NVIDIA and Microsoft are bringing local AI compute, cloud routing and agent access controls together on Windows PCs. At their October 7, 2026 event in San Francisco, the companies outlined RTX Spark hardware, Windows Hybrid Intelligence and Microsoft Execution Containers (MXC): three parts of a platform intended to let agents use a PC’s files and tools while keeping their actions under operating-system control. NVIDIA’s announcement and Microsoft’s Windows announcement describe the integration.
The capabilities arrive on different schedules. MXC is generally available on Windows 11. NVIDIA announced RTX Spark preorders on October 7, with laptops from Asus, Dell, HP, Lenovo and MSI beginning shipment October 16. Hybrid Copilot features are expected in the coming months, while hybrid coding features are scheduled for experimental preview later in October. For developers, the immediate distinction is between hardware they can order, containment they can use and routing features still approaching preview.
RTX Spark supplies the local compute
NVIDIA describes RTX Spark as a combination of a Blackwell RTX GPU and a Grace CPU, with up to 128GB of unified memory and one petaflop of FP4 AI performance. Its published specifications are:
| Component | NVIDIA’s stated specification |
|---|---|
| GPU | NVIDIA Blackwell RTX GPU with up to 6,144 cores |
| CPU | NVIDIA Grace CPU with up to 20 cores |
| CPU–GPU connection | 600 GB/s |
| AI compute | One petaflop of FP4 AI performance |
| Memory | Up to 128GB unified memory |
| Software platform | Full NVIDIA CUDA platform |
These are vendor specifications rather than independently measured application results. The compact desktop configuration uses the same RTX Spark superchip in a chassis designed for 24/7 operation, targeting agents that run continuously. NVIDIA’s hardware and software breakdown supplies these details.
Microsoft says the Surface Laptop Ultra can run models exceeding 120 billion parameters locally. NVIDIA names Qwen 3.8 Flash Next as an example, describing it as a 125B model with 51B n-gram that can run without sending data to the cloud. Local execution is a documented vendor capability claim; equivalent quality across workloads remains unestablished here.
Hybrid intelligence decides where work runs
Windows Hybrid Intelligence adds three connected capabilities to Copilot: using local context, taking actions on the PC and drawing on local models alongside cloud models. GeekWire’s event report describes these capabilities.
For coding, Microsoft says HydraFusion-powered hybrid intelligence is coming to the GitHub Copilot app, GitHub Copilot CLI and Visual Studio Code in experimental preview later in October. GeekWire describes the intended routing as sending simpler coding tasks to a model on the PC and harder tasks to the cloud. The evidence does not specify the routing thresholds or provide measured results for that division of work.
Microsoft’s local-model plans illustrate how it intends to fit large models onto PCs. It describes MAI Code 1.1 Flash as having 137 billion total parameters and 6.8 billion active parameters, using 3-bit precision locally to reduce model size by nearly 80% while supporting a 256K local context window. It also identifies an upcoming NVIDIA Nemotron model with over 70 billion parameters, quantized in 2-bit to use just over 20GB of memory. Microsoft separately announced llama.cpp support in Windows ML. These are Microsoft’s model and runtime claims, not independent workload tests. Microsoft’s announcement sets out the plans.
The architectural implication is that local inference is one execution option within a broader service. A PC able to run a model locally does not establish that every Copilot task stays on that device. The announced routing explicitly retains cloud execution, making the choice of execution location an important part of evaluating a workflow.
MXC governs what agents can reach
Microsoft Execution Containers address a separate requirement: controlling an agent’s access to files and networks. Microsoft says organizations can define those permissions and have policies enforced at runtime. Its Windows security approach combines containment, identity and manageability: limiting access, distinguishing an agent’s actions from a person’s actions, and connecting governance to Agent 365. Microsoft’s MXC description explains that model.
Microsoft lists existing MXC support from OpenAI’s Codex, GitHub Copilot, OpenClaw, Replit, LM Studio, NVIDIA OpenShell and Unsloth AI. GeekWire reports that Microsoft also announced planned support from Anthropic’s Claude Code and Perplexity. Meta’s Muse is coming as a native Windows app with MXC integration. Existing support and announced integrations should therefore be treated as different stages of ecosystem availability.
Compute and containment solve different problems. RTX Spark supplies resources for local models; MXC defines the resources an agent may access while acting. Neither the published specifications nor the containment announcement establishes that an agent will complete a task correctly. For platform teams, the useful evaluation is whether the particular agent supports MXC and whether its permitted files and networks match the intended workflow.
Devices, prices and performance qualifications
The Surface Laptop Ultra has a 15-inch display and up to 128GB of memory. GeekWire reports a $2,599 starting price and October 16 availability; TechCrunch reports a $2,600 entry model. The sources differ by one dollar on the entry price. For the Surface RTX Spark Dev Box, GeekWire reports $5,999 and TechCrunch $6,000; Microsoft says U.S. customer shipments begin in November. Those price differences are preserved rather than treated as separate confirmed configurations. GeekWire’s pricing report and TechCrunch’s device coverage provide the figures.
The Surface Laptop Ultra is available for preorder starting October 16.
Microsoft claims up to 2.1 times faster time to first token and up to 4.3 times faster AI image generation against a 16-inch MacBook Pro with an M5 Pro chip. These are attributed, workload-specific results, not evidence of a universal performance advantage. The supplied evidence does not establish sustained throughput, battery life or performance under concurrent agent workloads.
At the larger end, Microsoft and NVIDIA previewed DGX Station for Windows, with availability expected later this year. The announced system uses the GB300 Grace Blackwell Ultra Desktop Superchip, 748GB of coherent memory and up to 20 petaFLOPS of FP4 AI compute. Microsoft says it can support 32 or more simultaneous agents serving a team; that remains a vendor capacity claim. It extends the proposal from a personal laptop to shared local compute, with actual deployment choices still depending on workload measurements and the agent controls available when each feature ships.
