Hybrid AI Set the Terms
Microsoft opened its Windows and Surface event in San Francisco on October 7 with Satya Nadella, Microsoft CEO, and Pavan Davuluri, executive vice president of Windows and Devices, laying out the case for hybrid AI. Many in the room were there for the Surface Laptop Ultra. The opening framing explains why that device, and the RTX Spark laptops from Microsoft's partners, need to exist.
The argument for splitting AI work between the device and the cloud usually rests on privacy and economics: keep sensitive data local and stop paying for cloud inference on every task. Nadella and Davuluri added a third point. Hybrid will become the default workflow for users and agents alike, with work moving between the PC and the cloud depending on what each task needs. Microsoft calls this hybrid intelligence: a local model handles the work suited to the PC, and a cloud model adds capability when the task needs it. GitHub is building the idea into developer tools with HydraFusion, which routes work between local and cloud models in the GitHub Copilot app, Copilot CLI and Visual Studio Code, with an experimental preview due later in October. Accept that premise and the new hardware has a clear job. It is where the local half of that workflow runs.
Trust Starts at the Operating System
The second theme of the day was trust, and Microsoft approached it from two directions. The technical one is Microsoft Execution Containers, or MXC, which the company introduced at Build in June. MXC isolates agents from the user's desktop, clipboard and input devices, ties each agent to its own local or Entra-backed identity, and enforces the file and network access a developer declares before the agent acts. MXC became generally available on Windows 11 today, so organizations can now set which files and networks an agent can reach and have Windows enforce those limits while it runs. Microsoft also showed MXC integration for OpenClaw as part of a simpler setup for popular agents.
In the fireside chat with Nadella that closed the event, moderated by Sriram Krishnan, Jensen Huang, Nvidia CEO, called MXC the cornerstone of the agentic era. His endorsement carries weight because Nvidia silicon now sits inside Windows PCs built to run agents locally. A sandbox that IT departments can manage with the identity and policy tools they already use is what makes a powerful local machine one an enterprise will let agents operate on.
Trusting the Agent Itself
The second direction is harder to engineer. Users have to trust the agent they are running before they will let it act for them. That trust has three parts: confidence the agent can do the job, confidence it followed instructions correctly, and confidence the data it touched stayed private and secure.
Delegation between people works the same way. A manager gives a new hire small tasks, checks the output, and hands over more as the results hold up, until checking stops feeling necessary. Agents will earn autonomy on that curve. Each task done right lowers the bar for the next one, and each visible mistake raises it. MXC covers the security part. Capability and accuracy get earned one task at a time, which makes the earliest experiences disproportionately important.
Meeting Consumers Where They Are
Trust matters for consumers too, and most of them are still uneasy about agents. Microsoft is meeting them at that level with the new search bar in Windows, which can now take actions such as sending a message or changing a PC setting. Search actions start rolling out today to Windows Insiders in the Experimental channel, with examples including switching to dark mode, turning on Do Not Disturb and arranging windows. Optional Copilot integration in taskbar Search follows in select markets later this year.
Small, low-stakes tasks are the right place to start. Every time AI gets a simple action right, it builds the confidence consumers will need before they consider agents for areas such as health and finances. Those are also the areas where many people will want the work done locally, which brings the argument back to hybrid AI and to the hardware that makes local processing practical.
A Workstation for Local AI
The Surface Laptop Ultra starts at $2,599 with Nvidia's RTX Spark platform, an 18-core N1X CPU, a 5,120-core Blackwell GPU, 24GB of unified memory and 512GB of storage. Higher configurations step up to a 20-core Grace CPU and 6,144 GPU cores, and Microsoft says the machine can run local agent workflows on models of up to 120 billion parameters. The 15-inch PixelSense Ultra touchscreen reaches 2,000 nits of peak HDR brightness, the chassis is under 18mm thin and 4.5 pounds, and the SSD is removable and upgradeable to 2TB. The top configuration, with 128GB of memory and a 1TB SSD, reaches $5,899.99. Preorders opened today and units ship October 16, alongside RTX Spark laptops from ASUS, Dell, HP, Lenovo and MSI. The Surface RTX Spark Dev Box, a compact desktop with 128GB of unified memory that ships with Visual Studio Code, GitHub Copilot, WSL and other developer tools set up, follows in November at $5,999.
These prices put the category out of reach for most PC buyers. But, honestly, they are not for everyone. They are AI workstation. Sure The Surface Laptop Ultra can handle gaming well, but its primary role is as the home for local AI workloads, with up to 128GB of unified memory available to large models running on the device. That is the configuration Apple has owned with the MacBook Pro. Microsoft names the target outright: its launch promotion offers up to $1,000 back on a qualifying trade-in through November 23, marketed as breaking up with your MacBook. Windows buyers who want unified memory and serious on-device AI performance in a laptop now have a credible alternative, from six brands.
Reinventing the PC
Huang closed the conversation by telling Nadella, Davuluri and the Microsoft team that together they have reinvented the computer as we know it. Anyone judging RTX Spark machines on whether every Windows app runs natively on Arm is applying the test of the PC we have used for decades. The reinvention Huang described is about agents, and agents will get work done in a different way from how we move between apps today.
Huang also pointed out that agents can draw on every feature these devices have, while people are limited to the features they know about. Most of us use a fraction of what our hardware can do. Running on a machine like the Surface Laptop Ultra, an agent will extract far more from it than its owner ever could.