok ran at 1440p with ultra settings and DLSS at 41–47 fps.
That last point matters beyond gaming: CUDA is the lingua franca of AI development, so if you’re exploring custom AI models and want to experiment locally without paying cloud GPU bills for every iteration, this is the first Windows laptop that credibly lets you.
What reviewers are flagging: two dead units and real trade-offs
Now the part Microsoft didn’t put in the launch materials. StorageReview’s most urgent finding: two separate Surface Laptop Ultra review units died in the lab under the same sustained workload. During an OpenBLAS benchmark — an all-core double-precision floating-point stress test at problem size N=20000 — the first unit hung, developed a display defect near the exhaust edge alongside a reproducible CPU fault, and had to be replaced. The replacement then suffered three sudden, hard power-offs in a single evening on October 14.
What makes this unsettling: the same chip had already sat at 97–98°C through a 31-minute Cinebench loop and Blender without losing performance — not an ordinary thermal shutdown. At each power cut the CPU was at a merely warm 95°C, and Windows logged Kernel-Power Event ID 41 with no crash dump, no hardware error record, and no thermal trip event. StorageReview’s conclusion: the evidence points at power delivery or firmware protection, not heat.
Microsoft hasn’t publicly explained the failures as of this writing, and an extreme FP64 stress test isn’t a normal workload. But two samples in a row dying under sustained all-core load is a red flag for a machine whose entire pitch is sustained heavy compute.
The caveats keep coming from there:
- Weak single-thread, high memory latency. That same multi-core champion has the worst memory latency in StorageReview’s group (251 ns) and memory latency more than double the fastest x64 competitor. Single-threaded work — the stuff of everyday responsiveness — trails Intel and Apple parts by wide margins (Cinebench R23 single-core: ~1,317 vs ~2,164 for an Intel Core Ultra 9 285H in Notebookcheck’s testing).
- Battery life under stress is poor. Linus Tech Tips’ stress testing found the Surface lasted just 53 minutes under load, versus about 1.5 hours for a MacBook Pro 14 and over 2 hours for a Snapdragon-based Asus Zenbook with a smaller battery. You’ll be tethered to the wall for heavy AI work.
- The 32GB sweet-spot config has limits. The Shortcut’s testing of the $3,699.99 configuration found only 12–16GB of memory could actually be allocated to graphics; most games weren’t playable at the display’s full resolution even with DLSS, and running models through Microsoft’s Prism emulator came with extended load times and sporadic frame drops. Running a Qwen3.8-27B model with the GPU in AI mode sent fans to full speed and drained the battery fast.
- Windows on Arm caps the storage. The Gen5 SSD reads 14.1 GB/s sequentially, but random 4K I/O through the Arm storage and software-encryption stack tops out near 300K IOPS — roughly a fifth of the raw drive’s capability.
- No Copilot+ features at launch. Windows Central notes the laptop currently lacks all Copilot+ PC features, despite being Microsoft’s flagship “AI PC.” The newest Windows-exclusive AI functions simply aren’t there yet.
- The display is divisive. The display divides opinion: the 2,000-nit mini-LED panel earns praise for brightness and color, but Notebookcheck flags reflections and blooming as real problems for professional color work — with no stylus support.
The price ladder, honestly assessed
Microsoft’s US configurator lists eight configurations:
| Chip | Memory | Storage | Price |
|---|---|---|---|
| 18-core / 5,120-core GPU | 24GB | 512GB | $2,599.99 |
| 18-core | 24GB | 1TB | $2,999.99 |
| 18-core | 32GB | 1TB | $3,299.99 |
| 20-core / 6,144-core GPU | 32GB | 1TB | $3,699.99 |
| 20-core | 48GB | 1TB | $3,999.99 |
| 20-core | 64GB | 1TB | $4,299.99 |
| 20-core | 64GB | 2TB | $4,699.99 |
| 20-core | 128GB | 1TB | $5,899.99 |
The entry $2,599.99 model is the weak value pick: 24GB of unified memory disappears fast once Windows takes its share, and Microsoft hasn’t said whether the cut cores come from the performance or efficiency clusters. The $3,699.99 configuration — the one CNET, PCWorld, and The Shortcut tested — is the sensible middle of a very tall ladder.
The interesting pricing story is at the top: the 128GB configuration at $5,899.99 is about $800 cheaper than a 14-inch MacBook Pro with the same memory. I didn’t expect Microsoft to beat Apple at the top end, but here we are — for buyers whose models genuinely need that memory pool, the math works. There’s also a $5,999.99 RTX Spark Dev Box shipping in November for developers who don’t need portability, and a trade-in promotion (up to $1,000 cash back with a qualifying MacBook Pro trade-in, running through November 23 in the US and Canada).
Buying for a team of AI engineers? The same bulk procurement discipline that applies to servers applies here — pilot first-generation hardware before standardizing it.
Who should buy it — and who shouldn’t
Buy it if: you’re a developer or researcher whose daily work is local AI models above 24GB — the 35B-to-70B range that no other laptop can touch. You get the fastest agentic-AI test scores reviewers measured, genuine multi-core workstation speed, and a memory pool that no competing portable machine offers. You’re buying into a first-generation platform, and you should know that.
Skip it if: gaming is the priority — a dedicated gaming laptop costs less and performs better, with none of the emulation rough edges. Skip it if you want maximum battery life and polish — the MacBook Pro still wins there, and it’s not close under load. Skip it if your models fit in 24GB — you don’t need this machine.
Wait if: you can afford to. Dell, HP, Asus, Lenovo, and MSI all ship RTX Spark laptops too, and watching how they price and tune the exact same chip — plus whether Microsoft addresses the power-off reports — is the smart play. The Asus ProArt P16 with the same N1X already exists, and competition should sharpen both pricing and firmware.
Microsoft also claims the Ultra beats a MacBook Pro with M5 Pro by 6.2x in video generation and 4.3x in image creation — vendor benchmarks on vendor-chosen workloads. Independent reviewers confirm the direction, not the ratios.
The bottom line
The Surface Laptop Ultra is the most interesting Windows laptop in years — and the least finished flagship Microsoft has shipped in a while. Nvidia’s RTX Spark genuinely creates a new category: a portable machine that runs serious AI models locally. For how AI is reshaping everyday computing, that’s a meaningful milestone — the hardware to match the ambition.
But first-generation hardware earns its skepticism. Two dead review units, weak single-thread performance, poor stressed battery life, missing Copilot+ features, and a $5,900 ceiling add up to a machine that is brilliant at one thing and compromised at several others. If your work lives in that one thing — big local models — nothing else portable does it. For everyone else, the wise move is to watch the firmware updates and the OEM competition, and decide in the spring.
Sources and review data: StorageReview’s lab review (Oct 17), Notebookcheck review and battery analysis (Oct 19), The Shortcut’s 32GB review (Oct 17), Windows Central, Gizmodo (Oct 18), Gadget Hacks preview, m1k.tech announcement analysis, Microsoft’s published specifications and US configurator pricing. Microsoft’s performance claims are labeled as vendor claims where cited; review findings are attributed to the outlet that reported them.
