Satya Nadella and Jensen Huang: Hybrid Intelligence and the Agent PC

Satya Nadella and Jensen Huang shared a stage at Microsoft’s October 2026 Windows event to explain hybrid intelligence, the idea that AI agents on a PC should move between local and cloud models without the user noticing. The 24-minute conversation, published on the Windows channel and moderated by former White House AI policy advisor Sriram Krishnan, covers the new agent sandbox in Windows called MXC, the Nvidia chip inside the Surface Laptop Ultra, and who is liable when an agent spends your money.

TLDW

Huang says Nvidia exists because of Windows, and Nadella says an early bet on high performance computing in Azure is the accident that gave OpenAI its first supercomputer. Both argue that the PC now has two users, the person and the person’s agents, and that agents need the same primitives people do: a file system, an identity, and limits. Microsoft’s answer is MXC, a containment layer in Windows that Nvidia’s OpenShell plugs into, and Nvidia’s answer is a single chip that runs DirectX, OpenGL and CUDA natively on one machine. Asked who is responsible when a personal agent makes a purchase, Nadella gives the most interesting answer of the session: insurance will price it.

Thoughts

Nadella’s claim that insurance will price agent delegation deserves more attention than everything said about hardware. Krishnan asked the standard question (is it standards, is it protocols) and Nadella waved most of it away. Authentication is not the hard part, he said, and the protocols are good enough to adopt. What remains is a risk question: how much authority did I hand over, and what does it cost when the agent gets it wrong? That is how every other form of delegated authority got sorted out. Corporate cards, power of attorney and employee fraud are all handled by a mix of limits, audit trails and somebody underwriting the loss. If he is right, the companies that matter in consumer agents will include whoever can measure an agent’s error rate well enough to write a policy against it. It also explains why Microsoft cares so much about a separate agent identity. You cannot insure what you cannot attribute.

Huang’s line that people know 10 to 15 percent of an application’s features while an agent knows all of them is the real argument for the local PC, and it is a better one than speed or privacy. Blender, Unreal, Photoshop and the big engineering suites are decades of accumulated capability that almost nobody uses fully. An agent that can drive every menu turns that unused depth into output. The catch is that those tools live on Windows, are licensed per machine, and often depend on a GPU. So the value of the desktop shifts from being where you work to being where your agent’s tools are installed. Nadella’s closing remark about never leaving a piece of code behind stops being a boast about backward compatibility and becomes the moat. An agent platform with thirty years of software to operate is harder to copy than a model.

Nadella’s most reusable sentence is that adoption stalls when the offer is not a complete thought. Containment without identity cannot be audited, identity without policy cannot be governed, and none of it gets past a finance team that cannot see the token bill. He listed all of those as one package. It is worth holding him to that standard. Trade coverage of the launch, citing Microsoft’s own developer blog, says the process, session and WSL containers are generally available, while hardware-isolated microVMs are still experimental and the Intune, Entra and Agent 365 management pieces are listed as coming soon. By his own test, the enterprise version of this is not finished yet. Developers can use it now. IT departments are being shown the plan.

The hybrid pitch borrows its force from a good analogy. Nadella recalls a grad school setup with a terminal into a VAX on one side and DOS programs on the other, and says it took until the browser for those to feel like one thing. His prediction is that nobody will ask whether a task ran locally or in the cloud. That may be true for the user. It will not be true for whoever pays. Local tokens are, as Krishnan quoted from an earlier presenter, free once you own the machine, and cloud tokens are metered. The router that decides which model handles which step is therefore a pricing decision, and both men on stage sell the expensive side as well as the cheap one. Who controls that routing, and whether the user can see it, is the question the session did not reach.

The origin stories are worth a paragraph because they undercut the vision talk around them. Nadella describes the OpenAI supercomputer as happenstance: Azure had built InfiniBand clusters for a small HPC business, Sam Altman asked whether Microsoft had computing, and the answer was “yeah, maybe.” Huang’s chain runs Windows 95, DirectX, programmable shaders, CUDA, deep learning. Neither was a plan for AI. Both were infrastructure built for a narrower market that happened to be the right shape later. That is a more honest account of how platforms win than the keynote framing, and a reasonable way to read this launch too. The agent PC might matter most for something neither company named on stage.

Key Takeaways

Huang says Nvidia was founded because of Windows. In late 1992, with Windows 3.1 on the market, the founders bet that a PC with 3D graphics could do the work of a workstation.
He traces a straight line from Windows 95 to DirectX, to the programmable shaders Nvidia co-developed with Microsoft for DirectX 8, to CUDA, to the deep learning researchers who needed a computer to train on.
Nadella says Azure was the first cloud to use InfiniBand, for a high performance computing business, and that this is the only reason Microsoft had a supercomputer to offer OpenAI.
The agent PC started as a conversation between the two CEOs four years ago about what the PC should be once agents arrive.
MXC (Microsoft Execution Containers) is the agent sandbox built into Windows. Nadella says containment has to exist at every level: the microVM, the virtual machine, and the Windows session.
Containment is one of four parts. The others are an identity for each agent that is separate from the user, observability over everything the agent does, and policy that governs it. He adds tracking of token spend as a fifth concern.
Nvidia’s OpenShell agent runtime is integrated natively with MXC. Huang says MXC will change how agents are built the way DirectX changed how applications were built.
Nadella’s definition of hybrid intelligence: you will never check whether something runs locally or in the cloud, and you will expect it to be everywhere.
The demo both men kept returning to moved assets between ComfyUI, Blender, Photoshop and Unreal using local and cloud models, and cut a job of several days to hours.
Coding agents made the file system and the console fashionable again. The panel expects agents to become the heaviest users of file systems, which is why the storage layer needs the same security.
Huang argues agents will be better tool users than people because people learn 10 to 15 percent of an application and agents know every feature.
Huang says the chip in the Surface Laptop Ultra took more than 4,000 engineering years, delivers one petaflop, and has been tested against more than 1,200 applications. Press reports name the chip RTX Spark and note the petaflop figure is a low-precision AI measure.
On agents that buy things, Nadella says the first decision is how much authority you delegate, that trust grows with autonomy, and that the market will end up pricing the risk through insurance.

Chapters

0:47 Nvidia Was Founded Because of Windows

Krishnan opens by asking Huang for a favorite memory of working with Microsoft. Huang goes back to 1992, when he had never seen a PC and designed chips on workstations and supercomputers. Windows 95 made it possible to connect a GPU to a personal computer, Direct3D exposed it to applications, and DirectX 8 brought the first programmable GPU. He says that without GeForce there would have been no CUDA, and without CUDA the early deep learning researchers would not have had a machine to work on.

3:09 The HPC Accident Behind OpenAI’s First Supercomputer

Nadella credits Huang with the most consistent vision in the industry, a single idea about accelerated computing that outgrew any one form factor or chip design. His own first memory of the partnership is a deal to make high performance computing bigger in Azure. Huang adds that Nadella was the first to put InfiniBand inside a cloud. Nadella says that if Microsoft had not built that cluster, it would have had nothing to give OpenAI to train a GPT model.

5:15 A Four-Year-Old Conversation About the PC

Huang describes a conversation with Nadella about what the perfect PC looks like once an agent sits on it as a personal assistant. The agent wants every tool, but the tools were split between gaming PCs and workstations: Dassault, Siemens, Autodesk, Adobe, Blender, Unreal, Omniverse. The goal was one machine that runs all of them alongside the AI. He says it took four years.

6:32 MXC: Why Agent Security Belongs in the Operating System

Krishnan notes that people now give agents full disk access and let them act as the user, and calls MXC a new operating system primitive. Nadella says the need showed up first with coding agents, where the harness runs locally even when the model is in the cloud. He lists containment at each layer, agent identity, observability, policy and token tracking, and says leaving any one out destroys trust. Huang calls the architecture the foundation of the next generation of IT and says agents are a non-starter without it.

10:07 What Hybrid Intelligence Means

Krishnan describes the shift from one large cloud model to a choice of local and cloud models with routing between them. Nadella compares it to the late 1980s, when he had a terminal into a VAX, a PC at home and Usenet, and it took until the first browsers for those to become one experience. He walks through the creative demo step by step and says people will look back on it as the moment the pieces clicked. Huang says that one workflow drew on nearly every technology company at once.

14:03 Agents, File Systems and Local Tools

The models have become good at tool use, but the tools and the files are local. Nadella says coding agents reminded everyone how strong an abstraction the file system is. He points to a Copilot demo that gathered tax documents from downloads, cloud files and email into one contained folder and then worked on it with local and cloud models. His summary: I use the computer, my agents use the computer, and they need the same primitives around the clock.

16:44 One Chip for DirectX, OpenGL and CUDA

Asked about tokens produced next to you instead of in a data center, Huang describes a chip made of two large dies fused together that unifies what used to be separate architectures for gaming PCs, laptops, workstations and cloud systems. He compares its one petaflop to the DGX-1 that started the AI boom. He says it is the only computer that runs every DirectX, OpenGL and CUDA application natively, and that every laptop, desktop and workstation maker signed on.

20:28 Who Is Liable When an Agent Buys Something

Krishnan raises the debate over personal agents: when an agent makes a purchase, is it you or the agent, and how should a merchant respond when no human shows up on the site? Nadella says it begins with trust, which means containment and an agent identity separate from the user’s. The first decision is what authority you delegate. In the consumer market he expects insurance to price it. Huang says agents should get minimum rights, credentials and monitoring before they are allowed to act.

23:35 “We Have Reinvented the Computer”

Huang closes by saying the two companies have reinvented the computer. Nadella ends on compatibility: Windows has always been about not leaving any code or framework behind, and the new machine carries Windows apps, graphics apps, CUDA and OpenGL together with what he calls the token factory.

Notable Quotes

“If not for Windows, Nvidia wouldn’t have been founded.”
Jensen Huang, on starting the company in the Windows 3.1 era

“Life is full of accidents, but if we had not done that for HPC, we wouldn’t have had the supercomputer to give to OpenAI to go train a GPT model.”
Satya Nadella, on Azure’s early InfiniBand clusters

“The thing that blocks broad diffusion is not having that complete thought on what is required.”
Satya Nadella, on why containment, identity and governance have to ship together

“You will expect this hybrid intelligence to be everywhere and pervasive.”
Satya Nadella, on local and cloud models becoming one experience

“We know 10, 15% of the features of any application, but the agent knows every single feature.”
Jensen Huang, on why agents will use desktop tools better than people do

“I think of it as insurance will price this.”
Satya Nadella, on liability for purchases made by a personal agent

“Windows has been with me for 34 years. Looks like it’s going to be with me for 34 more.”
Satya Nadella, closing the session

Watch the full conversation here.