Learn. Share. Evolve…

0:00:06 we’ve entered a world where our computers are now being designed from the ground up
0:00:12 specifically for AI. Companies like Microsoft are building computers with special chips in them
0:00:18 that serve the sole purpose of running AI. So in this episode, I want to demystify the world of AI
0:00:24 powered computers. And to help us understand it all, I’ve invited Pavan Devaleri from Microsoft
0:00:29 to the show. He’s the corporate vice president of Windows and devices and is right at the
0:00:35 forefront of these incredible developments. We’ll dive into topics like privacy, especially with
0:00:40 features like recall, which automatically takes screenshots of your computer. We’ll talk about
0:00:46 how these new AI chips are making AI more accessible and affordable to everyone. And we’ll get a glimpse
0:00:52 into what PCs might be capable of just five years from now. And trust me, it’s moving faster than you
0:00:58 think. Also, while I was at Microsoft, I was given the opportunity to ask Satya Nadella, the CEO of
0:01:05 Microsoft. Just one question. And at the end of this episode, I’m going to share that question and his
0:01:11 response to it. So definitely stick around for that. But let’s go ahead and get right to it and dive into
0:01:13 my conversation with Pavan Devaleri.
0:01:21 Cutting your sales cycle in half sounds pretty impossible. But that’s exactly what Sandler training
0:01:27 did with HubSpot. They used Breeze, HubSpot’s AI tools to tailor every customer interaction without losing
0:01:35 their personal touch. And the results were pretty incredible. Click-through rates jumped 25%. And get this,
0:01:42 qualified leads quadrupled. Who doesn’t want that? People spent three times longer on their landing pages.
0:01:48 It’s incredible. Go to HubSpot.com to see how Breeze can help your business grow.
0:01:57 So, the next PC that people might buy, it might have an AI processor in it, the new NPUs.
0:01:57 Yes.
0:02:02 I’m curious, what sort of things that somebody that just uses it for maybe email and Netflix,
0:02:06 like what sort of benefits are they going to get by having an NPU in their computer?
0:02:11 It’s a great question. I think neural processing units are going to be, I think, a third processor
0:02:18 inside your computer. Just like we have CPUs and GPUs. I think NPUs will get added to that mix. And
0:02:25 primarily because NPUs will give you a lot more access to running AI efficiently on your device.
0:02:29 And our goal very much is to have these AI capabilities become available broadly to both
0:02:35 consumers and commercial customers and really build a platform where developers can build on top of it as
0:02:39 well. And I think it’ll show up in a couple of different ways. First, your own Windows device
0:02:45 experiences, whether it is something as simple as getting settings to be simpler and easier to use,
0:02:49 whether it is how you search for your files and folders in the operating system,
0:02:54 or whether it is how apps on top of Windows are built, will all change, I think, going forward.
0:02:58 I think at the end of the day, the goal is for those devices to become more simple, more intuitive,
0:03:04 become more thoughtful in terms of completing tasks and activities on your behalf. And at the end of
0:03:08 the day, just accelerating, I think, what you can find ways to make happen on your computer.
0:03:13 Right. Very cool. I know some of the stuff that has been teased has been, you know,
0:03:17 obviously NPUs. You can start to run AI locally on the device.
0:03:17 Yes.
0:03:21 And so I imagine, you know, people who are just using it for things like email and stuff like that.
0:03:24 Yeah. I mean, even email should get simpler for you, right?
0:03:27 Some of the kind of summarization that you can do that goes through the cloud now,
0:03:28 now we can do it locally.
0:03:32 That’s correct. Yeah. And you see this with features that we have on Copilot Plus PCs right now, right?
0:03:36 We have this idea of click-to-do, which gives you a one-click moment where we have understanding
0:03:42 of screen and context. And when you click write, we open up a variety of tools, kind of like your email
0:03:47 example. Just the ability to write, summarize, understand content, I think will become a lot more
0:03:52 pervasive. Certainly with emails, you might be offline, you might have encrypted content that
0:03:56 you only want to see summarized locally. The fact that those skills can become proactive on your behalf
0:04:01 are things I think customers and consumers will see on a broad basis going forward.
0:04:07 Definitely. Can I ask you the difference between, you know, we have CPUs, GPUs, NPUs, maybe?
0:04:07 Yeah.
0:04:11 For the laymen that don’t know, like, the difference between them, you know, me, I don’t
0:04:15 really totally know the difference. Can you help me understand the difference between all
0:04:15 three of them?
0:04:18 Yeah. You know, I think most of the world, you know, has built applications and devices and
0:04:23 experiences that utilize the CPU. Over the last couple of decades, GPUs have become really
0:04:29 important, especially when it comes for gaming, you know, using high-resolution displays, for CAD,
0:04:35 you know, type workloads where visualization is important. And I think the value of the NPUs, essentially
0:04:40 for client devices, for laptops and battery-powered devices, is to be able to give you the ability
0:04:46 to accelerate the ability to run these models and sort of lower the footprint and tax of
0:04:50 those models running on your device. And at the end of the day, we expect these NPUs will
0:04:55 make it just easier and lower cost for you to have models running on your behalf on a pervasive
0:04:56 basis inside the device.
0:05:01 Right, right. And so the NPU, it doesn’t necessarily mean that you don’t need a GPU anymore, right?
0:05:05 So you’re going to still have an NPU and a GPU, which is something we were talking about a little
0:05:11 bit last night, is you might be able to offload the sort of video processing to the GPU and still be
0:05:12 able to do things with AI using the NPU.
0:05:18 Bingo. I think NPUs will be a complement to GPUs and CPUs. And I think the reason why NPUs are useful is
0:05:23 because they’re very efficient, Matt, when it comes to energy efficiency and battery life. And so you can run,
0:05:27 you know, pretty powerful models, pretty capable models, but you don’t need the gigantic footprint of,
0:05:33 you know, thermals and battery life and heat sinks and all the stuff you’d expect for running a large
0:05:38 model. These NPUs really just get efficient in terms of running that model for you. The beauty with that is
0:05:42 concurrency. So you can have your apps, all the apps that you know and love today doing all the things that
0:05:47 they do and add new AI capabilities to those things and have those AIs, you know, be offloaded onto the NPU.
0:05:53 Not bogging down the GPU to do them. Correct. And, you know, another thing is, so I got a chance to
0:05:57 sort of check out the Applied Science Lab. It was great to have you. Yeah. One of the things you guys
0:06:03 talked about during the tour was that having this NPU really sort of democratizes AI. And I think that’s
0:06:08 a big concern, right, is people feel like, well, maybe only AI is going to be for people that have
0:06:13 a lot of money and, you know, the haves and haves not with AI. And it sounds like these NPUs are a little
0:06:19 bit less expensive to produce than GPUs. So maybe you can talk into the sort of like economics of that a little bit.
0:06:25 NPUs are more purpose-built for running AI models and workloads. And by virtue of being more purpose-built,
0:06:31 they’re inherently more efficient by way of the size and cost associated with building those NPUs. Right.
0:06:36 So the benefit for us there is we think we can deliver the NPU and the performant nature of the NPU
0:06:41 more broadly across devices, across a variety of devices and endpoints. In fact, we ourselves have started to do that with
0:06:47 Copilot plus PCs. Last year, when we introduced those Copilot plus PCs, we initially targeted a set of customers
0:06:51 that were more premium devices and, you know, prosumers. And this year now, we’re able to offer
0:06:56 those same class of NPU capabilities to a much broader footprint of devices and added, you know,
0:07:01 more mainstream price points. And it’s very much happening because NPUs scale better with price because
0:07:06 they have the ability to be focused on running AI compute and then be efficient and performant in that space.
0:07:10 And so our idea very much, especially with Windows, is to be able to bring the breadth of these features
0:07:14 and capabilities to the broad base of our Windows consumers, you know, globally. And having, you know,
0:07:20 price performance be great, have performance per watt be great is important for us. And NPUs are our
0:07:24 vehicle for making that come to life. Yeah, yeah. I noticed too, one of the things they showed us was
0:07:31 that they pointed like a temperature gun, like a FLIR kind of thing at the two computers. And one was
0:07:36 running the NPU and it was like 70 degrees. And then the other one, it was pointed at, it was like 113 degrees.
0:07:41 So that’s pretty crazy. Yeah. And I think it speaks to the fact that they’re just more efficient. Right.
0:07:46 And because they’re more efficient, they consume less energy, they generate less heat,
0:07:51 they give you a longer battery life, they give all of those attributes a lower price. And so we look
0:07:55 at them as a vehicle for them, you know, getting runway and scale with these devices. Right, right.
0:07:59 I want to talk about Recall real quick. When I first saw the announcement of Recall,
0:08:01 was it billed last year that they announced it for the first time? It was, yeah.
0:08:05 I thought that was like the coolest thing, basically, like having this, the whole history
0:08:10 of what was I looking at yesterday? And you can go back and find it. But I know that there was some
0:08:14 sort of privacy concerns and things that popped up around it that sort of freaked people out a little
0:08:19 bit. Yeah. So I’m curious, how has Recall evolved since then? Yeah, it’s a great question. We think of
0:08:24 Recall as one of several places where we think about the capabilities in the operating system evolving.
0:08:28 That capability and feature set, you know, surfaces and manifests itself in a variety of different ways.
0:08:34 Recall is one of them. Like we talked about earlier, search is another great one, for example. Click
0:08:39 to do is another experience. Camera stacks, audio stacks, paint having, you know, new capabilities,
0:08:43 photos being able to relight themselves. So AI is going to show up on the device and in the operating
0:08:48 system in a variety of different ways. Right. Recall was a great learning experience for us
0:08:53 in terms of understanding our customer needs and expectations, but we have privacy and them feeling
0:08:58 like they were in control. And it was a good experience for us to make sure the development process of
0:09:02 Windows allowed us to make sure we were taking advantage of those points of feedback. Right.
0:09:06 Which is exactly what we did. We had a several set of, you know, private release previews with
0:09:12 customers. We got great feedback through it and we’ve now successfully GA’d the product. And the early
0:09:16 signals we’re seeing so far is there’s a set of customers who opt into the device experience and it
0:09:21 really helps them kind of get into the flow of finding and searching and reliving points and times and
0:09:26 really augmenting their memory in a digital context. And we’re looking forward to the continued evolution
0:09:30 of that feature. Right. Right. So it’s not turned on by default on computers. Right. So like if you get a
0:09:36 new AI PC, it’s not. Right. You have to actually opt in. Right. Yes. As opposed to opt out. And also you
0:09:40 guys aren’t sending anything to the cloud. Right. It’s all staying right on the PC. Yeah. You nailed it.
0:09:45 That is a couple of important points. First, it is an opt-in experience. And after you opt-in, there are a variety of features in
0:09:49 the use of the product that are user defined and controlled. And so you have the ability to define
0:09:54 what you would like your recall experience to be. And then very importantly, the models and the data
0:09:58 stay local on the device. Right. And they’re all, they’re using the new NPU process. They’re using the
0:10:02 NPU, the models running on the NPU. Very cool. That’s right. So I want to look into the future a little
0:10:08 bit too. So five years from now. Yeah. What do you think we’ll be able to do with PCs that we can’t do today?
0:10:12 You know, we think about this quite a bit on the Windows team. And I feel like we make plans and what’s
0:10:15 surprising with the plans is the rate at which they are changing. Right. Right.
0:10:20 In some ways it is happening faster than we anticipated. What I think at the end of the day,
0:10:25 I think a core element of the Windows proposition is to make sure we’re in the business of empowering
0:10:31 our customers and consumers and developers and commercial, you know, information workers to be
0:10:35 able to do more with their computers, with their PCs and with Windows. I think that will be more so true
0:10:40 five years from now than today. By way of actual features and experiences, you know, I think we see a
0:10:45 world where Windows makes this evolution to the, you know, being an agentic OS very much like we talked
0:10:51 about at Build with the agentic evolution of the web itself. And I think that evolution of the OS itself
0:10:56 will be a platform construct. We ourselves will build a bunch of new experiences where you have models and
0:11:01 agents and capabilities running inside Windows in itself. And I think it’ll also be a world where
0:11:05 developers will be incented to build a bunch of new applications and experiences. Apps that you know and
0:11:09 love today will extend themselves with new capabilities. And the net new apps are going
0:11:13 to show up in the ecosystem that use things like model context protocol, for example, to be able
0:11:18 to talk across applications and talk to the OS in ways, quite frankly, we have probably not imagined yet.
0:11:22 Yeah, yeah. That’s kind of exciting in itself. It’s funny because I constantly try to make
0:11:25 predictions of where I think things are going. And I’m like, yeah, that’s probably three years out. And then
0:11:30 it happens three months later. That’s right. It’s kind of amazing. Yeah. One real example of that for us is the
0:11:35 performance and capability of these models are running on the NPUs. A year ago, we were kind
0:11:39 of wondering if we would have, you know, a billion parameter model run on the edge. And what we were
0:11:46 talking about earlier was we just last week had a 14 billion parameter model that has reasoning capability
0:11:50 running fully offloader to NPUs. And so what that means for a developer, what that means for the
0:11:55 Windows experience, I think super exciting for one, and it’s happening at a faster rate than we probably
0:12:01 could have imagined. Yeah, yeah. Is there any sort of misconceptions that you hear around like the AI PCs
0:12:06 that you want to sort of lay to rest? You know, I think the biggest thing is customers just knowing
0:12:13 that AI PCs are a full stack experience from the hardware, the device itself, they deliver great fundamentals
0:12:19 in terms of battery life and security and performance. And then all of that ladders up to serving a capability
0:12:25 or a platform that in turn has great AI experiences, I think is probably the most important things for
0:12:30 people to know. And so when you’re in your journey of having your next PC, you should expect this device
0:12:36 to be just a great device, you know, daily use. And also a durable construct in terms of future experiences
0:12:41 that are going to get unlocked, taking advantage of the platform. Right. So when it comes to AI right now,
0:12:45 it feels like we’re in this world where like everything is just like super fast. And it feels
0:12:50 like, you know, companies are sort of motivated to ship things really fast. How does Microsoft see
0:12:55 balancing, trying to keep shipping new features and keeping people sort of impressed with, you know,
0:13:01 the privacy security, the kinds of concerns people have? I think you nailed it. I think balance is the key
0:13:06 for us. And so in Windows for us, I think of it in a couple of different vectors for sure as a team that builds
0:13:12 products. We have a variety of mechanisms today for making sure we have active listening systems
0:13:17 across our ecosystem. And so we build a lot of these features using release previews in Windows where
0:13:21 we get feedback from insiders, we get feedback from the developer community, we get feedback from the
0:13:26 industry writ large, quite frankly. And so that’s one important aspect of our product development system
0:13:31 in Windows in itself that allows us to make sure we’re getting rich, robust feedback at the scale of
0:13:35 Windows. That’s one important piece, Matt. The second thing, kind of like with DMCP work that is
0:13:39 happening in Windows. It is happening quickly. For sure we are in a world the rate at which the
0:13:44 industry is evolving. In that example, the fact that the Windows team is a part of building these
0:13:49 new technologies, building these new standards, building protocols, allows us to go at day one,
0:13:53 build these capabilities into the base technologies in a way that will serve Windows customers in the
0:13:59 long arc of time in itself, I think. And the third one, I think, is some of this is an ecosystem
0:14:03 exercise where we will deliver some of these experiences for sure. And a lot of this is others who are going to
0:14:08 build on top of Windows and us getting signals from them on what they are seeing from their customers
0:14:12 and making sure we’re setting them up for success. So opportunity in multiple vectors and we have a
0:14:16 variety of tools in the toolkit to make sure we’re delivering meaningful value at the end of the day.
0:14:21 Very cool. Well, this is my last question. It’s sort of a two-part question. What’s something that
0:14:24 excites you about what AI can do today? And what’s something that excites you about what we’ll be able
0:14:31 to do with AI in the near future? That’s a great question. The things that I get excited about with AI today,
0:14:37 personally, which I found quite remarkable, is the ability for us to do things like deep research and
0:14:44 analyst work on the M365 co-pilot. It’s a capability that is an asynchronous task. It takes a while to
0:14:48 kind of run through. It requires a reasonable amount of domain knowledge. It requires an understanding of
0:14:53 your corporate environment and understanding of your team or your discipline or your department. And I’m
0:14:59 very excited with the quality of work that comes out of these high-performing agents that are running in
0:15:03 the Microsoft co-pilot environment, the M365 co-pilot environment. So that was the thing that I think
0:15:08 a year ago, to your point earlier, I don’t think I would imagine it’s simply just possible. And now we’re
0:15:13 getting to a place where they’re becoming a part of our collective team’s workflow when we do analysis,
0:15:18 when we do reports, when we synthesize feedback, when we make preparations for what future roadmaps are going
0:15:23 to look like. So that’s the thing that I’m kind of amazed with, quite frankly. And your second question,
0:15:27 you know, what’s coming down the pipe, what’s going to be kind of exciting? I think the singular thing I’m
0:15:33 excited about is what I consider to be sort of this 10x thing, where we have an unlock of what is possible
0:15:36 on the edge. I think you’ll be living in a world where the devices are going to get more performant.
0:15:41 We in Windows are spending a lot of time making sure the software tool chains and the run times and
0:15:45 environments for these models are getting more performant. I’m excited that the models themselves are
0:15:49 getting better, like adding reasoning on the edge as an example. And I’m also super grateful that we
0:15:54 have a set of class of developers who are building on top of these. And so I’m just excited that, you know,
0:15:59 for years we would invest in how much more experience and value can we get. And Kevin talked about, you
0:16:05 know, I was primarily relying on Moore’s law, that’s all we had. And now I think you have these compounding
0:16:09 effects of innovations happening across the entire, you know, device edge client computing stack that will
0:16:13 just unlock, I think, new things that are possible for customers. Amazing. Well, thank you so much for
0:16:17 spending the time. I appreciate it. Thank you. Thank you for having us. Lovely to be here and
0:16:19 build with you. Yeah, awesome. Thanks. Thank you.
0:16:26 Hey, we’ll be right back to the show. But first, I want to tell you about another podcast I know you’re
0:16:31 going to love. It’s called Marketing Against the Grain. It’s hosted by Kip Bodner and Kieran Flanagan,
0:16:36 and it’s brought to you by the HubSpot Podcast Network, the audio destination for business professionals.
0:16:41 If you want to know what’s happening now in marketing, especially how to use AI marketing,
0:16:47 this is the podcast for you. Kip and Kieran share their marketing expertise, unfiltered in the details,
0:16:52 the truth, and like nobody else will tell it to you. They recently had a great episode called
0:16:59 Using ChatTBT03 to Plan our 2025 marketing campaign. It was full of like actual insights as well as just
0:17:06 things I had not thought of about how to apply AI to marketing. I highly suggest you check it out.
0:17:11 Listen to Marketing Against the Grain wherever you get your podcasts.
0:17:19 All right. I mentioned in the intro that I had the opportunity to ask Satya Nadella just one question.
0:17:26 And what I wanted to know from him is, what does AI look like that could truly change the world?
0:17:33 So here’s the exact question I asked. If you can design an AI system that would fundamentally change
0:17:40 society beyond just answering questions and generating art, what would it look like and what risk and
0:17:46 responsibilities come with it? That was the exact question. And here was Satya’s response.
0:17:53 I would say the thing that I’m most inspired by was one of the demos I showed even today is in healthcare,
0:18:02 right? Because I feel like what touches all of us is this challenge of can we improve care and reduce
0:18:09 cost? So if there was one place where I would say this agentic AI has to make a real difference would be
0:18:14 take one of the challenges that we have as a society and go at it. And I think we’re at the
0:18:19 verge of it. Like what Stanford University was able to do by just essentially for something so high
0:18:27 stakes, like the tumor board meeting, and orchestrate all these agents and then ultimately empower the
0:18:33 caregivers there, right? The doctors, the nurses, all the specialists to be able to have a more successful
0:18:39 tumor board meeting and then improve care. That to me is where I think these systems are built
0:18:44 and then made available can make a huge difference. Awesome. Amazing. I really appreciate the time.
0:18:45 Thank you.