Surface Laptop Ultra Review: Local AI Performance, Arm Compatibility and $2,599 Starting Price

What you're actually buying​

PCMag lists the base model at $2,599.99, but that text comes from the PCMag packet, so here is how the tiers break down. The entry model has an 18-core Grace CPU and a 5,120-core Blackwell GPU, with 24GB of unified memory and a 512GB SSD. The tested unit uses the higher-tier chip, with 20 CPU cores and 6,144 GPU cores, plus 64GB of memory and a 1TB SSD, at $4,299.99.

The top memory option is 128GB, which costs $5,899.99. PCMag points out an odd rule: 64GB can be paired with 2TB of storage, but 128GB tops out at 1TB.

Microsoft says the OS varies by tier. The 24GB models ship with Windows 11 Home, and the 32GB, 48GB, 64GB and 128GB models ship with Windows 11 Pro. That matters for IT buyers who need management and security features.

Availability may have shifted since launch. Tom's Hardware reported that the $5,899 configuration it tested had lead times out to late December. Check current store listings before you plan a purchase.

Why unified memory is the headline​

The N1X combines an Arm CPU and a Blackwell GPU on one chip, and both draw from a single shared memory pool. PCMag compares this to Apple Silicon and to AMD-based machines such as the HP ZBook Ultra G1a 14. It names Nvidia's DGX Spark desktop as the closest relative. Because there's no separate VRAM, larger models and bigger context windows can fit in memory.

Microsoft's own materials make bigger claims. It says the laptop can run models of up to 120 billion parameters locally. Its footnote says that figure rests on theoretical FP4 throughput using sparsity. Treat it as a ceiling, not a promise. Microsoft also says the maximum amount of memory the GPU can address is less than the total system memory.

Windows on Arm: the compatibility question​

This is a Windows 11 on Arm machine. Microsoft says many apps run natively and the rest run through the Prism emulator. It advises checking compatibility for specialized drivers, plug-ins and older utilities before you buy.

PCMag's own testing ran into this. It skipped PCMark 10 because the suite isn't yet Arm-compatible. It also skipped its Photoshop benchmark because of problems on Arm systems. It left out gaming tests too, citing a limited native game catalog and the quirks of x86 emulation. This review therefore says nothing definitive about gaming. Other outlets have started testing games. Tom's Hardware's early verdict was that gaming is not the primary focus of this platform.

PCMag's benchmark results​

Treat these as one lab's results. They depend on the model, settings and software used.

The pattern is consistent: the Surface is competitive with other RTX Spark laptops, but Apple's silicon is still ahead on raw CPU and graphics scores.

Local AI results​

PCMag sorts AI PCs into three classes. At 64GB, this unit lands in Class 2. Class 3 requires at least 96GB of memory, so the review unit couldn't run those tests. The 128GB configuration would qualify, and PCMag didn't test it.

The tests use MLPerf Client 2.0.1:

• Class 1 (Llama 3.1 8B Instruct): time to first token was 0.35 seconds. That's slightly faster than the Asus ProArt P16, which uses the same platform. The HP ZBook Ultra G1a 14 was slowest at 1.52 seconds.
• Class 2 (Phi 4 Reasoning 14B, with 4,000 to 8,000 token prompts): the Surface was 0.15 seconds quicker to first token than the Asus. It generated 31.1 tokens per second against the Asus's 27.6. The MacBook Pro produced its first token in 0.41 seconds and nearly doubled the Surface's generation rate.

Other reviewers have tested different things. StorageReview's headline claim is that the RTX Spark N1X ran a 35B-parameter coding agent at 89 tokens per second. The same site reports that two Surface Laptop Ultra for Business units have failed in the same way, which it attributes to power delivery or firmware protection rather than heat. That is one outlet's finding, and I couldn't verify it independently. Business buyers should still watch for follow-up.

Display, battery and hardware​

Display. The screen is a 15-inch mini-LED touchscreen at 3,270 × 2,180 with a dynamic refresh rate up to 120Hz. PCMag measured 100% of sRGB and 98% of DCI-P3. Peak SDR brightness was above 700 nits, while competitors were around 400. Microsoft claims 2,000 nits of peak HDR brightness. PCMag didn't measure that, but it calls the claim plausible. Microsoft's footnote says the figure is measured over a 10 percent window with HDR content. Notebookcheck's headline, by contrast, says the 2000-nits Mini-LED panel is a problem. The details are behind that review, so I can't say why. Pen input isn't supported.

Battery. PCMag's local-video test ran 15 hours 40 minutes. Several other RTX Spark laptops lasted longer, and the MacBook Pro ran more than 27 hours. Microsoft's own claim is up to 15 hours of video and up to 12 hours of web browsing, from preproduction testing at 150 nits. Real-world use differs. The Shortcut reported eight- and 10-hour days before needing a charge.

Ports and design.

• The chassis is CNC-milled aluminum and weighs about 4.5 pounds in PCMag's description. Microsoft's own listing says under 4.5 pounds.
• There are three USB-C ports with USB4, one USB-A port, full-size HDMI, a UHS-II SD card slot and a 3.5mm jack. Only one USB-C port has the Magnetic Connect charging feature.
• Wi-Fi 7 and Bluetooth 5.4 are included.
• The haptic glass touchpad is the largest on a Surface. PCMag praises the keyboard too.
• The webcam is 1080p with IR for Windows Hello. PCMag wanted a higher-resolution sensor at this price.
• PCMag says the system can drive up to three external 4K displays at 60Hz. I couldn't confirm that count from Microsoft's own materials.

Marketing claims versus tested results​

Microsoft's October 7 Windows blog claims big gains over an M5 Pro MacBook Pro. It cites up to 2.1× faster time to first token, with higher figures for image and video generation. Its footnotes say these came from Microsoft-commissioned or Nvidia testing in September 2026 on preproduction systems. Each claim used a different model and setup. PCMag compared against the M5 Max, and Apple led in PCMag's tests. Don't mix the two sets of numbers.

Who should consider it​

It fits developers, AI builders and creators who want CUDA and a large shared memory pool, and who have confirmed their tools run on Arm. Microsoft's own advice is that mainstream users may be better served by the regular Surface Laptop.

It doesn't fit buyers who mainly want a gaming laptop, or who depend on legacy drivers and x86-only tools. Tom's Hardware's verdict is that a Windows power user won't get a ton out of RTX Spark besides better software and gaming compatibility than earlier Arm laptops.

Bottom line​

The Surface Laptop Ultra is the first full review of a new Windows hardware category. It has a standout screen and keyboard, a sensible magnetic charging idea, a decent port selection and credible local-AI results. It's also expensive, it trails Apple on measured performance and Windows-on-Arm compatibility is a real risk. The 64GB review unit sits in the middle of the range, and the 128GB model is the one that would test the platform's biggest claims. If you're buying for work, test your critical apps first. If you're buying for bragging rights, you'll pay a lot for them.