San Francisco, October 7. Microsoft and Nvidia turned the RTX Spark superchip from a June teaser into shipping hardware today, opening pre-orders for Surface Laptop Ultra and partner systems starting near $2,999 with mid-October delivery. The top configuration runs to roughly $7,000. Satya Nadella and Nvidia’s Jensen Huang shared the stage with Windows chief Pavan Davuluri at the company’s first dedicated Windows event in over two years, billed as a conversation on how local AI shapes the next chapter of the PC.
Two configurations, one memory story
Pre-orders opened today. Two chip configurations anchor the line.
| Config | CPU/GPU | Unified memory | Price |
|---|---|---|---|
| RTX Spark S2 | 18-core CPU, 5,120-core GPU | 24-32GB | from about $2,999 |
| RTX Spark S3 | 20-core CPU, 6,144-core GPU | up to 128GB | up to about $7,000 |
Dell, HP, Lenovo, MSI, ASUS and Acer all have designs coming, though none of the six laptops Nvidia has listed so far is a gaming model, a gap that has drawn comment from the enthusiast press. Microsoft also dropped the Copilot+ PC branding for this generation of Surface. Third-party OEM pricing for partner laptops was expected at the event as well, with most partner builds landing in creator and productivity lineups first. The Surface RTX Spark Dev Box, a compact developer mini PC shown earlier with a 100W thermal envelope and all-aluminum cooling body, was also expected to make an appearance with shipping details.
What RTX Spark physically is
The chip is a shrunk Grace Blackwell datacenter superchip: 20 Arm CPU cores and 6,144 Blackwell GPU cores on one die at 3nm, joined by NVLink-C2C and sharing up to 128GB of LPDDR5X at roughly 300GB/s of bandwidth. The shared pool is the point. A laptop can load a 120-billion-parameter model that needed a multi-GPU server rig two years ago, since the GPU does not copy model weights back and forth from separate VRAM, and it does so at up to only 80 watts of power draw. Microsoft claims around a petaflop of AI compute for the top configuration, a vendor headline figure based on theoretical FP4 TOPS with sparsity. Users also get manual controls for how much unified memory goes to graphics and AI on the fly. Nvidia says the top spec supports up to 1 million tokens of context for local agent sessions.
Adobe is rebuilding Premiere and Photoshop around the chip and claims double the speed on its AI features. Nvidia’s claims cover 12K 4:2:2 video editing, 4K AI video generation, 90GB-plus 3D scenes, and 1440p AAA gaming above 100 frames per second with ray tracing and DLSS. Windows on Arm support has widened too: Prism, the x86 emulation layer, now handles AVX and AVX2 instructions, and Epic’s Easy Anti-Cheat and BattlEye have shipped native Arm64 builds, with Valorant, League of Legends and PUBG confirmed for the platform.
CUDA is the wedge
Against an Apple MacBook Pro with an M4 Max, the raw comparison is not flattering on bandwidth: Apple delivers about 546GB/s to the RTX Spark’s roughly 300GB/s, and a 128GB MacBook Pro starts at $3,999, well under the comparable $7,000 Surface configuration. What the Windows machines get in exchange is the CUDA ecosystem. RTX Spark runs PyTorch with full CUDA acceleration, TensorRT, Llama.cpp, Hugging Face tooling, Unsloth and Kohya, and Apple’s Metal stack runs none of that. Much of the local fine-tuning and inference world still assumes CUDA, so the real decision between the platforms is framework compatibility against raw memory speed. Memory chip cost explains most of the price gap: 128GB of LPDDR5X is a serious line item no vendor escapes, and Microsoft has not said what the high-memory configurations cost once configured upward.
AMD’s Ryzen AI Max systems shipped earlier at lower price points, so Microsoft is not even first to the local-AI laptop category on Windows. The pitch is that nothing else carries the full Blackwell software stack, with sheer memory capacity deciding what models fit at all: a 128GB PC runs model sizes a 64GB one simply cannot load.
The Windows layer may matter more than the silicon
The quieter half of the announcement will land heaviest on software people. Windows 11 gains native Model Context Protocol support as a first-class OS feature, plus a File Explorer Connector and a Settings Connector so agents can read files or change system configuration under explicit permissions. Agent Workspaces put agents in their own sandboxed sessions with separate accounts and a virtualized desktop, isolated from the user’s main session. Microsoft Execution Containers, shown at Build, let developers declare what an agent can access while Windows enforces those boundaries at runtime. Nvidia’s OpenShell runtime adds identity, containment and policy layers on top, letting users define what agents can and cannot do, and mask personal information in queries routed to cloud models.
Microsoft’s Peter Dawould pitched the container work in an Nvidia video as making Windows the most secure place to run your own agents on an RTX Spark PC. The company also rebuilt parts of the Windows 11 task scheduler with Workload Profile Scheduling to spread work across all 20 Arm cores, and raised how much system memory the GPU can reach. Three in-house models run on the hardware: Aion 1.0 Instruct, Aion 1.0 Plan for reasoning and tool calling, and MAI Code Flash for code completion, all designed to run entirely on-device.
Nadella framed the launch plainly: “Our goal is to deliver unmetered intelligence to every home and every desk with Windows.” Privacy questions shadow the agent push, just as they did the Recall feature two years ago, and Microsoft has been down this road before with agent features that shipped sandboxed and reversible after early criticism. Whether a $7,000 laptop answers those questions, and sells, is the thing the pre-order numbers will start answering this month. Nvidia’s DGX Station for Windows extends the same architecture to enterprise desks for teams that outgrow the laptop configs.