InfrastructureCompute, chips & cloud
Nvidia's RTX Spark Arm laptops ship with Windows tuned for emulation, unified memory and local agents, but without Copilot+ features
Nvidia's first Arm-based Windows laptops go on sale with Microsoft tuning for x64 emulation, unified memory and sandboxed local agents. Early tests show strong large-model inference, while Copilot+ features such as Recall are absent at launch.
From the The Verge · AI article of October 15, 2026 and cited sources: Microsoft's platform work centers on Prism emulation, which is tuned to the RTX Spark CPU and supports AVX/AVX2. [1] [2] [4]
SOURCES[1] How Windows is being optimized for Nvidia’s RTX Spark… theverge.com[2] Windows Experience Blog — Introducing… RTX Spark (May 31) blogs.windows.com[4] Windows Experience Blog — Building Windows for hybrid… blogs.windows.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Microsoft's platform work centers on Prism emulation, which is tuned to the RTX Spark CPU and supports AVX/AVX2. [1] [2] [4]
- Original source
- The Verge · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 15, 2026 article published
- Covered
- Oct 16 edition
- Evidence
- 10 sources across 7 sites
- 1How Windows is being optimized for Nvidia’s RTX Spark…theverge.com
- 2Windows Experience Blog — Introducing… RTX Spark (May 31)blogs.windows.com
- 3Tom's Hardware — Computex 2026 unveilingtomshardware.com
- 4Windows Experience Blog — Building Windows for hybrid…blogs.windows.com
- 5Jon Peddie Research — RTX Spark laptops have arrivedjonpeddie.com
- 6StorageReview — Surface Laptop Ultra for Business reviewstoragereview.com
Reach: not measured yet
Microsoft's platform work centers on Prism emulation, which is tuned to the RTX Spark CPU and supports AVX/AVX2. It also changes how Windows handles unified memory and holds some memory back so the system stays responsive. Agents run in the MXC sandbox, and supported agents such as Hermes, Perplexity and OpenClaw install from the Store. [1] [2] [4]
With up to 128GB of unified memory, the laptops can run models that don't fit in a discrete GPU's VRAM. StorageReview measured Qwen3.6-35B-A3B at 89.2 tokens/s and gpt-oss:20b at 58.0 tokens/s, both at 64K context. On smaller models, an RTX 5090 laptop generated tokens 2.3× faster. Microsoft's MacBook Pro comparisons are vendor-reported and lack published benchmark details. [5] [6]
Some things are missing at launch. Recall and other Copilot+ features are absent, even though Microsoft said in May the laptops would join that category. Anti-cheat support for Call of Duty and Battlefield is not expected until 2027. Reviewers also disagree on gaming: one calls it good enough for play on the go, another says gaming is not its strength. [1] [2] [10]
Review benchmarks show 128GB of unified memory running models too large for discrete laptop GPUs at usable speeds. That points to a portable option for local AI and agent development, though smaller models still run faster on an RTX 5090 laptop.
The dossier is below. It draws on independent reviews and analysis plus Microsoft and Nvidia blog posts.
Executive brief
Nvidia's first Arm-based Windows laptops ship today, October 16, 2026. They ship without Copilot+ PC features such as Recall, even though Microsoft said in May they would join that category. A Microsoft marketing director told The Verge the features are expected to reach Windows Insiders "later this year". Microsoft's tuning work has gone elsewhere: emulation, unified memory, anti-cheat support and installing agents in two clicks. Independent tests show strong local AI with up to 128GB of memory. One review also recorded unexplained power-offs on two review units.
What changed and event timeline
Platform unveiled
At Computex, Nvidia and Microsoft announced RTX Spark: up to 20 Arm cores, a Blackwell GPU and up to 128GB of unified memory. Microsoft said the laptops would join the Copilot+ PC category.
More detail
Preorders and "hybrid intelligence."
Microsoft opened preorders and made Microsoft Execution Containers (MXC), its sandbox for agents, generally available. It also announced llama.cpp support in Windows ML.
Analyst framing
Jon Peddie Research describes the laptops as aimed at premium x86 machines. It notes that Microsoft's speed claims against the MacBook Pro come without detailed benchmarks.
Reviews and Windows details
The Verge reported Microsoft's tuning work and the delay to Copilot+ features. StorageReview published local LLM benchmarks and a reproducible power-off problem.
More detail
Shipping
The Surface Laptop Ultra and partner laptops go on sale, and Nvidia says hundreds of games are available at launch.
Capabilities and access
- Chip: the RTX Spark N1X has a 20-core Arm CPU co-developed with MediaTek and a Blackwell GPU with 6,144 CUDA cores.
- Devices and prices: the Surface Laptop Ultra runs from $2,599 to about $5,899 with 128GB (Tom's Guide).
- Agents: Hermes, Perplexity and OpenClaw are officially supported at launch and install through the Store (The Verge).
Read the full section
- Chip: the RTX Spark N1X has a 20-core Arm CPU co-developed with MediaTek and a Blackwell GPU with 6,144 CUDA cores. Nvidia rates it at 1 PFLOPS of FP4 compute with sparsity. (Jon Peddie)
- Devices and prices: the Surface Laptop Ultra runs from $2,599 to about $5,899 with 128GB (Tom's Guide). Asus, Dell, HP, Lenovo and MSI also sell RTX Spark laptops.
- Agents: Hermes, Perplexity and OpenClaw are officially supported at launch and install through the Store (The Verge). Microsoft lists Claude Code and Meta's Muse as coming soon (Windows blog).
Technical analysis for researchers and developers
- Prism emulation: Prism is Windows' emulator for x64 apps on Arm.
- Memory: Windows changes the page size it uses for unified memory and keeps some memory back so the system stays responsive.
- Toolchain: native PyTorch and Python 3.15 builds for Windows on Arm, llama.cpp support in Windows ML, and Nvidia's Switchyard for routing requests between local and cloud models (NVIDIA).
Read the full section
- Prism emulation: Prism is Windows' emulator for x64 apps on Arm. It is tuned to the RTX Spark CPU design, supports AVX/AVX2 instructions, and gets compatibility updates every month (Windows blog; The Verge).
- Memory: Windows changes the page size it uses for unified memory and keeps some memory back so the system stays responsive. A slider in Nvidia's app sets how much memory the GPU gets. StorageReview tested with 64GB of 128GB assigned to the GPU, adjustable to 96GB.
- Toolchain: native PyTorch and Python 3.15 builds for Windows on Arm, llama.cpp support in Windows ML, and Nvidia's Switchyard for routing requests between local and cloud models (NVIDIA).
- Storage: random I/O through Windows on Arm measured 300K IOPS, against 1.67M IOPS with raw device access (StorageReview).
Claims and evidence
- Independent: StorageReview measured Qwen3.6-35B-A3B at 89.2 tok/s and gpt-oss:20b at 58.0 tok/s, both at 64K context, plus Phi-4-mini at 97.5 tok/s.
- Vendor-reported: against a MacBook Pro M5 Pro, Microsoft claims 2.1× faster time to first token and 4.3× faster image generation (Windows blog).
- Reviewer's hands-on result: about 60fps in most games at 1440p, medium settings and DLSS Performance, with almost every game running emulated (The Verge).
Read the full section
- Independent: StorageReview measured Qwen3.6-35B-A3B at 89.2 tok/s and gpt-oss:20b at 58.0 tok/s, both at 64K context, plus Phi-4-mini at 97.5 tok/s. On smaller models that fit in 24GB of VRAM, a Razer Blade 16 with an RTX 5090 generated tokens 2.3× faster (StorageReview).
- Vendor-reported: against a MacBook Pro M5 Pro, Microsoft claims 2.1× faster time to first token and 4.3× faster image generation (Windows blog). Jon Peddie notes these claims come without benchmark details.
- Reviewer's hands-on result: about 60fps in most games at 1440p, medium settings and DLSS Performance, with almost every game running emulated (The Verge).
Context and prior work
Windows on Arm became credible with Qualcomm's Snapdragon X laptops, which pushed developers toward native Arm apps and improved Prism. RTX Spark builds on that ecosystem with a much stronger GPU. The idea of local agents in sandboxes extends Nvidia's earlier OpenShell and DGX Spark work (NVIDIA GTC blog).
Read the full section
Windows on Arm became credible with Qualcomm's Snapdragon X laptops, which pushed developers toward native Arm apps and improved Prism. Their GPUs were too weak for serious gaming (The Verge). RTX Spark builds on that ecosystem with a much stronger GPU. The shader-precompilation feature comes from the Xbox Ally handhelds. The idea of local agents in sandboxes extends Nvidia's earlier OpenShell and DGX Spark work (NVIDIA GTC blog).
Limitations, safety and contested findings
- Copilot+ status conflicts: in May, Microsoft said the laptops would join the Copilot+ category.
- Stability: two Surface units shut off identically during an OpenBLAS double-precision LU solve test.
- Gaming verdict differs: Tom's Hardware headlines gaming as "not its strong suit". The Verge calls it good enough for gaming on the go.
Read the full section
- Copilot+ status conflicts: in May, Microsoft said the laptops would join the Copilot+ category. At launch, Recall and other Copilot+ features are missing, with no reason given (The Verge).
- Stability: two Surface units shut off identically during an OpenBLAS double-precision LU solve test. The logs point to power delivery or firmware protection, not heat (StorageReview).
- Gaming verdict differs: Tom's Hardware headlines gaming as "not its strong suit". The Verge calls it good enough for gaming on the go.
- Anti-cheat gaps: Call of Duty (Ricochet) and Battlefield (Javelin) are not expected until 2027. Riot's games have no date.
- Setup: OpenClaw's local gateway needed an uninstall and reinstall. Large local models still respond more slowly than cloud models.
Business and practitioner implications
- Where it fits: 128GB of unified memory lets a laptop hold models that won't fit in a discrete GPU's VRAM.
- For IT teams: MXC policies decide which files and networks an agent can reach. Wait for firmware fixes before buying for sustained, compute-heavy work.
- For software vendors: native Arm64 builds avoid Prism overhead, and native PyTorch and Python remove a long-standing barrier to porting.
Read the full section
- Where it fits: 128GB of unified memory lets a laptop hold models that won't fit in a discrete GPU's VRAM. For smaller models, a discrete RTX 5090 laptop is still faster.
- For IT teams: MXC policies decide which files and networks an agent can reach. Wait for firmware fixes before buying for sustained, compute-heavy work.
- For software vendors: native Arm64 builds avoid Prism overhead, and native PyTorch and Python remove a long-standing barrier to porting.
- For buyers: don't count on Recall or other Copilot+ features at purchase. Check that your games' anti-cheat is supported.
The source trail.
Sources (10)
How Windows is being optimized for Nvidia’s RTX Spark laptops
Article text retrieved; extracted text may omit tables or interactive elements.
theverge.com