‘We’re building the OS for hybrid intelligence,’ says Pavan Davuluri, Microsoft’s executive vice president of Windows and devices.
Microsoft has opened pre-orders for new Copilot+ PCs, with shipping set to start Oct. 16, pledged to start shipping Surface RTX Spark Dev Box devices to U.S. customers in November and expects the first batch of Windows devices built with Nvidia’s DGX Station later this year.
Alongside showcasing the new device lineup during an online event Wednesday, the Redmond, Wash.-based executives said to expect Copilot “hybrid intelligence” features leveraging local data, local models and cloud-based resources “in the coming months.”
“We’re building the OS for hybrid intelligence,” Pavan Davuluri, Microsoft’s executive vice president of Windows and devices, said during the event. “It provides the most secure platform for agents, unlocks unmetered intelligence. And we’re bringing AI into the ambition of humans and their flow. And, of course, we’re delivering it all on PCs that bring it together.”
Surface Laptop Ultra starts at $2,599 manufacturer’s suggested retail price from Microsoft.com or select retailers worldwide. Surface RTX Spark Dev Box starts at $5,999 MSRP exclusively on Microsoft.com in the U.S.
[RELATED: Logitech Enterprise Products Reveal: 5 Things To Know]
Peter Rodenhauser, president and chief revenue officer at Fort Wayne, Ind.-based Microsoft partner Corsica Technologies—No. 440 2026 Solution Provider 500—told CRN in a recent interview that Copilot has been a major opportunity for the solution provider and the recent integrated security operations center (SOC) offering with Microsoft Sentinel was a key move by the vendor.
“When we’re talking about deploying agents, a good majority of the work we’ve done to date has been around Copilot,” he said. “I have an appreciation for how different enterprise businesses operate from mid-market. … What it means is the products and offerings need to be tailored to meet those demands.”
Corsica has sought to stand out from others in Microsoft’s 500,000-member partner ecosystem through bringing clients, especially the midmarket, more simplicity in adopting Microsoft’s latest technologies.
“Enterprises can handle a much higher degree of complexity and scale than mid-market organizations,” Rodenhauser said. “That’s the balance and the challenge that I see for vendors like Microsoft and the others. Not an easy one to solve, but I think they do a nice job of it.”
Windows As The OS For Agents
Davuluri explained that Microsoft is moving beyond operating systems built for applications to one built for agents working independently of humans for hours at a time.
Microsoft has invested in new primitives for agent-minded OS, giving agents flexible containment, distinct identities and separate permissions from the user.
“We are building these primitives directly into Windows, so agents can be secured, observed and governed, making Windows the most secure platform for agents,” Davuluri said.
New Nvidia-Powered Copilot+ PCs Ship Oct. 16
The new line of PCs available for pre-order are powered by Nvidia’s RTX Spark. They include Microsoft’s Surface Laptop Pro and devices from Dell Technologies, HP, Lenovo, Asus and MSI.
Users can also pre-order the Surface RTX Spark Dev Box aimed at developers and AI engineers who evaluate AI models, iterate locally and need local AI compute. The Dev Box will ship to U.S. customers starting in November.
Microsoft bills the Surface Laptop Ultra as its most powerful Surface device with 128 gigabytes of unified memory and the ability to run AI models exceeding 120 billion parameters locally.
The vendor has also taken shots at Apple’s M5 Pro processor, saying that with Windows PCs with RTX Spark, the PCs see about twice the speed to first token, about four times the speed in AI image generation and about six times the speed in AI video generation.
Microsoft has launched a promotional campaign giving Surface Laptop Ultra buyers up to $1,000 cash back when they trade in an eligible MacBook Pro. The offer is valid with qualifying new Ultra purchases until Nov. 23. Buyers must submit the online trade-in claim for their used device, plus proof of purchase for new device, according to Microsoft.
Over 40 percent of the laptops being built for business are Copilot+ PCs, comprising tens of millions of PCs a year, Davuluri said. Microsoft has seen more than 2 trillion inferences per month running locally to improve search, imaging, video calls and other experiences.
MAI Code 1.1 Flash
Microsoft said that devices will use 3-bit precision to reduce the model size of its Microsoft AI (MAI) Code 1.1 Flash model by about 80 percent.
The reduced Code 1.1 Flash still promises coding quality and support for a 256K context window locally.
Device users will also have access to Nvidia’s Nemotron model, which will offer more than 70 billion parameters quantized in 2-bit to use just over 20GB of memory, according to Microsoft.
The 284B parameter model DeepSeek V4 Flash will also be optimized for speed and performance for running locally on RTX Spark, according to Microsoft, with other models expected to receive this treatment in the future.
DGX Station Expands Local AI Compute Capabilities
Later this year, the Dell Pro Precision and HP ZGX Fury AI Station will make up part of the Windows devices built with Nvidia’s DGX Station that promise to enable frontier-class models and agentic workloads on devices instead of just through the cloud or data centers.
Microsoft bills these devices as bringing users last quarter’s frontier capabilities and the RTX Spark devices as bringing last year’s frontier capabilities. Users will have access to powerful capabilities in a deskside computer without needing to rent a cluster.
These DGX Station devices will leverage Nvidia’s GB300 Grace Blackwell Ultra desktop superchip and deliver data-center-class performance with enterprise-grade security, according to Microsoft.
Execution Containers Aim To Make Windows Safer For Agents
Today’s sandbox environments don’t sufficiently secure agentic workloads, Microsoft’s Davuluri said. Agents need containment to change with every tool call and every prompt.
At the heart of Microsoft’s architecture strategy for making Windows devices the safest for AI agent use are Microsoft Execution Containers, now in general availability on Windows 11.
MXCs gives AI agents a runtime composable sandbox where organizations can set policies for which files and networks an agent can access. Windows enforces those policies at runtime, Davuluri said. “It gives IT confidence that agents can only access what is approved,” he said.
Microsoft plans to “soon” make Intune policies available for managing MXC process containers used by MXC-integrated agents on Windows 11, giving IT administrators control over how Windows evaluates container creation requests made by agents.
With MXC, developers can contain model-generated output, plugins, tools, agent harness or the entire agent, according to Microsoft. If something goes wrong, the impact scope is limited.
Policies stay outside the agent workload’s control, preventing the agent or generated code from granting itself additional access—a concern likely on users’ minds with the multiple publicly disclosed rogue agent hacks.
MXC works with other OSes, but Microsoft executives Wednesday said that using the containers with Windows gives greater control over data and more containment options in process isolation, session isolation, Windows Subsystem for Linux containers (WSLc), virtual machines (VMs) and Windows 365 for Agents.
W365 for Agents is a specialized cloud platform that provides virtual cloud PCs that run Windows or Linux and are designed for autonomous AI agents instead of human users.
MXC coupled with Microsoft’s security stack will also help separate agent activity from user activity for improved observability. Microsoft Defender, for example, can detect and block attempted malicious actions by agents. And Windows can restrict access for the agent, not the user.
“Whether an agent is working under your direction or acting on its own, every action it takes should be attributed to the agent, not to you,” Davuluri said. “Only Windows enforces this attribution strongly at the operating system level. That attribution is what makes observability possible.”
Nadella added that assigning agents identities brings peace of mind to IT teams and cybersecurity teams, he said. And tracking token spend should help financial operations (FinOps) teams.
“Having seen all these adoption cycles, whether it’s the consumer adoption cycle, the enterprise adoption cycle, I would say the thing that blocks broad diffusion is not having that complete thought on what is required,” Nadella said. “If you do one and leave the other, there’s no trust and what have you. So I feel very, very good that they’ve brought it together in the cloud, in the the client, across IT, sec and FinOps. And it’s heterogeneous, right? It’s not about any one piece of Microsoft technology, but supporting it all.”
‘Hybrid Intelligence’ A Way To Lower Costs, Boost Performance
Following on the news from September that Microsoft is rebuilding the Copilot experience, Microsoft executives said Wednesday that Copilot on Copilot+ PCs will offer users more power through Windows hybrid intelligence.
Microsoft expects to start rolling out Copilot hybrid intelligence features in the coming months, Jacob Andreou, Microsoft’s recently promoted executive vice president over the unified Microsoft Copilot team, said during the event.
Users may need a Windows update to use those features, with feature availability and timing possibly vary by device, market and silicon platform, according to Microsoft.
Through hybrid intelligence, Copilot can leverage local context, files, and recent activity from the user’s PC. Copilot can organize files, assess device diagnostics, troubleshoot issues, code, carry out workflows and perform other tasks locally for users.
Copilot can also access local AI models running on the PC, using local compute for certain tasks and tapping into the cloud when it must, according to Microsoft.
Microsoft has also extended its GitHub HydraFusion task routing capability to Windows, allowing HydraFusion to use models running locally on devices.
HydraFusion is set to come to the GitHub Copilot application, command-line interface (CLI) and Visual Studio (VS) Code Oct. 15. This capability routes agent tasks to appropriate models to Windows, allowing HydraFusion to tap into models running locally, Davuluri said.
“When I talk to customers and partners, for scaling AI, they often ask, ‘How do I get the most of my token budgets without giving up frontier capability?’” Davuluri said. “With hybrid intelligence, there’s no need to choose. We’re moving toward a future where hybrid intelligence is always available, running locally where it makes sense, reaching to the cloud where it needs to, and seamlessly bringing the best of both.”
Windows Search Upgrades
Microsoft has started rolling out new Windows Search capabilities to members of its Windows Insider program in the experimental channel and plans to introduce an opt-in Copilot and Search integration to select markets later this year.
The goal is for a Search with better performance, speed and memory usage, according to Microsoft. The user interface (UI) is now in a single list for quicker scanning. Search should better understand what users are looking for through better typo and synonym matching, which Microsoft will fine-tune “over the coming weeks.”
Inline previews and richer file details will come later this year to help users differentiate similar results, according to Microsoft.
The new Search, built on WinUI 3, promises a faster experience, more efficiency and thousands of new actions available through the taskbar.
Users can switch to dark mode, enter airplane mode, turn on “do not disturb,” minimize all windows, arrange windows, dim screens, mute themselves, text people and perform other tasks without opening another app or screen, according to Microsoft.
For those who decide to go with the integrated Copilot and taskbar Search, users can receive quick inline answers from Search, jumping to the full Copilot application for follow-ups and more complex prompts.
