Jensen Huang & Satya Nadella on unmetered intelligence | Microsoft Build 2026

The PC becomes a personal AI

• The conversation began three years ago between Satya Nadella and Jensen Huang about building a new class of PCs for designers, creators, and AI.
• The PC has evolved from a personal computer to a personal AI — an autonomous agent can run on the PC and be texted remotely to do coding or design work, iterating with the user while they are away.
• RTX Spark delivers a petaflop of AI performance in NVFP4, a numerical format the two companies worked on together, with 128GB of memory that can fit a couple of hundred billion parameter model.
• Windows is also coming to the GB300, described as a data center on your desktop.

Data center evolution

• Microsoft and Nvidia built the first AI supercomputer together to train the GPT models, based on Ampere; Hopper focused on pre-training.
• Grace Blackwell shifted focus to post-training and reinforcement learning, enabling reasoning models based on mixture of experts that are intelligent and energy efficient but require giant systems — hence NVLink72, making the entire rack one computer.
• Microsoft deployed the largest and fastest number of Grace Blackwells in the world. Fairwater is completely liquid cooled, closed loop, uses almost no water, and increased token generation rate while reducing cost by an order of magnitude — some 30 times over Hopper.

Vera Rubin for agents

• Vera Rubin is designed to run agents, the same agentic computing pattern as RTX Spark but processing an enormous number simultaneously from different customers and partners.
• It innovates in confidential computing: storage is long-term memory, working memory is encrypted, and data is encrypted in transit and in use.
• Vera is a revolutionary CPU designed for agents — past CPUs were designed for humans, who are more patient; agents want extremely low latency.
• The two teams aligned long before the chips taped out; systems were stood up at Microsoft as they rolled off the lines.

Developer opportunity

• GitHub commits have gone completely parabolic, increasing by a factor of three in the last several months — evidence that agentic systems are useful and tokens are now profitable.
• All tools on Azure are being fully GPU accelerated — Fabric, data processing, SQL, Spark, semantic, vector, and graph-based — because agents are impatient and faster answers mean faster iteration and more profitable token generation.
• Nvidia models and tooling are coming to Foundry, and Nvidia software will help accelerate workloads including the data warehouse and Windows.