Microsoft Build 2025: CEO Nadella Takes Platform, Systems Approach To The ‘Agentic Web’
‘The big winners are going to be people like yourselves who are going to build applications, not just people creating platforms like us,’ Microsoft CEO Satya Nadella says.
Microsoft CEO Satya Nadella said that the artificial intelligence platform shift is “getting into the middle innings” and touted his company’s systems-based strategy of meeting every layer of the AI stack as the winning approach to democratizing the technology.
“We are taking, really, a systems approach, a platform approach–which you can expect from Microsoft–across every layer of the stack,” Nadella said during his keynote address at Microsoft Build 2025. “The big winners are going to be people like yourselves who are going to build applications, not just people creating platforms like us. The winners need to be across every sector of the economy and every part of the world. That’s our goal.”
Nadella shared the Redmond, Wash.-based cloud and AI products vendor’s accomplishments and vision during his keynote address at Microsoft’s annual developer-focused Build conference, which runs through Thursday in Seattle.
[RELATED: Microsoft Build 2025: The Biggest News In AI, Agents, Windows]
Microsoft Build 2025
Microsoft’s GitHub and GitHub Copilot are enabling an open ecosystem for software development in the AI era, Nadella said. Microsoft 365 Copilot, Copilot Studio and Teams are enabling agents in every job role and business process. Foundry is allowing for AI application and agent building with any data. And all of it applies rails, identity management and security needed by enterprises.
Those tools and more in the Microsoft stack will help build what Nadella called an open, scalable agentic web, speaking to AI assistants’ roles as, perhaps, the next main interface for the internet and for doing work.
“You can ask questions and AI assistants give us answers,” Nadella said. “You can assign tasks to agents and have them execute them. Or work side-by-side with AI to complete jobs and projects. And you can mix and match all of these form factors.”
Here’s more of what Nadella had to say during his keynote, edited for length and clarity.
Innovating GitHub Copilot
Visual Studio and family now has over 50 million users. GitHub has 150 million users. GitHub Copilot, in fact, has been used by more than 15 million developers. And we are just getting started.
As GitHub Copilot has evolved inside VS Code, AI has become so central to how we code. And that’s why we are open sourcing Copilot in VS Code.
We will integrate these AI-powered capabilities directly into the core of VS Code, bringing them into the same open-source repo (repository) that powers the world’s most-loved dev tool.
We will continue to build out GitHub Copilot, too. In fact, over the past few years, we’ve gone from code completions to chat to multi-file edits and now agents.
This same pattern is emerging more broadly across the agentic web. You can ask questions and AI assistants give us answers. You can assign tasks to agents and have them execute them. Or work side-by-side with AI to complete jobs and projects. And you can mix and match all of these form factors.
We are building app modernization into agent mode. Copilot now is capable of upgrading frameworks like … .NET 6 to .NET 9 and migrate any on-premise app to the cloud.
It creates a plan for your code, dependencies, suggests the fixes along the way, learns from changes you make and makes the entire process seamless.
New Microsoft AI Agents
Think about one of the pain points for any of us–getting woken up in the night to deal with a live site issue.
The (AI) SRE (site reliability engineering) agent starts automatically triaging, root causing, mitigating the issue, and then it logs the incidence management report as a GitHub issue with all the repair items.
And from there, you can even assign the repair items to GitHub Copilot.
(The full coding agent in GitHub Copilot takes) Copilot from being a pair programmer to a peer programmer.
You can assign issues to Copilot. Bug fixes, new features, code maintenance–and it will complete these tasks autonomously.
It’s setting up a branch. It … creates a virtual machine using GitHub Actions. Commits a draft PR (pull request) to session logs. In fact, you can go back to the session logs and continue to see all the draft PRs as it’s working.
Coding agent respects all the security measures while delivering a great developer experience.
It only uses MCP (Model Context Protocol) servers configured by a developer. We can get other agents to do code reviews and keep people in the loop before they run any CI/CD (continuous integration/continuous delivery) or merge.
Microsoft 365 Copilot Updates
I don’t think since Teams launched, we have had an update of this level. It really brings together chat, search, notebooks, create and agents all into this one scaffolding that’s intuitive.
I always say it is the UI (user interface) for AI. And chat, for example, is grounded both on web data as well as your work data. That’s the game changer, especially with (Copilot) pages.
Search works across all of your applications, whether it’s Confluence or Google Drive or Jira or ServiceNow. Not just M365 data. With notebooks, I can now create these heterogeneous collections of data. Right. In fact, I can have chats and pages and any documents, emails, all in that collection. I can get all these audio reviews or podcasts out of it.
I can use create to turn a PowerPoint into a new explainer video or generate an image. And when it comes to agents, we have a couple of special agents. Like researcher.
It’s been perhaps the biggest game changer for me because it’s synthesizing across both the web and enterprise sources, applying deep chain of thought reasoning to any topic or any project.
Analyst goes from raw data across multiple source files–I can just upload a bunch of excel files. It will get insights, it will do forecasts, it will do all the visualizations. These agents are all, at the end, about putting expertise at your fingertips.
We are at that age where we are going to put expertise at your fingertips.
Copilot Tuning
We are introducing a new class of enterprise-grade agents you can build using models fine-tuned on your company’s data, workflows and style. We call it Copilot Tuning.
It’s about tuning Copilot for every customer, every enterprise, every firm. Copilot can now learn your company’s unique tone and language. And soon it will even go further, understanding all of
the company-specific expertise and knowledge. All you need to do is seed the training environment with a small set of references and kick off a training run.
The customized model inherits the permissions of all the source control and once integrated into the agent, it can deploy to authorized users.
If you’re a legal firm, it will reason through past arguments and relevant literature and deliver answers and generate docs that are very specific to your firm. Or if you’re a consulting company that works across, let’s say, multiple vertical industries, you can now start to tune these models for each vertical industry to reflect the specific knowhow you know about the workflows in that industry.
It’s all about taking that expertise that you have as a firm and further amplifying it.
You can now think about these agents and multiagent frameworks orchestrating the workflows in an agentic way for every role, every business process.
Foundry: Intelligence’s Production Line
As models evolve faster and become more capable with new samples being dropped every couple of months, the apps will have to evolve to become these full, stateful applications that are multimodel and multiagent. That’s the big departure now. It’s not about one model with just a request response API call. We are building real stateful multi-model applications. And they have to be production ready.
That’s what is the motivation for building a first-class app server. Think of Foundry like a production line for intelligence. It takes more than a great model to build these agents and applications. The system around the model–whether they are evals, this orchestration layer or RAG (retrieval-augmented generation)–all really, really matter.
Foundry is that complete app platform for the AI age. Over 70,000 organizations are already using it across industries.
Enterprises are moving from doing POCs (proofs of concept) to enterprisewide deployments to unlock the ROI (return on investment) of AI. Over the past three months, we have processed over 100 trillion tokens, which is 5x year over year.
We already support 1,900 models, whether they are response models, reasoning models, task-specific, multimodal, you name it.
Picking a model can be a chore. You need to route your queries to the right one fast. We are making that easier, too. Our new model router will automatically choose the best OpenAI model for the job. No more of those manual model selections.
You can provision throughput once on Foundry and you can use that provision throughput across multiple models including Grok (by Elon Musk’s xAI).
You now have Mistral, which you can even provision with all the sovereign deployment in the EU (European Union) region, which increasingly becomes a massive consideration for people building applications around the world. I think increasingly there will be models that people prefer in different parts.
The Foundry Agent Service lets you build declarative agents, in fact, with just a few lines of code just in the portal. For complex workflows, it supports multi-agent orchestration. You can use the agent service essentially like a managed service. More than 10,000 organizations are already using it.
We are providing a full spectrum of compute so that you can hit that right price performance for any of your agentic scenarios.
We are making it straightforward, for example, for you to connect Foundry to your container app or functions and deploy any open source model into AKS, whether it’s in the cloud or in hybrid mode with Arc. Increasingly, you want to have models that get deployed at the edge. Foundry will support that. We are closing the gap and the loop between Foundry and Copilot Studio.
You can now take a model … post-train it in Foundry and then drop it right into Copilot Studio so that you can now use that post-trained model to automate a workflow or build an agent.
Security In The Agent Era
The other important consideration for an app server is observability. That’s why we now have new observability features coming to Foundry to help you monitor and manage AI in production. You can track the impact, quality, safety as well as cost all in one place.
In the future, we believe every organization is going to have people and agents working together. That means the systems that you are, today, using ubiquitously–for things like identity, management of endpoints, security–will now extend to agents as well. That’s a big deal. You want the same rails that you use today at scale to work across people and agents.
With Entra ID, agents now get their own identity, permissions, policies, access controls. The agents you build in Foundry and Copilot Studio show up automatically in an agent directory in Entra. We’re also partnering with ServiceNow and Workday to bring automated provisioning and management to their agents via Entra.
When it comes to data governance, (Microsoft) Purview now integrates with Foundry. When you write an agent, automatically because of Purview, you can ensure end-to-end data production. Another massive safety consideration.
On the security side, Defender now integrates with Foundry. That means your agents are also protected just like an endpoint would be from threats like wallet abuse or credential theft.
Local AI Development
(Foundry Local) includes a fast, high performance run time, models, agents-as-a-service and a CLI (command-line interface) for local app development. And, yes, it’s fully supported on Windows and the Mac.
If you’re a developer, Windows is the most open platform with massive scale with over 1 billion users and 1 billion devices that you can reach.
In the past year, we have seen developers from Adobe to Zoom use the on-device AI capabilities on Windows to ship some amazing applications.
The Windows AI Foundry is what we used, in fact, ourselves internally to build features on Copilot+ PCs for things like Recall or even Click to Do. All of these now are built using the same runtime and the SDK. Now we’re extending this platform to support the full dev life cycle. Not just on Copilot PCs, but across CPUs, GPUs, NPUs and in the cloud. So you can build your application and have them run across all of that silicon.
Foundry Local is built into Windows AI Foundry so you can tap into this rich catalog of these pre-optimized open source models that you can run locally on your device.
With Windows AI Foundry you can customize … our built-in Phi Silica SLM (small language model) using LoRA (low-rank adapters) to basically meet any specific need of your application.
If (Microsoft-backed OpenAI’s) o1 and DeepSeek marked the beginning of, basically, inference or test-time compute in the cloud, I think Phi Silica is going to completely revolutionize what inference compute on the PC is going to look like. You all as developers are going to truly exploit that to build some amazing experiences.
Windows will include several built-in MCP servers like file systems, settings, app actions.
We are adding (a) native MCP registry that lets MCP-compatible clients discover secure MCP servers that have been vetted by us for security performance all while keeping you in control.
The NLWeb Project
(NLWeb) democratizes both the creation of intelligence–for any app, any website … it democratizes the aggregation of this intelligence for any developer.
Give me a reasoning model on NLWeb and I can take an intent and start composing using the reasoning model to go compose and synthesize across that distributed intelligence.
What is search, what is a feed–any of these things gets completely changed in terms of how one goes about building it. This is a platform that we want to create together. I think something big is going to shake out of this.
The last 10, 15 years, I know, have always been about aggregator power. I think something big is going to shift.
Natural Language In Data Analysis
We are integrating (Azure) Cosmos DB directly into Foundry. Any agent can store and retrieve things like conversational history. Soon they will be able to use Cosmos for all their RAG application needs.
Now inside a PostgreSQL query, you can have LLM responses directly integrated. You can have natural language and SQL together.
(Fabric) is at the heart of our data and analytics stack. Fabric brings together all your data, all your workloads together into this one unified experience. Last fall, we put SQL into Fabric.
We are bringing Cosmos DB to Fabric, too. AI apps need more than just structured data. They need semi-structured data, whether it is text, images, audio.
With Cosmos and Fabric and your data instantly available alongside SQL, you can unify your entire data estate and make it ready for AI.
We are even building our digital twin builder right into Fabric. Now you can very easily take digital twins with no code, lo code.
You can map the data from your physical assets and systems super fast.
(Copilot in Power BI) allows you to chat with all of your data. You can ask questions about data, visually explore, analyze it across multiple Power BI reports and semantic models.
This agent will be available in Microsoft 365 Copilot. The power of all the work you did–building out the semantic models, building out those dashboards in Power BI–and now being able to put reasoning models on top of it in a chat interface, think about what a game changer that will be.’-----
Microsoft’s Complete AI System
As a developer you face that classic optimization problem between delivering the best AI experience in terms of performance and latency–and then of course you have to deliver it with monotonically decreasing costs.
That’s why we are taking a systems approach working across the entire industry to optimize the full stack–whether it’s the data center, silicon, system software, or the app server. Bringing it all together as one system and optimizing and using the power of software.
Our aim is to offer the lowest cost, highest scale infrastructure to build both cloud and (the) next generation of AI workloads. It all comes down to delivering the most tokens per watt per dollar. That’s sort of the ultimate equation for us.
The largest GB200-based supercomputer is going to be Azure. We’re excited about scaling this and making it available to all of you as developers.
We now have 70-plus data center regions. More than any other provider. Over the past three months alone we have opened 10 data centers across countries and continents.
We are building a complete AI system. That means it includes cooling to meet the demands of AI workloads. With Maia, we have introduced and engineered this sidekick liquid cooling unit that also supports GB200 in this closed loop fashion so that you can really consume zero water.
On the network side, our newest data centers built for AI have more fiber optics than we added in all of Azure before last year.
We are connecting our DCs with this AI WAN 400-terabyte backbone.
Whether it’s inference or training needs, when you distribute them, the AI WAN connects this data center footprint.
Last October, we launched Cobalt, our Arm-based VMs. They are already powering a lot of our own workloads with massive scale. Teams runs on it. Defender runs on it.