Surface Laptop Ultra Reviews: Strong Developer Hardware, but Local AI Models Fall Short

What Microsoft says it is​

Microsoft announced pre-orders on October 7, 2026, with availability beginning October 16 per its launch post, at a starting MSRP of $2,599. The Microsoft Store page lists $2,599.99.

The machine is built around NVIDIA RTX Spark. Microsoft says it combines a Blackwell RTX GPU with up to 6,144 cores, a Grace CPU with up to 20 cores, and up to 128 GB of unified memory.

Other manufacturer-stated specifications:

• Memory and OS: Configurations run from 24 GB to 128 GB. The 24 GB model ships with Windows 11 Home. The 32, 48, 64 and 128 GB models ship with Windows 11 Pro.
• Storage: SSDs range from 512 GB to 2 TB, and Microsoft says storage is user-removable.
• Size: Under 18 mm thin and under 4.5 lb, with a 15-inch PixelSense Ultra touchscreen.
• Thermals: Microsoft claims up to 2.5x the thermal capacity of its current Surface Laptops.
• Display brightness: Up to 2,000 nits peak HDR, measured over a 10% window. Microsoft says sustained and SDR brightness are lower.
• Battery: Up to 15 hours of local video playback or 12 hours of active web browsing, from Microsoft's own tests on preproduction units.
• Ports: Three USB-C ports with USB4, USB-A, HDMI, a 3.5 mm headphone jack and an SD card reader.
• Magnetic charging: The included cable attaches magnetically to one USB-C port, which still works for video and data.

One caveat from Microsoft's own footnotes: the total unified memory is shared between CPU and GPU. The maximum the GPU can address is less than the total and depends on configuration and workload.

What the reviewer found​

Weinbach tested the 128 GB model with the 6,144-core GPU. His impressions are qualitative, and he published no benchmark numbers.

• Build: He praises the aluminum body, tight hinge, haptic trackpad, keyboard and side-firing Dolby Atmos speakers. He says it is a bit heavy, but he doesn't mind.
• Cooling: Air comes in at the sides and exits at the back, not out of the bottom onto your lap. The speakers have no grilles because the side intake ports double as the speaker openings.
• Developer work: He used it for about a month to contribute Windows on Arm optimizations to OpenCode and T3 Code. He says it was fast, the fans rarely spun up and the battery life was great.
• Windows on Arm: He says it behaves like any other Windows PC now. Microsoft's own advice is more cautious. Apps not built for Arm run through the Prism emulator, and anyone who relies on specialized drivers, plug-ins or older utilities should check compatibility before buying.
• Display: This is his main complaint. He finds the Mini-LED panel sharp, which suits text, but says its dimming zones are too few and too slow. He sees brightness ghosting, partly from the active window being kept slightly brighter than background windows. He says turning that off or enabling HDR seemed to help, and that OLED might have been better.

Fast models, limited usefulness​

He reports running MAI-Code Flash 1.1 inside GitHub Copilot and calls it very fast. That fits Microsoft's pitch of local models with GitHub Copilot for on-device agentic coding.

His sharper point is that speed is not capability. In his assessment:

• The best US-lab local model for this machine is perhaps Recursive Flint, a 40B-parameter mixture-of-experts model that he says fits a 64 GB configuration.
• Qwen 3.8 Flash Next from Alibaba is slightly smarter but needs the 128 GB model.
• Both handle small, bounded tasks well, such as adding support for a new model to an existing agent.
• Larger agentic work failed in his tests.

Those model rankings and memory requirements are his own judgment. Microsoft's material doesn't validate them. Microsoft says the hardware can run models above 120B parameters. That is a statement about what the hardware can load, not about whether any given model will run well or be useful on a given configuration.

The token-savings argument​

Weinbach argues you should not buy this laptop to save on cloud AI costs. He cites a figure he saw of roughly $1,200 of API-priced tokens generated locally. He compares that with consumer subscriptions of $20 to $200 a month, which he says offer far more usage. He gives no method for these numbers, and I can't confirm them, so treat them as his opinion. His more defensible conclusion is that local compute makes sense for privacy, tinkering, fine-tuning, rendering and other GPU-heavy work. It is not clearly a money saver.

How other reviewers compare​

Other outlets give a broadly consistent but more mixed picture:

• Tom's Hardware says its tests show RTX Spark coming out ahead of Apple's M5 Max and AMD's Ryzen AI Max+ 395 in local AI. It also says gaming is not the platform's primary focus.
• Notebookcheck is harsher on the screen. It calls the Mini-LED panel a problem and lists PWM flickering and no local dimming (high black level) among its cons. It also notes the weight of about 2 kg. This partly echoes Weinbach's display complaint, though the two disagree on the details.
• PCWorld praises the build and audio but flags a very high base price, an average webcam and slow single-core performance, with initial bugginess that went away.
• StorageReview was impressed by the multi-threaded CPU. It also reported that two Surface Laptop Ultra units have now failed identically, which it attributes to power delivery or firmware protection rather than heat. That is a single outlet's report, but anyone buying should watch for follow-up.

Who should consider it​

• Good fit: Developers who work on Windows on Arm, need CUDA-style GPU access, do creative or 3D work, or want to experiment with local models. Microsoft also lists Adobe app optimizations and gaming support as secondary uses.
• Think twice: Anyone who expects local models to replace cloud coding agents, who depends on x64-only drivers or plug-ins, or who cares about display behavior.
• Choose memory carefully: Weinbach says the best model he tried needs the 128 GB version. Microsoft says only part of the unified memory is GPU-addressable. Match the configuration to the specific models you plan to run.

Bottom line​

The reviewers broadly agree the hardware is strong and well built. The open questions are the display, the single-core feel, the price and whether local models will be useful enough to justify the memory. Weinbach recommends it, but his own findings say the sensible use today is as a fast machine for development, creative and GPU work, with cloud models still doing the heavy agentic lifting.