Microsoft benchmarks RTX Spark laptops at 64GB, not the 128GB it sells

Microsoft sells its new RTX Spark laptops on the promise of running AI models that won’t fit on a traditional machine. The footnotes under its own announcement show the three speed claims it published against a MacBook Pro were measured on machines carrying half the memory it advertises.

The post published Wednesday by Pavan Davuluri, Microsoft’s executive vice president of Windows and Devices, claims up to 2.1x faster time to first token, 4.3x faster AI image generation and 6.2x faster AI video generation than an Apple MacBook Pro 16-inch with M5 Pro. Footnote five says the Apple machine had 64 GB of memory. Footnotes six, seven and eight each put the RTX Spark machines it was tested against at 64 GB too. Microsoft’s Surface Laptop Ultra page sells up to 128 GB.

That matters because the memory ceiling is the pitch. Davuluri said as much on stage, in remarks written up by NVIDIA.

With up to 128 gigs of unified memory and up to a petaflop of AI compute, you can run models on this laptop that simply don’t fit on a traditional machine.

Pavan Davuluri, EVP Windows and Devices, via NVIDIA

But none of the three tests ran a model like that. The footnotes name every one of them, and they run from 4 billion to 27 billion parameters, each quantised down from full precision. They also name who did the measuring, which is where the second problem sits.

Claim	What the footnote says ran	Who measured it
2.1x faster time to first token	Qwen3.5 27B (Q4K-Medium) in llama.cpp, fixed 8,192-token prompt, 64 GB units	Microsoft-commissioned
4.3x faster AI image generation	FLUX.2 Klein 4B at NVFP4, four sampling steps, 1024 x 1024 in ComfyUI	NVIDIA
6.2x faster AI video generation	LTX 2.3 22B at NVFP4, 121 frames at 25 fps, eight steps, 1280 x 720 in ComfyUI	NVIDIA

So the chip vendor measured two of the three wins itself, and every Windows machine in the comparison was preproduction, tested in September 2026. Those three figures don’t appear in either the TechCrunch report or the CNBC report on the event. Apple’s memory ceiling is higher still: we reported a 512GB maximum when the M5 Ultra launched, on Apple’s own figures.

And Microsoft’s own two pages don’t agree on how large a model the laptop takes. The blog post says it can run models “exceeding 120 billion parameters” locally. The Surface product page says CUDA accelerates local agent workflows “powered by models up to 120B parameters”, which is the opposite boundary.

The prices are firmer, though they still shift between outlets. Microsoft’s store lists the Surface Laptop Ultra at $2,599.99 for preorder, and CNBC reported the start price as $2,599. TechCrunch put the second base configuration at $3,700 and said prices rise to $5,900 with more memory and storage, and that the highest-end device is already out of stock. The Surface RTX Spark Dev Box starts at $6,000, per TechCrunch.

Another claim narrows once you read the terms behind it. TechCrunch described up to $1,000 off for trading in a MacBook Pro, which reads as an Apple-specific swipe. Microsoft’s terms name no brand, saying only “qualifying device”, and the cash back varies with the device bought and the trade-in value. The offer runs only through the online Microsoft Store in the US and Canada between 7 October and 23 November 2026.

Still, the hardware underneath is real enough. Jensen Huang told the event that the Blackwell RTX Spark GPU matches NVIDIA’s DGX-1 server from 2016, which cost $250 million, as CNBC reported. Microsoft also made Execution Containers generally available on Windows 11, the sandbox layer that lets agents run under operating system control, with support listed from OpenAI Codex, GitHub Copilot, OpenClaw, Replit, LM Studio and Unsloth AI.

Shipping dates are the next thing to watch, and the two companies don’t quite line up there either. Microsoft says every RTX Spark laptop begins shipping on 16 October, while NVIDIA puts the compact desktops in November. Until someone publishes an independent run at 128 GB on a shipping unit, the case for buying one over a local machine with no token fees rests on tests nobody outside the two vendors has repeated.