Microsoft has started shipping a Surface laptop unlike anything it has sold before. The Surface Laptop Ultra starts at $2,599.99 and swaps the usual Qualcomm or Intel setup for an NVIDIA “superchip” platform built around an Arm-based Grace CPU and a Blackwell RTX GPU. Microsoft calls it the Surface Laptop Ultra; outlets covering the October 7, 2026 preorder opening, including Fortune, have framed it as the company’s answer to a question more buyers are asking in 2026: can a laptop actually run a large language model on its own, without sending a single token to the cloud.
The timing is not an accident. Preorders opened October 7, 2026, and general availability followed on October 16, 2026, right as DRAM and NAND shortages are pushing PC prices up across the industry. Microsoft is betting that a laptop with up to 128GB of unified memory and a Blackwell GPU rated at up to one petaflop of FP4 compute can justify a price tag that puts it well above a typical Copilot+ PC, while undercutting the cost of a dedicated AI workstation.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the Surface Laptop Ultra actually is
The Surface Laptop Ultra is built on NVIDIA’s RTX Spark N1X platform, the same superchip family NVIDIA has been positioning as a bridge between consumer laptops and data-center-grade AI hardware. Instead of a conventional x86 or Snapdragon processor, the machine pairs an Arm-based NVIDIA Grace CPU, available in 18-core or 20-core configurations, with a Blackwell GPU carrying either 5,120 or 6,144 CUDA cores depending on configuration. Microsoft and coverage of the launch describe the AI performance ceiling as up to one petaflop using FP4 precision, a figure pulled from GPU throughput rather than a conventional NPU TOPS rating.
Memory is the headline spec for anyone interested in local AI. The base $2,599 configuration ships with 24GB of unified memory shared between the CPU and GPU, paired with an 18-core Grace CPU and 512GB of storage. Step up to the mid-tier configuration and the price reportedly jumps to $3,999.99 with 48GB of unified memory and a 1TB SSD, while the top-tier build is listed on the Microsoft Store at $5,899.99 with a full 128GB of unified memory and a 1TB SSD, the same configuration Tom’s Hardware tested. That top configuration is what Microsoft points to when it describes the laptop’s ability to keep models as large as 120 billion parameters resident in memory for local inference.
The chassis itself borrows from the existing Surface Laptop Ultra specs that leaked ahead of launch, including a 15-inch PixelSense Ultra mini-LED touchscreen running at 120Hz with up to 2,000 nits of peak HDR brightness and a 3:2 aspect ratio. Ports include three USB-C connections, one USB-A port, HDMI, a full-size SD card reader, and a 3.5mm headphone jack. Microsoft says the laptop, which comes in Platinum and Nightfall finishes, weighs under 4.5 pounds and measures under 18mm thick, with a redesigned two-fan cooling system the company claims delivers up to 2.5 times the thermal capacity of the 15-inch Surface Laptop 7th Edition.
Surface Laptop Ultra price and configuration breakdown
Pricing scales steeply with memory, which makes sense given how dependent local LLM performance is on how much of a model can sit in unified memory at once. The table below lays out the three configurations as reported around launch.
| Configuration | CPU | GPU | Unified Memory | Storage | Starting Price |
|---|---|---|---|---|---|
| Base | 18-core Grace CPU | Blackwell, 5,120 CUDA cores | 24GB | 512GB SSD | $2,599.99 |
| Mid-tier | 20-core Grace CPU | Blackwell, 6,144 CUDA cores | 48GB | 1TB SSD | $3,999.99 |
| Top-tier | 20-core Grace CPU | Blackwell, 6,144 CUDA cores | 128GB | 1TB SSD | $5,899 |
That spread means the entry-level Surface Laptop Ultra is priced closer to a high-end creator laptop than a premium AI workstation, while the 128GB configuration lands in territory shared with small-batch AI dev boxes like Microsoft’s own Surface Dev Box, which starts at $5,999 for a desktop form factor built on the same RTX Spark silicon. It is worth noting that 24GB of unified memory, the base configuration’s allotment, is a meaningfully different proposition than 128GB when it comes to which local models actually fit and run at usable speed.
Running large language models locally: what’s actually confirmed
Microsoft’s pitch centers on local inference without a cloud round-trip. The company’s Windows AI documentation points developers toward Foundry Local to run open-source large language models on-device, and Microsoft describes the Surface Laptop Ultra as its “most powerful Surface laptop ever, built for local AI and 3D graphics.” The strongest specific capability claim tied to the launch is that the 128GB configuration can keep models as large as 120 billion parameters loaded for local inference, a figure made possible by the laptop’s unified memory architecture pooling CPU and GPU access to the same pool of RAM.
What is not confirmed is just as important for anyone deciding whether to buy one. There is no verified, standardized benchmark list showing tokens-per-second results for named open models such as Llama, Mistral, Qwen, or DeepSeek running on the Surface Laptop Ultra. No independent lab has published side-by-side numbers for the 18-core versus 20-core configurations. Actual throughput for any given model will depend on quantization level, context window length, how well frameworks support NVIDIA’s Arm-based Grace platform, and how the two-fan cooling system holds up under sustained GPU load rather than short bursts. Buyers treating the “120 billion parameters” figure as a guarantee of a specific speed or accuracy are extrapolating beyond what’s been verified.
This is also a different approach than running models through Ollama on consumer hardware, where users are typically working within 8GB to 24GB of VRAM on a discrete GPU. The Surface Laptop Ultra’s unified memory pool removes the usual bottleneck of a model being too large for VRAM but too slow to run split across system RAM, which is the core architectural bet behind the pricing premium.
Satya Nadella on hybrid intelligence
Microsoft has tied the Surface Laptop Ultra launch to a broader “hybrid intelligence” framing for Windows, the idea that some AI workloads run locally on-device while others still route to the cloud depending on task complexity. Speaking around the launch, Microsoft Chairman and CEO Satya Nadella said “just having the [AI] model doesn’t do much for anything,” arguing that raw model access matters less than the surrounding software experience built around it. Nadella also said “Copilot now truly understands your PC,” describing new local-context capabilities that let Copilot reason about files, settings, and activity on the device itself rather than relying entirely on cloud inference.
Microsoft’s own messaging reinforces the same point from the platform side. In material published alongside the announcement, the company said “hybrid intelligence brings frontier AI capabilities to Windows locally,” positioning the Surface Laptop Ultra as the flagship hardware proof point for that strategy rather than a one-off high-end SKU.
Copilot+ PC certification gets complicated
Microsoft’s Copilot+ PC program was built around a specific bar: a neural processing unit capable of at least 40 trillion operations per second, or 40 TOPS. Qualcomm’s Snapdragon X-series chips were among the first to clear that threshold, and Microsoft’s own business product listings label six-core Snapdragon X2 Plus configurations as Copilot+ PCs rated at 80 TOPS.
The Surface Laptop Ultra doesn’t fit neatly into that same framework. Its RTX Spark platform is marketed around Blackwell GPU throughput and the one-petaflop FP4 figure rather than a conventional NPU TOPS rating, and reporting around the launch does not list a TOPS number for the laptop at all. That distinction matters: TOPS measures nominal integer-operation throughput on a dedicated NPU block, while FP4 petaflop figures describe a different category of GPU-driven AI compute. The two numbers are not interchangeable, and it should not be assumed the Surface Laptop Ultra automatically qualifies for every Copilot+ feature simply because its aggregate AI compute is higher on paper than a typical Snapdragon X2 machine. Copilot+ eligibility is ultimately a function of Microsoft’s platform certification and feature support, not raw horsepower.
How it compares to the rest of the Surface lineup
The Surface Laptop Ultra is a different category of product than the Surface Laptop 13.8-inch (8th Edition) with Snapdragon X2 and the Snapdragon-based Surface Pro lineup Microsoft has been selling throughout 2026, including the Surface Pro built on Snapdragon X2 Plus that starts at $1,149. Those machines are built around integrated Qualcomm NPUs, thinner chassis, and battery-first efficiency. The Surface Laptop Ultra inverts those priorities: it trades thinness and all-day battery efficiency for a discrete-class Blackwell GPU, a much larger memory ceiling, and a cooling system designed around sustained AI workloads rather than short bursts of on-device inference.
Microsoft does say the Surface Laptop Ultra retains more than 99% of its performance while running on battery, and describes all-day battery life in its materials, though a specific watt-hour capacity or standardized video-playback runtime figure has not been published alongside the launch. That omission is notable given how central battery claims usually are to Surface marketing.
Surface Laptop Ultra vs competing AI PCs
The AI PC category has gotten crowded fast in 2026, and the Surface Laptop Ultra is entering a field that includes Apple’s unified-memory Apple Silicon Macs, Qualcomm Snapdragon X2 Copilot+ machines from HP and Lenovo, and a wave of RTX Spark-based machines from PC makers racing to ship NVIDIA’s superchip platform, including the Dell XPS 16 RTX Spark Creator Edition, which starts at $3,799 with 18 cores and a 99Wh battery.
| Machine | Platform | Max Memory | Starting Price | Primary AI Angle |
|---|---|---|---|---|
| Surface Laptop Ultra | NVIDIA RTX Spark N1X (Grace CPU + Blackwell GPU) | 128GB unified | $2,599.99 | Large local LLM inference, up to 120B parameters |
| Dell XPS 16 RTX Spark Creator Edition | NVIDIA RTX Spark N1X | Not confirmed in full configuration | $3,799 | Creator workloads plus local AI |
| Surface Pro (Snapdragon X2 Plus) | Qualcomm Snapdragon X2 Plus NPU | Lower consumer tier | $1,149 | Copilot+ efficiency, 80 TOPS NPU |
| Apple Silicon MacBook Pro | Apple unified memory architecture | Varies by generation | Varies | Mature local-AI software ecosystem, no verified head-to-head benchmark yet |
The honest read here is that no independently verified benchmark currently pits the Surface Laptop Ultra against a MacBook Pro, a Dell XPS, an HP, or a Lenovo AI PC running the same local model at the same quantization level. Editorial framing that calls the Surface Laptop Ultra a “MacBook Pro killer” should be read as marketing shorthand, not analyst consensus, since no named analyst quote evaluating the product directly has surfaced in coverage of the launch.
The memory crisis shaping every AI PC’s price tag
The Surface Laptop Ultra isn’t launching into a stable memory market. Reporting around the 2026 DRAM and NAND shortage describes memory prices rising as much as 500% in some segments, with 128GB DDR5 reportedly reaching $3,399 at the module level, driven largely by AI data-center demand absorbing supply that would otherwise go to consumer devices. Micron has said its memory supply is already largely sold out heading into 2027, which helps explain why Microsoft’s pricing jumps so steeply between the 24GB base configuration and the 128GB top tier: the company isn’t just charging for silicon, it’s passing through genuinely scarce component costs.
That scarcity is also reshaping the broader PC market the Surface Laptop Ultra is selling into. Global PC shipments fell sharply in Q3 2026, and industry estimates put average PC selling prices rising as much as 8% for the year, with a more moderate scenario in the 4% to 6% range, as DRAM and NAND supply growth of roughly 16% to 17% fails to keep pace with combined AI and consumer demand. A laptop built around a 128GB unified memory pool is, by definition, more exposed to that pricing pressure than a conventional 16GB or 32GB machine, which is part of why the entry Surface Laptop Ultra configuration caps out at 24GB rather than offering a cheaper 64GB or 128GB tier at a lower price.
Why Microsoft is betting on NVIDIA instead of Qualcomm or Intel
Microsoft’s decision to build its flagship AI Surface around NVIDIA’s Grace CPU and Blackwell GPU rather than Qualcomm’s Snapdragon X2 or an Intel Panther Lake chip is itself a signal. Qualcomm’s Snapdragon X2 Plus remains the company’s Copilot+ workhorse, and no confirmed Surface Laptop Ultra configuration uses Intel Panther Lake or any AMD Ryzen AI silicon. The RTX Spark platform instead borrows architecture concepts directly from NVIDIA’s data-center Grace CPU line, scaled down into a laptop chassis. That’s a meaningfully different bet than putting a mobile-first NPU in a thin chassis: it treats the laptop as a shrunk-down AI workstation rather than an efficiency-first productivity machine, which is exactly the framing behind Microsoft’s own description of the device as built for local AI and 3D graphics rather than all-day battery life first.
Historical context: from Surface Pro 3 to an AI workstation
Microsoft’s Surface line started in 2012 as an attempt to prove Windows hardware could compete on design with Apple’s MacBooks. For most of the following decade, Surface devices chased thinness, battery life, and pen input rather than raw compute. That changed gradually as Copilot+ PCs arrived in 2024 with Qualcomm’s Snapdragon X Elite, introducing the 40-TOPS NPU baseline that defined the category. The Surface Laptop Ultra marks the most dramatic departure yet from that formula: instead of iterating on NPU efficiency, Microsoft partnered with NVIDIA to put genuine discrete-GPU-class AI compute into a laptop chassis, chasing a different buyer entirely, one who wants to run a 70-billion or 120-billion parameter model locally rather than simply get faster background Copilot features.
Market impact: pricing pressure and the AI PC category
The Surface Laptop Ultra’s launch lands at a moment when the broader AI PC category is still searching for its defining use case. Copilot+ PCs sold on the promise of faster background AI features, but few buyers reported a single killer app tied directly to the NPU. The Surface Laptop Ultra is a bet that developers, researchers, and power users who want to run models like Llama or Qwen variants locally, without API costs or data leaving the device, represent a real and growing market segment distinct from mainstream Copilot+ buyers. If that bet pays off, expect competing PC makers shipping RTX Spark-based machines, including Dell, HP, Lenovo, and ASUS, to lean harder into the “local AI workstation” positioning rather than competing purely on NPU TOPS figures.
The risk is that the memory-driven pricing makes these machines a niche product for a market still recovering from a steep 2026 PC shipment decline. A $2,599 entry price for 24GB of unified memory is a tough sell against a $1,149 Snapdragon X2 Surface Pro for buyers who don’t have a concrete local-LLM use case, and the jump to $5,899 for the 128GB configuration that actually unlocks 120-billion-parameter model support puts the most capable version of the Surface Laptop Ultra well outside mainstream laptop budgets.
What developers should actually expect
For developers evaluating whether the Surface Laptop Ultra is worth the premium over a conventional Copilot+ PC, the practical answer depends heavily on which configuration they’re considering. The 24GB base model is realistically suited to smaller quantized language models and typical local developer workloads, not frontier-scale models. The 32GB to 64GB range handles larger quantized models and simultaneous AI applications more comfortably. Only the 128GB configuration is positioned for the headline 120-billion-parameter use case Microsoft is marketing around.
Software compatibility is the other open question. Running models through Ollama or similar local-inference tools on an Arm-based Grace CPU paired with CUDA requires ecosystem support that is still maturing. Developers used to running inference on a standard x86 machine with a discrete NVIDIA GPU should expect some friction porting workflows to the Grace-Blackwell architecture, even though the GPU itself is CUDA-compatible Blackwell silicon.
Predictions: where the Surface Laptop Ultra and AI PC category go from here
- Expect Microsoft and NVIDIA to publish standardized tokens-per-second benchmarks for named open models within the next one to two quarters, once reviewers get hands-on units after the October 16, 2026 shipping date.
- Competing OEMs shipping RTX Spark machines, including Dell, HP, and Lenovo, will likely introduce their own 128GB-class configurations through 2027 as the “local AI workstation” segment separates from mainstream Copilot+ laptops.
- Memory pricing pressure will probably keep entry configurations capped around 24GB to 32GB through 2027, since DRAM supply growth of 16% to 17% is unlikely to fully absorb combined AI and consumer demand before then.
- Microsoft will likely clarify Copilot+ certification rules for GPU-driven AI PCs like the Surface Laptop Ultra, since the current 40-TOPS NPU framework doesn’t map cleanly onto FP4 petaflop figures.
- Apple is likely to respond with marketing emphasizing its own unified-memory architecture for local AI workloads, particularly if Surface Laptop Ultra sales data suggests real demand for a 128GB local-LLM laptop.
Specs at a glance
A quick reference for the confirmed baseline configuration versus what’s still unverified, now that the laptop has been available since October 16.
- Confirmed: $2,599.99 starting price, October 7 preorders, October 16 shipping, 18/20-core Grace CPU options, 5,120/6,144 CUDA core Blackwell GPU, 24GB to 128GB unified memory, 15-inch mini-LED 120Hz display up to 2,000 nits, 512GB base storage, up to 1 petaflop FP4 AI compute, support for models up to 120 billion parameters.
- Not yet confirmed: a conventional NPU TOPS rating, exact battery watt-hour capacity, a verified named-model tokens-per-second benchmark list, independent head-to-head results against MacBook Pro, Dell XPS, HP, or Lenovo machines, and any named analyst consensus on the product.
Frequently asked questions
What is the starting price of the Surface Laptop Ultra?
The base configuration starts at $2,599.99, with 24GB of unified memory, an 18-core NVIDIA Grace CPU, and a 512GB SSD.
When does the Surface Laptop Ultra ship?
Preorders opened October 7, 2026, with general availability beginning October 16, 2026.
Can the Surface Laptop Ultra really run large language models locally?
Microsoft says the 128GB configuration can keep models as large as 120 billion parameters loaded for local inference. No independent, standardized tokens-per-second benchmark for specific named models has been published yet.
What chip powers the Surface Laptop Ultra?
It uses NVIDIA’s RTX Spark N1X platform, combining an Arm-based NVIDIA Grace CPU (18 or 20 cores) with a Blackwell GPU carrying 5,120 or 6,144 CUDA cores, depending on configuration.
Is the Surface Laptop Ultra a Copilot+ PC?
That isn’t clearly confirmed. The RTX Spark platform is marketed around FP4 petaflop GPU throughput rather than the conventional 40-TOPS NPU threshold Microsoft uses to define Copilot+ PCs, so its exact certification status is unclear from available reporting.
How much does the top-tier 128GB Surface Laptop Ultra cost?
The fully configured version with 128GB of unified memory and a 1TB SSD is reported at $5,899.
How does it compare to the Surface Laptop 7?
It’s a different product category entirely. The Surface Laptop 7 and related Snapdragon Surface devices prioritize thinness, efficiency, and battery life with an integrated NPU, while the Surface Laptop Ultra prioritizes raw AI compute and memory capacity with a discrete-class Blackwell GPU and a more aggressive cooling system.
Why is the Surface Laptop Ultra priced so much higher than a typical Copilot+ PC?
Part of the premium reflects the NVIDIA Grace-Blackwell platform itself, and part reflects the 2026 memory shortage, which has pushed DRAM and NAND costs sharply higher industry-wide, particularly for machines with large unified-memory configurations like the Surface Laptop Ultra’s 128GB top tier.
