Start with PCWorld, because Mark Hachman's review is the most hands-on of the bunch and the most skeptical. His test unit was the 20-core configuration with 32GB of memory at $3,699, and his first impression was not the one Microsoft wanted: the machine arrived effectively defective. Apps crashed or refused to load, Media Player froze twice during battery testing, and 3DMark painted the screen like a dial-up loading bar. Microsoft told him the unit was likely bad and sent a 48GB replacement, which arrived too late to test. The review went ahead on the original machine after a security update settled the instability, which is worth keeping in mind for everything that follows.
The price is the first thing Hachman dwells on, and he does not soften it. At $2,599 to $5,899, he calls the pricing totally unrealistic in the 2026 economy, and the review keeps coming back to the comparison with gaming laptops. That comparison is the review's organizing idea: do you need a monstrously pricey Surface Laptop Ultra when monstrously pricey gaming laptops already exist? He lined up an $1,899 Acer Nitro V with an RTX 5060, a $3,799 MSI Raider with a last-gen RTX 4090, and a $4,299 ASUS ROG Strix with an RTX 5090, and ran the Surface against all three.
On build and display, the Surface mostly holds up. Hachman calls it the most solid laptop he has used, with no keyboard give and a novel magnetic USB-C connector. The Mini-LED display is terrific, but his meter found 792 nits where Microsoft claims 1,000. The webcam is a letdown: an ordinary 1080p unit with no Windows Studio Effects, because as far as he could tell there is no functioning NPU in the system at all. Windows reports one, the spec sheet does not list one, and he never saw it active. For a machine sold partly on AI, that is an odd gap.
Performance is where the review gets interesting. Out of the box, with Microsoft's default efficiency-biased power slider, the machine felt laggy in single-core work. Cranking the slider to Best Performance transformed it, which is now the official advice: if you buy this laptop, change that setting on day one. Under sustained load the machine fades a little, dropping 6.3 percent in a ten-minute Cinebench 2024 loop and failing 3DMark's stress test with a 6.8 percent graphics decline. Battery life landed at 7 hours 57 minutes of 4K streaming and 10 hours 3 minutes in the Procyon office simulation. Respectable, not remarkable.
Then came the surprises. In PugetBench's Photoshop test, the Surface finished last at default settings and first with performance maxed out. Hachman was blunt about what that means for the target buyer. In Premiere it was solid throughout. In Cyberpunk 2077 at 1080p Ultra, the machine was competent, and with DLSS frame generation enabled it jumped to 104 frames per second, 130 at maximum power. For a thin Arm laptop, those are real numbers.
But the power meter is the part of this review people will be quoting. Hachman wired all four laptops up and reran the benchmarks: the Acer held 115 to 117 watts, the MSI blasted through 290 to 305 watts, the ASUS fluctuated between 261 and 290, and the Surface Laptop Ultra drew 133 to 144 watts throughout. That is dramatically better than the gaming laptops, and it is the strongest efficiency argument the platform has. It is also, Hachman notes, a lot of watts for a device Microsoft wants you to think of as a portable AI workstation. He went looking for whether an airplane seat can actually power one of these things churning through offline AI agents with the fans howling, and his research says no, though he could not prove it definitively. Microsoft, he writes, said they would get back to him. That unanswered question is the review's parting shot, and it is a good one.
The broader roundup, collected by TechTimes from ComputerBase, HotHardware, Tom's Guide, Trusted Reviews, Windows Central, and KitGuru, fills in the platform-level picture, and its headline finding is the one buyers should sit with: the GPU is genuinely the most capable integrated graphics ever put in a Windows laptop, and the CPU is built on ARM cores that are already a generation old. The Cortex-X925 and Cortex-A725 inside the N1X are 2024 designs. ComputerBase pointed out, with feeling, that MediaTek, NVIDIA's own CPU partner on this chip, has already shipped ARM's newer C2 architecture in its Dimensity 9600 Pro. Engineering samples of the N1X reportedly date to 2024, which suggests a development timeline that slipped.
In benchmarks the age shows exactly where you would expect. Single-core scores land around 2,600 to 2,710 in Geekbench 7 and 555 to 558 in Cinebench 2026 single-core, competitive with Snapdragon X Elite territory but behind Apple's current generation. Multi-core is a different story: HotHardware found throughput competitive with the Apple M4 Pro and Intel's Core Ultra 9 285K desktop chip, and the GPU numbers are unambiguous. In 3DMark Time Spy Extreme the N1X scored 6,060 against 4,689 for AMD's Strix Halo and 3,226 for Intel's best mobile chip. ComputerBase measured more than double the efficiency of a discrete RTX 5070 laptop GPU setup at around 60W of system power.
On power, the roundup corroborates PCWorld. Peak system draw hit 138W transiently on the Surface, which means the 80W TDP is a sustained rating rather than a ceiling, and first-wave OEMs built in burst headroom. On battery, the Surface gave up about 18 percent of CPU performance in a browser benchmark versus plugged in, which is better than most competitors but a qualifier on the full-performance-on-battery claim. HotHardware measured 14 hours and 35 minutes of video playback on one unit, so the battery story varies by configuration and workload more than usual.
Two platform details from the roundup deserve attention. First, CUDA is confirmed working on production hardware. The pre-production CUDA failures seen on prototypes in July were driver issues and are resolved, and TensorRT, PyTorch's CUDA backend, and llama.cpp all run natively now. That ecosystem depth is the platform's structural advantage over AMD's ROCm-based alternatives. Second, the 55 TOPS NPU in the N1X is reserved exclusively for Microsoft Copilot+ features and is inaccessible to third-party software, a constraint NVIDIA confirmed to ComputerBase. Between that and PCWorld's reviewer being unable to find a working NPU at all, the on-device AI story has a gatekeeper problem: the neural hardware exists to serve Microsoft's features first.
The memory bandwidth caveat also got sharper in wave two. AI inference is memory-bandwidth-bound, and the N1X's 300 GB/s to its LPDDR5X pool is the ceiling on token generation speed. Apple's M4 Max pushes roughly 540 GB/s, so for models that fit on both platforms, Apple generates tokens faster. RTX Spark wins on capacity, models too large for anything Apple ships will run here and only here, and on CUDA. Which constraint matters depends entirely on the buyer's workload, and reviewers are now saying so explicitly instead of letting the memory number speak for itself.
Pricing across the first-wave OEM lineup is filling in. ComputerBase tested the ASUS ProArt P16 with the 20-core N1X, 64GB, and 1TB at 4,299 euros, with a 32GB ProArt P14 starting at 3,299 euros. The US floor remains $2,599 for a 24GB configuration. NVIDIA also confirmed an updated DLSS 5 model arriving today, with a smaller parameter count and 20 percent lower 4K VRAM usage, plus Multi Frame Generation up to 6x, which matters for this platform specifically because generating display frames from rendered ones reduces effective bandwidth demand per frame. It is a software mitigation for the hardware's defining constraint.
Windows on Arm continues its slow maturation. ComputerBase called compatibility "not a catastrophe," which is about as warm as that assessment gets. NVIDIA confirmed day-one native ARM anti-cheat for Counter-Strike 2, Dota 2, Fortnite, Marvel Rivals, Rocket League, and The Finals. The notable holdout is Call of Duty, which arrives for the platform in 2027. Microsoft is shipping general availability of Execution Containers for agent sandboxing and a Get Started experience that auto-configures AI agent platforms on first boot.
So where does wave two leave the story wave one started? StorageReview established the ceiling: 35B models at up to 89 tokens per second, chart-topping multi-core, and two units cutting power under extreme load. Wave two establishes the floor: a machine that needs its power slider fixed on day one, draws 144W doing it, may not run on airplane power, carries a CPU designed in 2024, and keeps its NPU behind Microsoft's velvet rope. The GPU and the memory architecture keep surviving contact with reviewers. Everything around them keeps getting more expensive and more qualified.
For the forum's purposes, the open questions are now concrete. Can anyone actually run one of these at full tilt on a plane? When does the NPU open up to third-party software, if ever? And how long does a 2024 CPU stay competitive at $3,699 and up? Those are the threads worth pulling as shipping units get into real hands starting today.