Interview, Fireside Chat
Satya Nadella – How Microsoft thinks about AGI
- Microsoft intends to pivot from an end-user tools model to an infrastructure business designed to support autonomous agents, provisioning computers, security, and identity on a per-agent basis rather than per-user.
- Training capacity is projected to increase 10x every 18 to 24 months, with future models potentially requiring two whole regions and aggregated flops across sites like the Fairwaters campus and the Wisconsin data center to run single jobs.
- The Fairwater 2 data center will deploy network optics nearly equal to the total count of all Azure data centers from two and a half years ago, while Fairwater 4 under construction nearby will connect to it via a one petabit network.
- Microsoft plans to build a flexible campus capable of running training jobs, pods, and super pods, avoiding optimization for a single model specification to prevent obsolescence from MOE-like breakthroughs or new chip requirements like the Vera Rubin Ultra.
- A "speed of light execution" partnership with NVIDIA is planned to optimize Total Cost of Ownership (TCO), alongside internal silicon development using Maia models for cost-optimized or latency-friendly tasks, with a goal of achieving high researcher-to-GPU ratios.
- The company expects to utilize the OpenAI GPT family for the next seven years and access all OpenAI IP rights except consumer hardware, while reserving the OpenAI stateless API exclusively for Azure and launching an omni model combining audio, image, and text domains on the Maia side.
- Microsoft anticipates the AI market for coding and software factory categories could exceed knowledge work, with growth rates of 10% to 20% and GitHub expecting one new developer per second joining, 80% of whom will use Copilot workflows.
- Plans include building an "Agent HQ" to function as a central control point for agents from various providers, creating a hybrid environment where human and agent tools communicate artifacts and observability layers, alongside databases that handle structured and unstructured data joins.
- Hyperscalers are expected to spend $500 billion on CapEx next year, a scale unmatched by previous revolutions, while true economic growth from AI is predicted to take 20 to 25 years to diffuse, compressing the Industrial Revolution's 200-year transformation.
- Infrastructure growth will focus on gigawatts of capacity rationalized by demand and geotype, with strategies to buy capacity through leasing and sites as a service rather than building everything, and welcoming Neoclouds into the Azure marketplace.
- Microsoft aims to build a world-class superintelligence team to pursue breakthroughs over the next five to eight years and expects to continue building a hyperscale fleet supporting multiple model families, including open source and OpenAI models.
- Sovereignty requirements are expected to drive global adoption, leading to the expansion of sovereign clouds and services in France, Germany, and the EU to navigate a bipolar geopolitical landscape involving the US and China.
- Microsoft expects to grow from 20 to 26 million subscriptions in the last quarter and views the industry structure as forcing specialization, where it competes at each layer by merit rather than vertically owning every category.
- The company views trust in American tech and institutions as a critical competitive advantage, planning to remain a knowledge-intensive business that optimizes throughput in tokens per dollar per watt through software capability.