Microsoft Puts Nvidia in the Surface Laptop Ultra Flagship

Microsoft opened pre-orders October 7 for the Surface Laptop Ultra, an Nvidia RTX Spark laptop that starts at $2,599 and ships October 16. Nvidia chief Jensen Huang shared the San Francisco stage with Satya Nadella. Intel and Qualcomm are not in this machine.

The same afternoon, Microsoft priced a screenless Surface RTX Spark Dev Box at $5,999, for shipment in November in the United States only. Both boxes are sold as hardware for “hybrid intelligence,” local models first and Azure when a task still needs the cloud.

The $2,599 Surface Laptop Ultra

Brett Ostrum, corporate vice president of Surface, posted pre-orders for two new devices built around Nvidia’s RTX Spark superchip. The laptop unites an Nvidia Grace CPU with up to 20 cores, a Blackwell RTX GPU with up to 6,144 cores, and up to 128 GB of unified memory shared by both. Microsoft says that top setup can run models exceeding 120 billion parameters on the device, with up to 1 petaflop of AI speed. That petaflop figure is theoretical FP4 math that uses sparsity.

The base unit is not that machine. For $2,599 you get the cut-down 18-core CPU, a 5,120-core GPU, 24 GB of unified memory, 512 GB of storage, and Windows 11 Home in Platinum. Pavan Davuluri, executive vice president of Windows and Devices, said Microsoft built it for developers and creators pushing performance limits. The store ladder makes that audience plain.

SURFACE LAPTOP ULTRA STORE PRICES

The chassis is a 15-inch PixelSense Ultra touchscreen in a body less than 18 mm thin and under 4.5 lb, in Platinum or Nightfall. Microsoft rates peak HDR at 2,000 nits, 25 percent brighter than MacBook Pro M5 as of October 1, and says a new thermal stack moves 2.5 times more heat than current Surface laptops. The trackpad is 30 percent larger than on Surface Laptop 8th Edition, storage is user-removable, and a Magnetic Connect port is the first built-in magnetic USB-C charge port on a laptop. Two more USB-C ports, HDMI, USB-A, an SD reader, and a headphone jack can drive three 4K screens. Pen input is not supported.

Microsoft Surface posted the launch film the same afternoon.

A new generation of Surface is here. pic.twitter.com/TEsVzs8Pma

— Microsoft Surface (@surface) October 7, 2026

A $5,999 Box With No Screen

Surface RTX Spark Dev Box uses the 20-core chip, 6,144 GPU cores, and 128 GB in a compact desk unit with 2 TB of removable storage. It is $5,999, sold only on Microsoft.com in the United States, and ships in November. Visual Studio Code, Git, GitHub Copilot, WSL, Python, and Node come installed. Buyers bring their own monitor and keyboard. Microsoft is also offering up to $1,000 cash back on an eligible MacBook Pro trade-in with a Laptop Ultra bought between October 7 and November 23, in the United States and Canada only.

Nvidia Walks Into the Windows Flagship

For decades Nvidia sold add-in graphics into PCs that Intel, and later AMD or Qualcomm, still owned at the system level. RTX Spark is a full Arm superchip: Grace CPU and Blackwell GPU linked at 600 GB/s, CUDA native, DLSS, Reflex, and G-Sync on a Windows on Arm laptop. Huang told the room Nvidia was founded because of Windows, and that without Windows there would be no GeForce. He said the chip took more than 4,000 engineering years.

Davuluri tied the laptop to that silicon in one line, saying Microsoft built Surface Laptop Ultra around RTX Spark, with up to 128 GB of unified memory and up to a petaflop of AI compute, so models that do not fit a traditional machine can run on this one. Nvidia named the same-day laptop wave from Acer, ASUS, Dell, HP, Lenovo, MSI, and Gigabyte, with compact desktops due in November. Gears of War: E-Day is a native title on the chip. Call of Duty is listed for 2027.

WHO GETS THIS FLAGSHIP

• Nvidia: A Windows system chip with CUDA on Microsoft’s most powerful Surface, plus a seven-brand OEM wave shipping October 16.
• Microsoft: A $2,599 to $5,899 “world makers” story, a $5,999 desk box, and a MacBook trade-in aimed at creators who already pay Apple money.
• Intel: No socket in Surface Laptop Ultra, after years as the default inside Surface notebooks.
• Qualcomm: Still the chip in the cheaper Copilot+ Surfaces, and absent from this one.
• 24 GB buyers: Flagship money for a config that cannot load the 120 billion parameter models in the keynote.

AMD still sells Windows chips with large unified memory of its own. They are not in this Surface either. The signal is the stage: Huang stood next to Nadella while Davuluri described a superchip Intel does not make.

Apple Already Sells Cheaper Local AI Desks

Apple put new Mac mini and Mac Studio machines on sale September 22, two weeks before the San Francisco event, and sold them as on-device AI boxes that do not meter tokens. Johny Srouji, Apple’s chief hardware officer, called Mac Studio the desk for on-device AI and heavy pro work. Mac Studio with M5 Max starts at $2,499 with up to 128 GB of unified memory. Mac Studio with M5 Ultra starts at $5,499 and reaches 512 GB, with that top memory config due in late October. Mac mini with M6 starts at $899, and the M5 Pro mini starts at $1,699.

LOCAL AI BOXES AT LIST PRICE

The $2,599 Surface is a laptop, so the Mac Studio row is not a like-for-like spec war. The list still shows the hole in Microsoft’s pitch. Apple already sells a 128 GB local-AI desk for $2,499, $100 under the cheapest Surface Laptop Ultra, and a 512 GB Ultra desk for $5,499, $500 under the Dev Box that tops out at 128 GB. Microsoft’s answer is CUDA, Windows agents, and a $1,000 MacBook bounty, not a lower sticker.

Qualcomm’s Copilot+ Deal Stopped at the Flagship

Microsoft spent 2024 telling the market an AI PC was a Snapdragon X laptop with a 45 TOPS Hexagon NPU, 16 GB of memory, and a $999 start. Those Copilot+ Surfaces went up for pre-order in May 2024 and began shipping June 18, 2024. Intel and AMD later cleared the same TOPS bar in thin x86 notebooks. The volume story was a $999 to $1,500 everyday machine that would run small models offline and keep Copilot close to the keyboard.

THE ROAD TO THIS CHIP

1. May 20, 2024: Microsoft opens Copilot+ Surface laptops and Surface Pro models on Qualcomm Snapdragon X, from $999, with a 45 TOPS NPU as the entry ticket.
2. June 18, 2024: Those first Copilot+ Surfaces begin shipping, with Windows on Arm and Prism emulation for x86 apps.
3. 2025: Intel Lunar Lake and AMD Strix-class chips join the Copilot+ label, so Qualcomm no longer owns the badge.
4. October 7, 2026: Microsoft puts Nvidia, not Qualcomm or Intel, in Surface Laptop Ultra and prices it from $2,599.
5. October 16, 2026: The laptop and the wider RTX Spark OEM wave begin shipping.
6. November 2026: Surface RTX Spark Dev Box begins shipping in the United States at $5,999.

The $999 Copilot+ PC still exists in the catalog. It is no longer the machine Microsoft flies Huang across the country to bless. Qualcomm keeps the thin-and-light slot. Nvidia gets the SKU Microsoft calls its most powerful Surface laptop ever. That split is the product map now: efficiency Arm from Qualcomm for battery life, compute Arm from Nvidia for models that need tens of gigabytes in one pool.

The $1,600 Memory Step Most Buyers Will Skip

Unified memory is the whole local-AI bet, because a 120 billion parameter model will not fit in the 8 GB or 16 GB frame buffer on a normal laptop GPU. Microsoft’s own footnote says the 128 GB pool is shared by CPU and GPU, and that the GPU cannot address all of it. The store prices show what that pool costs once you leave the demo unit.

MEMORY STEPS THAT CHANGE THE DEMO

• 24 GB base: $2,599, the config most shoppers will actually open, far below the 120 billion parameter claim.
• 64 GB: $4,299 with 1 TB, the last step before the demo memory class.
• 128 GB: $5,899, a $1,600 jump from the 64 GB 1 TB row for the extra memory.
• Dev Box: $5,999 locks 128 GB with no screen, so the memory is the product.

Nvidia is already selling a related desk box under a different name. DGX Spark 64GB starting at $4,999 arrives October 23 from Acer, ASUS, Dell, Gigabyte, HP, and MSI, rated for models up to 100 billion parameters on device. Two of those units can be cabled together to pool 128 GB. The Windows laptop that is supposed to make that class of model ordinary still asks $5,899 before tax for the 128 GB SKU. The loud complaint after the $2,599 post was not battery life or Arm apps. It was that the memory step between the machine you can order and the machine in the keynote is a second laptop.

Microsoft AI said MAI-Code-1.1-Flash will run on-device inside GitHub Copilot, with no inference charge for local model calls. That pitch only holds on silicon that can hold the weights. On the 24 GB row, Copilot still phones the cloud, and the cloud still bills.

Hybrid Intelligence Still Bills the Cloud

Nadella’s line at Dogpatch Studios was that users will stop asking whether a job ran locally or in the cloud. Davuluri’s hardware pitch is the other half of the same idea: keep more work on the device, spend fewer cloud tokens, and let Azure take the jobs that still do not fit. Windows is shipping Microsoft Execution Containers as the cage for agents that stay running in the background, with Agent 365 controls meant to tell an agent apart from a person.

The trajectory for me is so clear. There will never be a moment where we will go and look and say: ‘Oh, this runs locally, this runs in the cloud.’ You will expect this hybrid intelligence to be everywhere and pervasive.

Satya Nadella, CEO, Windows and Surface event in San Francisco

Nvidia’s own example on the same stage was Qwen 3.8 Flash Next, a 125 billion parameter model it says can match many cloud models without sending data off the PC. That is a 128 GB story, not a $2,599 story. Huang also previewed DGX Station for Windows, a GB300 deskside box with 748 GB of coherent memory and up to 20 petaflops of FP4, enough in Nvidia’s telling for models at a trillion parameters. No ship date was attached. The Station is the honest scale of “frontier on the desk.” Surface Laptop Ultra is the billboard that gets CUDA into Best Buy.

Copilot still leans on OpenAI and Anthropic models in Microsoft’s cloud for the hard prompts. Microsoft has been trying to shrink that dependence, and the on-device MAI-Code demo is part of that. Until the 128 GB SKU is the one people buy, hybrid intelligence is a routing rule: local when the weights fit, Azure when they do not. Intel and Qualcomm lost the flagship photo. Nvidia won a Windows superchip. Shoppers who pay $2,599 still rent the model that made the slide.

Surface Laptop Ultra begins shipping October 16, with partner RTX Spark laptops the same day. Review units are still under embargo until then, so the 24 GB base has not been timed against the 120 billion parameter claim in independent tests. The $5,899 row is the config that claim was written for.