Surface Laptop Ultra Review: RTX Spark Local AI Benchmarks, Battery Performance and 64GB vs 128GB

Microsoft and Nvidia have spent the better part of 2026 plotting their counteroffensive, and on October 16, that counterpunch officially began shipping. Enter the Surface Laptop Ultra, a Windows on Arm powerhouse engineered from the silicon up around Nvidia’s N1X-based RTX Spark Superchip. Combining an Nvidia Grace CPU (up to 20 cores), a Blackwell RTX GPU (up to 6,144 CUDA cores), and up to 128GB of unified LPDDR5X memory, Microsoft is pitching this machine directly at creatives, software engineers, and AI builders who are tired of paying cloud API tolls for every single token.

Marketing slides and theoretical petaflops are one thing; running actual local coding agents in a hotel room or on an airplane is quite another. A week-long, hands-on evaluation published on October 16 by Frederic Lardinois at The New Stack provides one of the clearest looks yet at how the Surface Laptop Ultra behaves when you throw real-world local AI models at it—and where the marketing promises collide with the laws of physics.

Under the Hood: The RTX Spark Platform and Pricing Tiers​

To understand what makes the Surface Laptop Ultra such a radical departure from previous Surface devices—including the ill-fated or thermally constrained Surface Laptop Studio series—you have to look at the silicon architecture.

Instead of pairing an x86 processor from Intel or AMD with a separate Nvidia mobile GPU over a PCIe bus, the Surface Laptop Ultra uses Nvidia’s RTX Spark platform. Like Apple’s M-series chips or AMD’s Strix Halo, RTX Spark uses a unified memory architecture. The CPU and the Blackwell GPU share a single pool of system RAM dynamically. According to Microsoft’s official product documentation, while the system can be configured with up to 128GB of total unified memory, the maximum amount directly addressable by the GPU is somewhat less than the total system RAM and varies depending on system configuration and workload.

Here is how the hardware and pricing break down based on Microsoft's official specifications and The New Stack's testing:

Microsoft’s Devices Blog notes that pre-orders opened on October 7, 2026, alongside a companion desktop unit—the $5,999 Surface RTX Spark Dev Box (slated to ship in November)—and an aggressive trade-in promotion offering up to $1,000 cash back for buyers trading in an eligible MacBook Pro.

But does a $4,300 Windows on Arm laptop actually dethrone a desktop Mac Studio or high-end MacBook Pro for local AI work?

Local AI Benchmarks: Why Model Architecture Matters More Than Petaflops​

Microsoft’s headline marketing claim for the Surface Laptop Ultra is that it delivers up to 1 petaflop of AI performance and can run models exceeding 120 billion parameters locally. However, as Microsoft’s own footnotes clarify, that 1-petaflop figure represents theoretical FP4 (4-bit floating point) performance utilizing hardware sparsity. In day-to-day local inference, raw compute is rarely your primary bottleneck—memory bandwidth and model architecture are what determine whether your coding assistant feels snappy or sluggish.

In his testing for The New Stack, Lardinois loaded up his $4,300 review unit (64GB RAM / 1TB SSD) with standard local AI developer tooling, including Unsloth Studio, Ollama, and Hermes Agent, while also evaluating workloads suitable for OpenCode. His baseline for comparison was his daily driver at home: an Apple Mac Studio equipped with an M4 Max chip and 48GB of unified memory.

While plugged into the included 140W power adapter, the Surface Laptop Ultra delivered revealing results across four distinct model configurations:

• Nemotron 3.5 Bolt (Heavily Quantized, 9B Active Parameters): Tested shortly after the model's launch, this lightweight configuration screamed along at 56 to 60 tokens per second (tok/s), occasionally peaking even higher. At nearly 60 tok/s, local inference feels virtually instantaneous, making it well-suited for agentic coding workflows in Hermes Agent or OpenCode.
• Qwen 3.8 27B (Dense Model, 8-bit Quantization): When shifting to a traditional dense model where every parameter is activated for every token generated, the hardware hit a wall. Running in Unsloth Studio at 8-bit quantization, Qwen 3.8 27B maxed out at just 8 to 9 tokens per second. Lardinois noted that this pace is barely usable for interactive work and pointed to memory bandwidth as the likely culprit—a classic bottleneck for dense LLMs on mobile unified memory buses.
• Qwen 3.8 27B (Dense Model, 4-bit Quantization): Dropping that exact same 27-billion-parameter dense model down to 4-bit quantization halved the memory footprint and memory bandwidth pressure per token. Throughput jumped roughly threefold to a very respectable 25 to 28 tokens per second.
• Gemma 4 26B-A4B (Mixture-of-Experts, 8-bit Quantization): If there is one golden takeaway for Windows AI developers, it is this test. By using a Mixture-of-Experts (MoE) architecture—which holds 26B total parameters in memory but only activates around 4B parameters per token—the Surface Laptop Ultra achieved roughly 40 tokens per second even at 8-bit quantization.

The MoE Advantage and Cross-Platform Context​

What do these numbers tell us? From an engineering standpoint, unified memory on a laptop gives you the capacity to hold large models, but it does not give you the multi-terabyte-per-second memory bandwidth of a desktop RTX 5090 or data-center H100.

Consequently, dense models at higher precision (like 8-bit or FP16) will choke waiting for weights to travel from RAM to the compute cores. Mixture-of-Experts (MoE) models bypass this bottleneck brilliantly: the 64GB or 128GB unified memory pool holds the massive total parameter count, while the memory bus only has to shuttle a fraction of those weights (such as 4B or 9B active parameters) for each token. It is no coincidence that Microsoft’s own recently highlighted on-device coding model, MAI Code 1.1 Flash, uses a 137B-parameter MoE architecture with just 6.8B active parameters compressed to roughly 3.3 bits per weight (about 53GB total).

Subjectively, Lardinois found that the Blackwell GPU inside the Surface Laptop Ultra felt uncompromising and was generally a bit faster than his M4 Max Mac Studio across the tools he tested. While he did not have an M5-based Mac on hand for side-by-side testing, separate testing by Jeffrey Kampman at Tom’s Hardware on a 128GB Surface Laptop Ultra corroborated the platform's strength in local AI. What is the takeaway for developers? If you pair CUDA acceleration on Blackwell with 4-bit dense models or 8-bit MoE models, the Surface Laptop Ultra is a legitimate mobile workstation.

Unplugged Reality Check: Battery Throttling, Heat, and Fan Noise​

Every laptop vendor loves to claim that their latest mobile workstation defies thermodynamics when you yank the power cord out of the wall. According to The New Stack, Microsoft claimed its internal benchmarks showed the Surface Laptop Ultra retaining 99.6% of its performance on battery power.

If you are checking email or running a brief burst test, maybe that holds up. If you are sustaining heavy local LLM inference on the GPU, however, The New Stack’s hands-on testing tells a very different story:

• Qwen 3.8 27B (4-bit) on Battery: Dropped from 25–28 tok/s (plugged in) down to roughly 18 tokens per second on battery.
• Nemotron 3.5 Bolt on Battery: Dropped from 56–60+ tok/s (plugged in) down to 42 to 45 tokens per second on battery.

Across both AI inference workloads and gaming, Lardinois observed a consistent 20% to 25% performance penalty when running unplugged. Now, let’s be fair: 18 tok/s on a 27B model or 45 tok/s on a fast coding agent while sitting in an airport terminal without an outlet is still completely usable. You can get real work done on battery. However, buyers should expect a quarter of their throughput to evaporate away from the 140W brick, regardless of what the marketing slides suggest.

Endurance and Acoustics​

What about battery life and fan noise? Microsoft rates the Surface Laptop Ultra for up to 12 hours of active web usage (tested at 150 nits screen brightness across open browser tabs on preproduction 64GB/1TB units). While Lardinois did not conduct formal battery rundown benchmarks, he noted that based on his observations during travel, consistently hammering the Blackwell GPU with local AI workloads will likely drain the battery in just a few hours.

Then there is the heat and noise. Microsoft engineered a new thermal system for the Ultra that boasts up to 2.5 times the thermal capacity of the Surface Laptop 15-inch (8th Edition). During normal day-to-day office and web tasks, the dual fans stay completely off, and the laptop remains silent. Even after a heavy load finishes, the system spins down and quiets rapidly.

When an agentic tool like Hermes Agent is grinding through a multi-step coding query, however, those dual fans spin up noticeably. The New Stack describes the acoustic profile under sustained AI load as not extreme, but loud enough to "get a bit on your nerves while you are working". There is a reason Microsoft's desktop Surface RTX Spark Dev Box is covered in air vents: dissipating sustained Blackwell GPU heat in a sub-18mm laptop chassis requires moving serious air.

The 64GB Dilemma: Sizing Your Unified Memory​

One of the sharpest practical insights from The New Stack’s review revolves around memory configuration pricing.

On paper, 64GB of unified memory sounds like a sweet spot. In practice, Lardinois points out that 64GB puts AI developers in a slightly awkward middle ground right now. Most popular mid-sized open-weights models cluster around the 27B to 30B parameter mark (which easily fit into 24GB to 32GB when quantized), whereas the next major tier jumps up to 70B+ or 120B+ models. There are relatively few mainstream models specifically targeting the 32GB to 48GB footprint.

Of course, having 64GB of RAM is far from wasted:

1. It leaves massive headroom for a large KV cache (allowing much longer context windows without crashing or swapping to the SSD).
2. It allows you to keep Docker containers, VS Code, a browser with fifty tabs, and multiple smaller specialized models loaded into memory simultaneously.
3. As seen with Microsoft's 53GB quantized MAI Code 1.1 Flash model, vendors are actively beginning to target this exact memory envelope.

Still, if your goal is to run massive 70B-to-120B parameter models with generous context windows alongside your IDE, you will want the 128GB configuration—and jumping from 64GB ($4,300) to 128GB pushes the price to a wallet-searing $5,900. At that price point, you have to ask yourself whether you truly need 128GB in your backpack, or if you would be better served by a desktop workstation or the $5,999 Surface RTX Spark Dev Box.

Living With the Surface Laptop Ultra: Hardware, Ports, and Windows on Arm Gaming​

Beyond the terminal window and token counters, how does the Surface Laptop Ultra hold up as an everyday flagship laptop? Microsoft clearly listened to years of complaints from power users who hated carrying dongle octopuses.

Chassis, Display, and I/O​

• Weight and Dimensions: At 4.41 lbs (2 kg) and under 18mm thin, this is a dense, substantial 15-inch machine. If you are coming from an ultraportable like a MacBook Air, you will feel it in your shoulder bag.
• The Display: The 15-inch, 3:2 PixelSense Ultra mini-LED touchscreen features a 3270 × 2180 resolution (262 PPI, 120Hz) and hits a searing 2,000 nits of peak HDR brightness. Lardinois praised its brightness and visual pop, though he noted that the glass is highly reflective, which can be distracting in bright environments. Note also that per Microsoft's specifications, this touchscreen is optimized for touch only and does not support Surface Pen input.
• Magnetic USB-C and Port Selection: Microsoft debuted Magnetic Connect, which the company calls the world's first built-in magnetic USB-C charging port on a laptop. The 140W cable snaps in magnetically like an old-school Surface Connect or MagSafe charger, yet the port itself still functions as a full-featured USB-C port for data and video. Alongside it, Microsoft notes the system includes two additional USB-C ports, a legacy USB-A port, a full-sized HDMI port, an SD card reader, and a 3.5mm headphone jack.
• Haptics, Audio, and Repairability: The haptic touchpad is 30% larger than the one on the Surface Laptop 8th Edition and supports tactile timeline scrubbing in apps like DaVinci Resolve. Front-facing side speaker vents project a remarkably wide stereo soundstage. Best of all for IT admins and tinkerers, the SSD is user-replaceable via four screws on the bottom panel without voiding your warranty (provided you follow Microsoft's Service Guide and don't damage internals during the swap).

Windows on Arm Maturity and Gaming​

A year or two ago, recommending a Windows on Arm laptop to a developer came with a laundry list of caveats. Today, the ecosystem has matured rapidly. Between native ARM64 builds for tools like VS Code, Python/Anaconda, Git, and OpenCode, plus full Nvidia CUDA support wired directly into the RTX Spark platform, local AI workflows run natively without feeling like second-class citizens.

And what happens when you close the code editor and launch Steam? While Windows on Arm gaming still relies on Microsoft’s x86-to-Arm emulation layer for many titles—and CPU-heavy games can show emulation overhead—the raw horsepower of the 6,144-core Blackwell GPU carries GPU-bound titles surprisingly well. In The New Stack’s testing, Cyberpunk 2077 ran at roughly 60 FPS at 2K resolution with ray tracing enabled (and could even stretch to 4K with DLSS frame generation and settings tweaks), while the remastered The Witcher 3 held a smooth 60 FPS at 4K with ray tracing turned off. Given the native 3270×2180 screen resolution, sticking to 2K/1440p-class resolutions on the built-in display makes the most sense anyway.

Practical Guidance: Should You Buy One for Local AI?​

If you are an IT decision-maker or independent developer weighing the Surface Laptop Ultra, keep these practical takeaways in mind:

1. Match Your Models to the Bus: Do not buy this machine expecting to run unquantized or 8-bit dense 30B+ models at blazing speeds. Build your local stack around Mixture-of-Experts (MoE) models (like Gemma 4 26B-A4B or MAI Code 1.1 Flash) or 4-bit quantized dense models, where the RTX Spark chip comfortably delivers 25 to 60+ tokens per second.
2. Plan for Wall Power on Heavy Jobs: Expect a 20% to 25% drop in token generation speed and only a few hours of runtime if you run sustained GPU inference on battery. Pack the 140W Magnetic Connect brick for serious road warrior sessions.
3. Evaluate the 64GB vs. 128GB Price Jump Carefully: At $4,300, the 64GB model easily handles 27B–30B quantized models and MoE coding agents with plenty of room for context and apps. Only pay the $5,900 premium for the 128GB tier if your specific workflow mandates keeping 70B to 120B+ parameter models resident in memory on the go.

Summary​

With the Surface Laptop Ultra, Microsoft and Nvidia have finally delivered a legitimate Windows-based rival to Apple’s MacBook Pro and Mac Studio for local AI development. It isn't magic—dense models still run into memory bandwidth realities, sustained agent runs will make the fans audible, and unplugging from the wall trims a quarter off your AI throughput. But by combining a 6,144-core Blackwell GPU, unified memory, native CUDA support, and a thoughtfully equipped chassis with replaceable storage and real ports, the Surface Laptop Ultra proves that Windows on Arm is ready for serious local AI builders.