Public transcript

Satya Nadella and Jensen Huang discuss Hybrid Intelligence and the future of PC

By: Windows
Length: 24 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

Satya Nadella and Jensen Huang sit down at a Microsoft event to discuss their decades-long partnership, the new MXC security primitive for agents in Windows, and the shift toward hybrid intelligence that blends local and cloud AI workloads. They trace how Windows enabled Nvidia's founding and how an Azure HPC supercomputer led to training OpenAI's GPT models. The conversation closes with the new Surface Laptop Ultra workstation chip and how trust, identity, and containment will define the future of personal agents.

Key points

Jensen Huang says Nvidia was founded because of Windows in late 1992, and DirectX enabled the GPU revolution that later produced CUDA.
Satya Nadella recalls that an InfiniBand-based HPC supercomputer built in Azure became the machine used to train OpenAI's GPT model.
MXC is described as a fundamental new operating system primitive for agents, providing containment, identity, observability, governance, and token spend tracking.
Hybrid intelligence means users will no longer distinguish between local and cloud compute, with models and tools working seamlessly across both.
Jensen says agents will be better users of tools than humans because agents know every feature of an application while humans know only 10-15%.
The new Surface Laptop Ultra runs on a super chip delivering one petaflop, matching the DGX1 that started the AI revolution, and runs DirectX, OpenGL, and all CUDA applications natively.
The hardware project took over 4,000 engineering years and has been tested with more than 1,200 applications.
On agent liability, Satya suggests insurance will ultimately price the risk of agents acting on a person's behalf, while Jensen stresses trust, containment, and monitoring must come first.

Questions it answers

00:00Origins of the partnership

How did Microsoft and Nvidia's partnership begin?

Jensen explains that Nvidia was founded because of Windows in late 1992, and Windows 95 plus DirectX enabled the GPU in personal computers. The two companies later co-invented the first programmable GPU with programmable shaders, which evolved into CUDA.

Nvidia was founded in 1993 because of the vision of Windows-based PCs.
DirectX and programmable shaders co-invented with Microsoft led to CUDA.
Without CUDA, deep learning pioneers would not have found a computer to train on.

03:32Azure HPC and OpenAI

How did Azure's HPC work connect to OpenAI's GPT training?

Satya recalls building an InfiniBand-based HPC supercomputer in Azure, which later became the machine offered to OpenAI to train a GPT model. He credits Jensen's consistent vision of accelerated computing as a secular trend.

Satya was the first user of InfiniBand inside a cloud.
The HPC work enabled the supercomputer later given to OpenAI.
Jensen's decades-long consistency of vision on accelerated computing is credited by Satya.

06:50MXC and agent security

Why is the operating system the right place to secure agents?

Coding agents running locally with file system access created the need for security at the OS level, which led to MXC. It provides containment at every level, agent identity for observability and governance, and tracking of token spend.

MXC is a new OS-level primitive for agent security.
Containment spans microVM, VM, and Windows session levels.
Agent identity enables observability, governance, and token spend tracking.

10:38Hybrid intelligence

What does hybrid intelligence between local and cloud models open up?

Satya compares it to the pre- and post-browser era, where local and cloud computing merged into one seamless experience. A demo showed assets flowing through Comfy UI, Blender, Photoshop, and cloud models like Astra without the user caring where anything ran.

Users should never have to think about where compute runs.
A demo moved assets through Comfy UI, Blender, Photoshop, Unreal, and cloud models seamlessly.
Hybrid intelligence will become pervasive everywhere.

13:57Agents and local tools

Why does tool use need to happen near local data?

Agents are becoming the largest users of file systems and need secure access to local data and tools. A demo organized tax files into a container and used local and cloud models on them, showing the PC as a computer both people and agents use.

Agents will be the largest users of file systems.
Agents know every feature of an application, unlike humans who know 10-15%.
Primitives should be available 24/7 and roam with the user.

16:47The new hardware

What is special about the new Surface Laptop Ultra workstation chip?

Jensen describes a super chip made of two fused dies delivering one petaflop, matching the DGX1 that started the AI revolution. It is the only computer that runs DirectX, OpenGL, and every CUDA application natively, took over 4,000 engineering years, and has been tested with over 1,200 applications.

The super chip delivers one petaflop, matching the original DGX1.
It is the only computer running DirectX, OpenGL, and all CUDA apps natively.
Over 1,200 applications have been tested and optimized.

21:00Trust and personal agents

Who is liable when a personal agent makes a purchase?

Satya says it starts with trust: containment, agent identity separate from the user, and delegation authority. He suggests insurance will ultimately price the risk of agents acting on someone's behalf, while Jensen stresses that MXC-level trust, containment, and monitoring must come first.

Agent identity must be separate from the user identity.
Delegation authority will grow as trust increases.
Insurance-style pricing may govern agent liability in the marketplace.

Notes

Nvidia and Microsoft's shared history

Jensen Huang says Nvidia was founded because of Windows in late 1992; the company started in 1993.
Windows 95 and DirectX (Direct 3D) made the GPU possible in personal computers.
With DX8, Nvidia and Microsoft co-invented the first programmable GPU with programmable shaders, which later evolved into CUDA.
Without CUDA, figures like Alex Krizhevsky, Ilya Sutskever, and others would not have found a computer to do deep learning.

From Azure HPC to OpenAI

Satya Nadella recalls making HPC a bigger thing in Azure; Jensen notes Satya was the first user of InfiniBand inside a cloud.
The InfiniBand-based supercomputer built for HPC became the machine offered to OpenAI to train a GPT model.
Satya credits Jensen's decades-long consistency of vision around accelerated computing as a secular trend.

MXC: a new OS primitive for agents

Coding agents running locally with file system access drove the need for security at the OS level.
MXC provides containment at every level: microVM, VM, and the Windows session.
Agent identity enables observability of the agent's full exhaust, governance through policy, and tracking of token spend.
Jensen calls MXC as foundational for agents as Windows and DirectX were for applications.

Hybrid intelligence

Satya compares the moment to the pre-browser era, when local and cloud computing merged into one seamless experience.
A demo moved assets through Comfy UI, Blender, Photoshop, cloud models like Astra, and Unreal without caring where compute ran.
Satya predicts there will never be a moment where users distinguish local from cloud; hybrid intelligence will be pervasive.

Agents as tool users

Agents will be the largest users of file systems and need secure, heterogeneous access.
A tax demo organized downloads, cloud files, and email into a file system container and used local and cloud models on it.
Agents know every feature of an application, while humans use only 10-15% of features.

The new hardware

The Surface Laptop Ultra's super chip is two giant dies fused together, delivering one petaflop — the same as DGX1, which cost a quarter of a billion dollars and weighed 500 lb.
It is the only computer that runs DirectX, OpenGL, and every CUDA application natively on the device.
The project took over 4,000 engineering years; more than 1,200 applications have been tested, performance tested, and optimized.

Trust and the future of personal agents

Trust starts with containment, agent identity separate from the user, and delegation authority.
Satya suggests insurance will ultimately price the risk of agents acting on a person's behalf.
Jensen stresses that minimum rights, credentials, monitoring, and trust must be secured before agents can act autonomously.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.