Microsoft wiring AI models into Excel to turn it into infrastructure for future agents: Satya Nadella

Microsoft CEO Satya Nadella has explained how building AI models directly into Excel could turn the spreadsheet software into a key platform for future autonomous AI agents. Speaking on a podcast with CEO of SemiAnalysis, Dylan Patel, and host Dwarkesh Patel, he said Excel is moving beyond its role as a standard productivity tool and becoming part of the core system that AI will use to carry out complex tasks on its own.

During the conversation, Nadella also shared an exclusive preview of Microsoft’s new Fairwater 2 datacentre. He described it as an important part of Microsoft’s plan to create the computing power needed for AGI across all parts of the company’s technology stack.

He said Microsoft is building several Fairwater centres, each containing hundreds of thousands of GB200s and GB300s. When the buildings are connected, they will provide more than 2 GW of capacity. According to Nadella, even one Fairwater building is stronger than any AI datacentre currently operating.

More from Tech

The discussion then turned to how companies can succeed in the AI model race. Dylan Patel pointed out two viewpoints: one that says the tools and scaffolding around models will decide who wins, and another that says the models themselves are the real source of value. He noted that OpenAI’s revenue rose sharply once it launched a competitive code model, and that Anthropic’s inference margins grew from well below 40% to over 60%, even with more competitors entering the market.

Nadella said the idea that companies can simply “wrap a model” and build a business is no longer true because models have become far more capable. He used Excel Agent as an example. Instead of putting AI only at the user interface level, Microsoft is placing AI models in the middle layer of Office 365. With GPT-family technology, the AI is being taught to understand Excel’s internal structure, identify formulas, correct mistakes and work directly with Excel’s built-in tools. Nadella compared this to giving the AI a detailed guide on how advanced Excel users think and work. In this setup, Excel effectively comes with an AI analyst built into it.

Microsoft’s tools become the base for autonomous agents

Nadella said that strong competition among model providers will let developers switch between models if prices rise. He compared this to how large cloud companies compete, with open-source models acting as an additional balancing force. Microsoft plans to stay in the hyperscale business and continue supporting several models. It will keep access to OpenAI models for seven more years, build its own through MAI, and create model-focused application layers for security, knowledge work, coding and science. These applications, he said, will have models integrated deeply into their structure rather than attached as simple add-ons.

Microsoft building AI analyst

Dwarkesh Patel asked whether future AI models will eventually operate computers just as well as human workers, which could make such deep integration unnecessary. Nadella replied that Microsoft is building an AI analyst that already knows how to use the tools people rely on. He said the company imagines two futures: one where humans still guide tools like Excel, and another where organisations assign computing resources to fully autonomous AI agents. These agents will have built-in access to the same tools and will use them more efficiently.

Nadella finished by saying Microsoft’s current business of providing tools for human users will shift into an infrastructure layer for agents that carry out work on their own. The systems built under M365 — storage, archiving, search and management — will still be needed, even when AI becomes the main user. He described this as the creation of a new kind of infrastructure.