At Microsoft Build, Satya Nadella sat down with Forward Future’s Matthew Berman to explore how AI is transforming software from the ground up.

They cover everything from the collapse of SaaS to the rise of personal agents, the economic implications of zero-cost intelligence, and why Microsoft is redesigning its entire stack for an agent-first future.

00:00 – Intro
Setting the stage: energy usage, intelligence at scale, and the future of the OS.

00:50 – Microsoft All In on AI
Why Azure is becoming an “AI factory”—and what that means for infrastructure.

03:53 – AI Changing Microsoft Products
Microsoft 365 is transforming into an intelligent IDE powered by agents.

05:56 – Is SaaS Dead?
Satya on why the app layer is collapsing and orchestration is the new frontier.

09:24 – Who Owns the Agents?
Governance, identity, and IP in a world where agents do the work.

10:58 – Personal Agents in the Workplace
How Microsoft separates personal and corporate agent environments.

12:18 – Zero Cost Intelligence
When intelligence becomes free, what happens to productivity—and power?

14:32 – AI’s Economic Impact
From oncology to small business: how AI justifies its energy footprint.

17:25 – Future Operating System
Satya explains why OS-level code may soon give way to generative systems

Matthew Berman introduces the core questions around AI’s energy use, the future of intelligence, and rethinking the software stack.

Satya Nadella explains how Microsoft is re-architecting Azure and the entire infrastructure layer to serve AI workloads.

Microsoft 365 enters the agentic era with chat-first UIs, embedded copilots, and a new kind of productivity layer.

Agents may replace traditional software interfaces—so what happens to vertical SaaS?

As agents become part of the workforce, companies must treat them like employees—governing identity, access, and intellectual property.

What happens when employees bring personal agents to work? Satya outlines Microsoft’s approach to separating identities and preventing data leakage.

What happens when intelligence becomes nearly free? Satya shares Microsoft’s vision for how abundance can drive economic growth and innovation.

From oncology to small business productivity, Nadella explores how AI can drive better outcomes while justifying its energy demands.

Satya describes a future OS where deterministic code blends with stochastic models. The key? Guardrails, sandboxing, and traceability.

Matthew Berman:
The energy usage they’re projecting is going to significantly affect our planet. Agents require more energy than any previous workload—but at a different scale. What excites you most about a future where the cost of intelligence approaches zero?

Right now, to tame inflation or improve economic growth, we need help. Do you ever envision a future where the operating system is essentially rebuilt—with little or no traditional underlying code? Every layer of the tech stack would need to be reimagined.

You’ve previously said the software application layer is going to collapse into agents. What does that mean for vertical SaaS companies?

Alright, Satya—thanks for chatting with me, and congratulations on everything you announced at Build. I had a few questions. You've overseen some major transitions at Microsoft: the shift to cloud, the embrace of open-source—and now we’re entering this next phase. With the rise of powerful AI agents, how are you thinking about investing in that future while still maintaining your current product suite and managing the changes ahead?

Satya Nadella:
Yeah. First of all, thanks so much for being at our developer conference. The way I think about it is: first, you have to embrace what’s new. Even though we’re already two to three years into this AI era—depending on how you count—it’s becoming clearer what it means to build agents and apps.

You really have to reevaluate your tech stack. What you built for a previous generation of workloads now needs to be rethought from first principles. Take the infrastructure layer. We’re very proud to have 70 Azure regions around the world, but now we have to retrofit them to function as AI factories.

Take an app like ChatGPT or Copilot. Yes, it requires a lot of GPUs or AI accelerators, but it also needs tons of storage—during both training and inference. It also needs a ton of traditional compute—not AI-accelerated compute—to host agent environments.

So, interestingly, the infrastructure we’ve built over the last 15 years might be more relevant now than ever—because agents require more of it than any previous workload, just at a different scale.

Satya Nadella:
It’s the same story with data. Traditionally, data has been about structured databases—people, places, things. But now, you can bring intelligence directly to the data. A reasoning engine can sit on top of it.

One of the coolest demos we showed was with Postgres. It’s so modular now that you can mix an LLM response into a SQL query. Think about the kinds of query plans you can generate. So I really believe every layer of the tech stack needs to be reimagined. At the same time, it means we can compound on the great work we’ve already done over the past 15 years—so developers can benefit from it.

That’s how we’re thinking about it: reexamining every layer of the tech stack from first principles for new AI workloads, and stitching it together in a way that meets the real-world needs of customers.

Matthew Berman:
Yeah. So for end users—especially those using products like Office 365—things must be changing fast. What does that acceleration look like?

Satya Nadella:
It’s fascinating. If you look at Microsoft 365, I’d say there are three main modes of use.

First is the brand-new mode: a new UI for AI. It’s a new scaffolding that includes chat, search, notebooks—a place where users can collect heterogeneous data and do things like podcasts, audio overviews, and more. We have agents now—researchers, analysts—task-specific agents users can delegate to. And we have Copilot Studio, so users can build their own agents. So that’s the new thing: we now have a UI for AI and agents.

Second, Teams takes all of that into multiplayer mode. All those agents are available in your channel, in your meeting. Teams becomes the scaffolding where AI works collaboratively with you.

And the third mode is what I call "heads-down" mode—just like how in GitHub Copilot with VS Code, you’re coding, but you have agents helping. In Excel, you’re working on a spreadsheet, and Copilot is right there. That’s like having a data scientist beside you. Or when writing, you’ve got a researcher at your side. We’ve effectively turned every Office canvas into an IDE with chat.

So the value of the M365 system has compounded because intelligence is now embedded in every layer.

Matthew Berman:
I want to continue on that note. You previously said the software application layer is going to collapse into agents, and I made a video about it called “SaaS Is Dead.” It got a lot of attention—some great thoughts from folks—but I want to hear your take.

The idea is that we’ll have an agent layer, and beneath it, a grounded database that agents can read from and write to. What does that mean for vertical SaaS companies? How should they prepare for this future?

Satya Nadella:
Yeah, I think the way to approach this—look at the demo we showed today, for example. It was Dynamics 365 using an MCP server with Copilot Studio to orchestrate a multi-agent application. It spanned CRM and other systems of record and completed a complex business process. That’s not hypothetical—it’s already here.

So, yeah, that’s going to require a pretty radical shift. If your whole model is: “I’m the system of record” or “I handle workflows on top of my own data”—that approach isn’t going to persist. We all have to be open to participating in this new orchestration layer of the agentic web.

That layer is going to have multiple backends. Your SaaS app will be one of them. You’ll need to support something like MCP to participate. And maybe something like the NL Web will help reduce friction from all these connectors.

In enterprises, connector friction is a huge issue. Something like the NL Web could be transformative even just for internal systems. So yes, I think SaaS applications as we know them will have to radically evolve to adapt to this future.

Matthew Berman:
So for these SaaS companies—do you see them becoming curators of “ground truth” data for their customers, while platforms like Microsoft provide the agents?

Satya Nadella:
That’s a good question. I’m not sure how it’ll fully shake out. We tend to overstate the importance of the thing we control today. But in a platform shift, value often gets created somewhere else.

At the end of the day, the job to be done is to complete a business process—not to protect a single system of record, or one workflow, or one agent. It’s about the full flow. So the question is: how do you move with that flow, rather than trying to protect a moat or build a facade around it with an agent front-end that doesn’t meet the actual customer need?

Matthew Berman:
I really liked what you said about different types of agents interacting—and how it doesn’t matter which database they’re talking to, because it’s just an abstraction layer. That’s exciting.

In another interview, you mentioned that when you hire someone, you’re essentially hiring their future basket of agents. That really stuck with me. But I wanted to clarify—won’t companies want to own those agents the same way they own traditional intellectual property? How do you see that playing out?

Satya Nadella:
That’s a great point, and actually, you’re exactly right. That’s how we view it too. The intellectual property of a company includes the work product of its employees. That’s going to remain true in the agent era as well.

That’s why we’ve extended our identity and security frameworks to include agents. For example, agents now have an Entry ID. You can manage conditional access for agents just like you would for employees, using tools like Microsoft Purview.

And if agents are accessing sensitive data, they need to be governed by the same data protection policies and access rights. From a security perspective, you’ll manage the agent’s environment like you would any endpoint. That’s why tools like Microsoft Defender are so critical—they ensure there’s no credential theft or security breach from agents.

So everything we’ve built for identity management and endpoint protection in a people-centric IT environment—we’re now extending that to agents and their IT environments.

Matthew Berman:
That makes a lot of sense. I also suspect people will start building personal agents for their own lives. Do you envision a future where people bring those personal agents into work?

Satya Nadella:
Yeah, that’s an important scenario. If someone brings a personal agent into the workplace, we have to make sure there’s no data leakage between personal and corporate environments.

It’s similar to how we treat email today—there’s a clear separation between personal and work accounts. That separation exists for both privacy and intellectual property reasons, and it’s a helpful model to follow here too.

That’s why we’ve built systems like Microsoft Entra for enterprise identity, alongside Microsoft accounts for personal use. For example, I use Microsoft Edge with two profiles—one tied to my personal Microsoft account, and one tied to my work identity via Entra. That separation keeps things clean and manageable.

If we conflate those two worlds, it becomes very easy to get tangled. Maintaining distinct boundaries—both technically and mentally—is going to be critical.

Satya Nadella:
I think that makes a lot of sense. So let me share a bit about my vision. The cost of intelligence really is dropping fast—hopefully approaching zero—and that’s going to unlock an entirely new world.

What use cases am I most excited about? Honestly, anything that drives productivity and economic growth. That’s where this abundance of intelligence can make the biggest difference.

Look around—whether it’s inflation or economic stagnation, we need help. AI can play a huge role. At our developer conference, we showed an example from Stanford Medicine—something high-stakes, like oncology and tumor board meetings.

They used a multi-agent framework inside Microsoft Foundry to orchestrate data across pathology, clinical trials, and other systems. The outcome was a better tumor board discussion—and the insights from that session were automatically shared through PowerPoint, Teams, and even used for teaching.

That’s the kind of impact I’m really excited about. Healthcare accounts for nearly 20% of GDP, and so much of that cost lies in the complexity of workflows. If providers can use AI to improve outcomes and reduce cost, the societal impact would be massive.

Matthew Berman:
Absolutely. I’m extremely excited about hyper-personalized healthcare too. I already use ChatGPT Copilot to help me answer personal health questions. It’s really promising.

And that material science breakthrough you showed—immersion cooling—was discovered using AI, right?

Satya Nadella:
Yes! That was such a cool example. The breakthroughs in material science, thanks to AI, are phenomenal. There’s so much opportunity in areas like that.

Matthew Berman:
Something I’ve been hearing from younger people: they’re hesitant to use AI—or avoid it altogether—because they’re concerned about the environmental impact. They worry about energy use. What would you say to them?

Satya Nadella:
First of all, I find it inspiring that younger generations care so deeply about this. That’s the right kind of pressure—it pushes all of us to ensure what we’re building serves meaningful societal outcomes.

Whether it’s healthcare, education, or financial access—it all comes back to economic prosperity and solving real challenges. That has to be the foundation: we’re not just building tech for the sake of it, but for people and the planet.

The second part is equally important: we must pursue sustainable abundance. One of the ways I think about this is: tokens per dollar per watt. How efficiently can we use energy to generate intelligence that, in turn, improves lives?

Tech today uses maybe 2–3% of total global energy. It may double, yes—but for that to be acceptable, it must create significant real-world value. That’s how we earn social permission to continue expanding.

That’s also why we’re among the largest buyers of renewable energy. In fact, many new projects are subsidized by companies like ours. But the bottom line is: we need to keep pushing on efficiency and impact.

If AI is powering better healthcare, more productive small businesses, new scientific breakthroughs—then the energy is well spent.

Matthew Berman:
I’m glad you said that. I’ll definitely share this part of the conversation with people who are concerned about the environmental side of AI.

Another big shift we’re seeing is in computing architecture. The line between deterministic and non-deterministic systems is blurring.

I saw a demo where someone recreated the game Doom using a diffusion model—every frame was predicted. Do you ever see an operating system that works similarly, with little or no traditional code?

Satya Nadella:
That’s a fascinating idea. We did something similar with the Muse model, a world-action model we trained on gaming data. You could use an Xbox controller to feed actions into it, and the model would generate the next scene. It’s very similar to robotics in that way.

So yes—everything is, in a sense, generated.

Even what we call “deterministic systems” are not perfectly deterministic. One of the core challenges in computer science is that you can’t fully prove what a program will do.

So, yes, these are stochastic systems. But they still need to behave in deterministic ways—ways we can inspect, understand, and bound.

As Elon said when I interviewed him, we need to understand the “physics of intelligence.” When we stitch together these complex systems, we need guardrails. We need to sandbox them, understand them, and control how they operate.

Take the coding agent we launched. It runs in a controlled environment powered by GitHub Actions. You define the virtual machine, its permissions—whether it has internet access, access to tools, etc.—and you log every action.

That’s how we’ll learn to integrate traditional imperative code with these new agents. It’s about monitoring and managing that interaction carefully.

Matthew Berman:
That’s really exciting. In your keynote, you said we’re in the “middle innings” of this transition. It’s an incredible time to watch how software evolves from here.

Thank you again for taking the time to speak with me.

Satya Nadella:
Thank you. It’s been a pleasure. I appreciate you coming—and I’m looking forward to what’s ahead.

For more in-depth interviews with the people shaping AI, follow us on X and subscribe to our YouTube channel.