The Microsoft Surface Laptop Ultra can run local AI workflows offline, and PCMag’s 64GB review unit delivered 31.1 tokens per second in a Class 2 test. But the battery figure comes from a separate video-playback test, while a reported power-loss issue involved different business review units under a specific sustained workload. The laptop makes the most sense if you’ll use its large shared memory for demanding AI or creative work, not just everyday apps.
Choose a memory tier for the models and workloads you expect to run: extra unified memory can determine what fits, while a benchmark on one configuration does not predict every task.
Configurations, memory, and US prices
The Surface Laptop Ultra pairs an Arm-based Grace CPU with Blackwell graphics in NVIDIA’s RTX Spark N1X system-on-chip. Its unified memory is shared by the processor and graphics, so the amount available matters when loading larger AI models.
| Configuration | CPU and graphics cores | Unified memory | SSD |
| Base | 18 CPU cores, 5,120 graphics cores | 24GB | 512GB |
| PCMag review unit | 20 CPU cores, 6,144 graphics cores | 64GB | 1TB |
| Highest-memory option | 20 CPU cores, 6,144 graphics cores | 128GB | 1TB |
In its review published November 4, 2026, PCMag listed the US base configuration at $2,599.99, the 64GB/1TB configuration at $4,299.99, and the 128GB/1TB option at $5,899.99. Those prices correspond to the configurations in that review.
PCMag’s Class 3 AI tests require at least 96GB of memory. Its 64GB review unit therefore did not qualify; the 128GB configuration meets that memory threshold, though PCMag did not run its Class 3 tests on that configuration. For a closer look at the choice between the 32GB and 128GB models, see our earlier configuration comparison.
What offline AI work looks like
Kevin Stratvert demonstrated document chat in LM Studio with Wi-Fi disabled, then used ComfyUI for local image and video generation on his 64GB review unit. Microsoft supplied that device for his review, and Stratvert disclosed that he could keep it.
His five-second image-to-video task took 2 minutes 31 seconds while plugged in and 2 minutes 34 seconds on battery. The displayed charge went from 99% to 96%. That is a useful snapshot of one creative task on one setup, not a measure of how long the laptop lasts through a workday.
Battery results and sustained-load behavior
PCMag measured 15 hours and 40 minutes in a local 720p video-playback test. The review unit started fully charged; the test used 50% screen brightness, 100% volume, Wi-Fi off, and keyboard backlighting off. That result describes the stated playback conditions, not a mixed workload with AI generation or other demanding apps.
StorageReview reported a different kind of result on two Surface Laptop Ultra for Business review units. Both lost power during an all-core OpenBLAS LU N=20000 workload; on the second unit, three of four attempts failed. StorageReview’s logs indicated a hard power cut, and the outlet did not identify the cause. This is a specific observation under sustained processor load, not a measure of how often retail units fail in normal use.
The benchmarks also answer different questions. PCMag’s Class 2 test used Phi 4 Reasoning 14B: the Surface scored 31.1 tokens per second, compared with 27.6 on the Asus ProArt P16. The MacBook Pro 16-inch with M5 Max posted nearly twice the Surface’s rate in that comparison. Those results apply to that benchmark and its tested systems, not to every local model or creative app.
Display and connections
The 15-inch mini-LED touchscreen has a 3,270 × 2,180 resolution, a 3:2 aspect ratio, and a dynamic refresh rate up to 120Hz. PCMag measured 100% sRGB coverage, 98% DCI-P3 coverage, and peak SDR brightness above 700 nits. Microsoft specifies peak HDR brightness of 2,000 nits, a figure for a different display mode. Finger touch works, but Surface Pen and Slim Pen input are not supported.
All three USB-C ports support USB4, 40Gbps, DisplayPort 2.1, and up to 140W power delivery. Magnetic Connect, the magnetic charging connection, is available on one of those ports. The laptop also has USB-A, HDMI, a full-size UHS-II SD card slot, and a 3.5mm audio jack.
Who should consider it
The strongest case is for someone who wants to run substantial local AI or creative workflows away from a network connection and will benefit from higher unified-memory capacity. Stratvert’s demonstration shows document chat and image-to-video generation running offline; PCMag’s Class 2 result gives a separate, standardized performance datapoint. Neither tells you how every model or application will behave.
For routine email, office work, or services that already run AI online, the high US-listed prices make the hardware harder to justify. If a workload depends on fitting a larger model in memory, compare that requirement with the configuration before choosing. And if your work includes sustained all-core processing, StorageReview’s specific power-loss report is worth weighing alongside the more typical benchmark results.
