All right, hello everyone. All right, it's fantastic to be back here in India and back here in Bengaluru. It's always unbelievable to come back and see the energy and the excitement, especially at a time like this. You know, we're perhaps entering this next phase where we're going to go from talking about AI and admiring some of the new capabilities, whether they're in infrastructure or models, to doing things with AI that are bold and big. And that's to me what's really in the air when I come here and I see people excited about what's happening and what you're doing. So I want to spend the next half hour or so just giving you at least what I see out there as the possibilities and what we at Microsoft are focused on in terms of building platforms. Because at some level, to me, Microsoft has always been fundamentally about two things: we are a platform company and we are a partner company. And even in the age of AI, that's what's going to be true. But whenever you talk about these platforms and platform shifts, I think I lost my... okay, you always need to be grounded in what the foundational force that's driving the platform shift. In fact, when I look back even my 35 years in tech, it's been about one fundamental force, which has been Moore's law. I always recount, you know, Bill used to bring a bunch of us together every year and he would just literally put the Moore's law and then what's happening with memory and say, 'Go fill it with software.' That was just the single instruction to the entire company. And it's very true even today, right? When you think about the scaling laws that are powering AI and pre-training in particular, it's really Moore's law at work again. And it started in 2010 with DNNs and then obviously the GPUs. They inflected again perhaps with Transformers just because of the efficiency of data parallelism with Transformers. And what was happening perhaps of doubling of capacity every 18 months started to double every 6 months. That's really what the scaling laws were. And by the way, there's all this debate: what's happening with scaling laws for pre-training? Will they continue? Have we hit the wall? We fundamentally believe the scaling laws are absolutely still great and will work and continue to work. But they do become harder, right? As the sizes of data become higher, the parameter counts become higher, the systems problems are bigger. So these synchronous data parallel workloads are new workloads, and so therefore that'll continue. But the more interesting thing that you are now starting to see is another scaling law with the inference time or test time compute scaling law. Right? I mean, at some level pre-training had that sampling step. This is sort of always about using that sampling step more efficiently. I think I've lost the mic. Just give me a hand mic. Just... good, no problem. So it's about being able to scale with inference time that I think is going to really take this to the next level. So we're very, very excited about, I mean, o1 and think hard, like to me, Copilot Think Hard has become the thing that I go to all the time. And that shows not just for sampling for additional pre-training, but to be able to use it during inference time to think harder and get you better results. So I think that these are capabilities that are just going to increase. Ultimately, three things: one is this multimodal capability that we now have as the interface to all software. Right? I recently set up, I guess there's this, I don't know what they call it, the action button I think on the iPhone, it's now set up to Copilot for me. And now I can sort of confidently speak in Hyderabadi Urdu to it and it's beautiful. It understands me. It's like speaking to my high school buddies. And the fundamental idea that you can now have that familiar simple interface with all of computing, I think it's going to change every software category. Then you couple it with these planning and reasoning capabilities. Right? So whenever you go to GitHub Copilot Workspace, right, that thing is invoking that ability to think about planning and executing that plan as a multi-step process. Right? It's the beginnings of what is truly agentic behavior. And then the other side of it is to be able to stitch these things that are outside the system or outside of the model: memory, context, tools use, and entitlements. So in fact, if anything, in the next 12 months, every developer is going to be really focused on how do I take my model and make it aware of the tools it can use, just beyond even function calling. How can I make sure it has the right understanding of its entitlements? How do I make sure it has memory and long-term memory? And that to me, I think, is what's going to truly help us create this rich tapestry of agents. Right? When we think about agents, you stitch the multimodal capability, planning and reasoning, and memory and tools use in particular, plus entitlements, you can start building personal agents, team agents, enterprise-wide and cross-enterprise agents. So that agentic world is what we are looking forward to sort of all building. Of course, at Microsoft, for us, it's never quite frankly about any of these technologies on their own. It's a means to the end, which is about empowering every person and every organization on the planet to achieve more. That's what that sense of empowerment that this platform can provide, I think, is going to be the next level. And that's what we're really, really focused on. And to that end, we're building three platforms: Copilot, Copilot and AI stack, as well as Copilot devices. And so what I want to do is just kind of give you the broad contours of what these three platforms all entail. The best way to conceptualize Copilot is it's the UI for AI. So one of the ways you think about even in a very rich agentic world, remember the AI will need to interface with us, that needs a UI layer. And that's why I think this organizing of Copilot becomes even more important in a world where there are many, many agents doing autonomous work. I think that's the best way to think about it. Now, the approach we have taken is to build Copilot into the existing workflow. One of the best examples I've seen quite frankly is one of the high-stakes things of knowledge work. Right? So here there's a doctor, she's getting ready for a tumor board meeting. Right? Think about a tumor board meeting, it's a high-stakes meeting. That means you would have had to read all of the reports, know exactly how much time to spend on each one. So that means this creation of the agenda is a reasoning task. So it creates an agenda which knows which one is the more complicated case that needs more time. Then you go into a Teams meeting. So all these docs are now having a conversation on all the cases. They're able to focus on the case as opposed to taking notes because there's an AI that's taking detailed notes on all of it. Then at the end of it, she's a teaching doctor, right? So she wants to be able to take what happened in the tumor board, go to class. And so that means she's able to take the notes, putting into Word, from Word to PowerPoint, go to the class. That simple workflow, right, that doctors everywhere are doing that has impact on lives, can be enabled with AI just being built into the workflow. So that's just a good example of sort of how to think about AI being infused into current workflow. Now, let's take it to the next step. Now with Pages and Chat with Web and Work Scope, I think it's sort of completely about thinking of new types of workflows. Right? Fundamentally, I'm now able to access information, whether it's web information or it's information inside of my Microsoft 365 graph. Right? One query at a time, I can go get back all the data. I can then promote that data into this interactive AI-first canvas called Pages. And once I have it in Pages, I can then use Copilot inline in Pages to keep modifying it. And I sort of use this as a metaphor where I'm thinking with AI and I'm working with my colleagues. Right? So think about that, like that's the new workflow where I think with AI, I promote things into Pages, I invite others, I collaborate with others. And by the way, AI is present even on that canvas. So Chat plus Pages is going to become this new AI hub for just like how Word, Excel, PowerPoint back in the day changed how I worked. Now Chat plus Pages, which will be a new module that effectively enhances how you work with AI. Now we're not obviously stopping there. The next thing that we are thinking about is, you know, extensibility. So how do you extend AI going forward? It starts with something called Copilot Actions. You know, for those of you who are big users of Outlook rules, which I was for a long time until the complexity was too much, you know, this is think of it as rules for the AI age. But they don't work one app at a time, they work across the entire M365 system. So that's the beauty of Actions. So if you think about, so much of our workflow is gathering information, distributing information, it's about connecting people to artifact, to other people, to other artifacts. Right? That's a lot of what knowledge work is all about. I can now set these up essentially as Copilot Actions. So that's the first extensibility. Now, of course, you can build full agents. And we ourselves are building many of these agents that have scopes, right, at group level, process level. Right? You can have a project agent. We have agents that are working inside of Teams, like an interpreter, like a facilitator. This is just like having an additional team member that is helping you with your tasks inside of your team. SharePoint, in fact, every SharePoint now has an agent. And so think of it as an intelligence layer on top of SharePoint that's just built in. So I just want to roll a video to give you a flavor for these agents that are being built in M365.
New agents in Microsoft 365 are transforming how humans and AI collaborate. Meet agents in SharePoint, which unlock high-value insights that are connected to your organization's documents. Created with just a few clicks, these agents enhance knowledge sharing. They can be customized with additional data sources in Copilot Studio and can be shared anywhere, like a Teams chat. Next is the facilitator agent, joining meetings to manage tasks like the agenda, real-time notes, and action items, allowing the team to stay focused on the discussion. And the facilitator agent can also streamline communication in chats by providing real-time summaries and responding to questions, so the team can focus on what matters. With the new interpreter agent, language barriers are a thing of the past, enabling real-time speech interpretation so everyone can speak and listen in different languages. Thoughts from marketing about the upcoming campaign? If we get a game plan... And the project manager agent creates project plans, assigns tasks, and even completes them on behalf of the team, keeping everyone informed and collaborating effectively. Finally, for specialized business processes in HR and IT, the new Employee Self-Service agent in Copilot Business Chat enables employees to get instant answers and take action, such as logging a help desk ticket. And it can be customized in Copilot Studio using pre-built workflows and more. New agents in Microsoft 365, supercharging productivity to reinvent how work gets done.
So that's just an example of essentially agents that we have built into the system. But the real exciting thing is, of course, you all being able to build agents. And that's where Copilot Studio comes in. And our vision with Copilot Studio is simple: this is the low-code, no-code tool for building agents. So think of it like when you had Excel and you could build spreadsheets. To us, building agents should be as simple as building spreadsheets. Copilot Studio is about helping every one of us to have real agency to shape and reshape the workflows around what we're doing as knowledge work. That's essentially what we want to be able to... it's the swarm of agents around what we do that is helping us get work done, create more flow, less drudgery. And Copilot Studio, like take something like field service, right? It's as simple as first giving it a prompt, giving it the instructions on what the agent is all about. Then it's about grounding it in knowledge sources, like in this case, it's about pointing it to the right SharePoint source. And then once you do that, it just creates an agent out of the box for you. Right? So that simplicity of being able to create agents that are essentially low-code, no-code, programmable is what we're doing with Copilot Studio. So now that you have this Copilot UI for AI, you have the ability to extend it with actions, you have the ability to use agents that are built in and build your own agents, you have a complete system. Now the question is ROI and measurement, because that's the other question, right? Which is, okay, how... because one of the fundamental things we want to also ensure is that real motivation for change. After all, what am I doing that is better, that is helping improve not only my own productivity but my organization's outcome? And that's where this measurement comes in. And we're building out these Copilot analytics so that every individual is not just a top-down thing, right? A sales territory manager can now go in and take an output metric, something like increase sales, increase yield, and correlate it back to specific usage of all of these Copilot features. So to me, that's another one that drives the adoption cycle. So you're not waiting, but you're able to see in real time how with increased usage you're able to drive business results. So that's really the first platform we're building: Copilot as the UI for AI with extensibility and measurement. Now we're already seeing fantastic results inside of Microsoft. Basically, we're baking in double-digit strong improvements to productivity across every business process: customer service, HR self-service, IT ops, finance, supply chain, marketing. Right? I mean, think about marketing where there is, you know, there's sophistication in buying, but there is a lot of places where there's a lot of inefficiency around content creation. There's massive amount of leverage and operating leverage we get there. So significant use cases across the length and breadth of our own company. And of course, when I come even to India and I get a chance to meet with everyone here, it's no longer... the diffusion is so fast, right? It's no longer about waiting multiple years before it becomes mainstream. I learn a lot by watching many of you deploy this at scale. In fact, this morning I had a chance to talk to folks at Cognizant. They were telling me about how they've deployed it across the entire sort of employee base. The knowledge turns, one of the things that Andy Grove way back in the 90s used to talk about was knowledge turns, which was about being able to create knowledge fast and really be able to then diffuse that knowledge. Right? That's what, like just like supply chain turns people talk about them inside of retail, this is about knowledge turns in any knowledge industry. And another example was what Persistent is doing. They built a contract management essentially agent. In fact, you can just go address it, you know, at Persistent, whatever, a contract AI inside a Copilot, you can get access to it. And that agent is available to you throughout the entire lifecycle of a contract management, where even a single change can have significant impact. So these are just a couple of examples of people already deploying this Copilot system at enterprise scale. So now I want to talk about the next platform, which is the Copilot stack and the AI platform. Now, to us, we've always conceptualized and built Azure as the world's computer. We continue to be super committed to it because one of the fundamental realizations is AI doesn't sit on its own, right? It requires the entire compute stack. And so we're building that out at a worldwide level. We have 60-plus regions, we have 300-plus data centers around the world. In India, we're excited about all the, you know, the regions we have, right? We have Central India, South India, West India. And then we also have the capacity that we built up with Jio. So we have a lot of regional expansion happening. And I'm really, really excited today to announce the single largest expansion we have ever done in India by putting three billion additional dollars to expand our AI capacity. I had a chance to meet the Prime Minister Modi ji yesterday. It was fantastic, was great to listen to all his examples, you know, his vision around how he wants to drive through AI Mission. But it's that combination of really the yojanas he has, the India stack, the entrepreneurial energy in this country, and the demographics both on the consumer and the business side that is all getting into a virtuous cycle. So that's why we feel fantastic about bringing core compute capability for the next generation AI. And so now with infrastructure, there is in some sense a new formula for quite frankly for any country or for any company. I think of that formula as tokens per dollar per watt. That's it. Tokens per dollar per watt. Right? For two years from now, five years from now, ten years from now, we will be talking about the correlation quite frankly between GDP growth in any community, in any country, in any industry, or even in a company level, fundamentally their own growth on how efficiently are they able to drive that equation. And to that end, that means infrastructure, infrastructure, infrastructure needs to be the highest priority. And we are innovating in every layer of it. Right? Think about it at the data center level. Like these data centers are just, you know, everything from how we think about even the construction of a data center that is optimized for liquid-cooled AI accelerators is a new engineering feat. And so that's what we're doing. Everything from how do we then work with upstream from us with renewable energy folks to get onto the grid base load that comes to our data centers, that then has the right cooling infrastructure with zero waste and zero water usage. How do we really make sure all of that gets built into at a system level? And then of course, silicon innovation. We are innovating with NVIDIA. In fact, today I think we have our first GB200 clusters up live in one of our data centers. Very excited about what that would mean. We then also working with AMD. We are building out our infrastructure with, or we're building our own silicon with Maia. All of these, in fact, Maia today is taking a lot of the customer service traffic in microsoft.com. So we're building, I would say, world-class AI accelerator infrastructure, but the entire system stack about it, optimized for training, optimizing the kernels for inferencing. So that's significant investment and innovation that's happening through us, our partners. And so we're very excited. In fact, I think of this as a golden age for systems when it comes to innovation. Now, if you have the infrastructure, the next big, big, big consideration, in fact today I met many, many partners, many customers, the first thing everybody wants to talk about is, it's fantastic we're talking about AI, but how do I get my data in shape? Right? I mean, data is the only way to create AI. It's not just for the pre-training, right? We know for RAG you need data, you need data for post-training, you need data for doing sampling, for doing inference time compute to improve pre-training. So data pipelines and data is everything. And so the first thing to do is to rendezvous the data with the cloud. And that's where we are building out our data estate such that you can bring all of your data. Right? So you can bring whether it's Snowflake, whether it's Databricks, whether it's Oracle, whether it's our own SQL, what have you, bring it to the cloud. We have fantastic operational stores that are being plumbed for AI. Right? Whether it's Cosmos DB or whether it is SQL Hyperscale or Fabric for analytic workloads, all of these are AI-ready. In fact, if you look at even ChatGPT, they are some of the biggest users of Cosmos DB because where's there stateful applications, right? Where is all the user state of ChatGPT? It's in Cosmos. So to me, the data layer is a super important layer and we are doing everything to make sure that we can help get the data in shape for you to be able to then use in conjunction with these models and to build models. Those are the two things, right? There is models that are being trained on the data, but you're also doing things like retrieval augmented generation using data. And so that's sort of why the proximity on the data gravity is a huge locality of data will matter. Now, once you have the infrastructure and data, the third thing you have to do is to have the AI app server. In fact, if you look back, you know, when the web happened, what did we do? We built IIS as the app server. When the cloud happened, what did we do? We started building the cloud-native app server. Same thing with mobile. So every generation has required an app server. And that's what we're doing with Foundry. And with Foundry, it starts with models. Right? We have the rich, and obviously with OpenAI, and OpenAI's innovation now with o1, we're excited about what's going to come with o3, with obviously GPT-4o, all of that's available. Plus all the open-source models, right? Whether it's from Llama or Mistral, we have industry-specific models, models that are being built out of India, where people are building for Indic languages, for India's vertical-specific needs. It's fantastic to see the amount of innovation that people are doing around models all over the world. And so we want to have the richest model catalog. And some of the more popular models will be available even as models of service, right? So they'll just be underneath a facade of an API that you can then go access. Now, once you have the models, you want to deploy these models, you want to be able to fine-tune these models, you want to be able to distill these models, you want to be able to do evaluations on these models, you want to be able to do groundedness tests, you want to be able to do safety. All of that, instead of building them all separately, we're building them all into the app server. In fact, evals are going to be the most important thing. So even for me, the guidance I give our teams is simple, which is stay on the frontier of the new model and then make sure that you have agility in the app server layer so that you can keep moving with models. Right? You'll move, you'll use the first, like the latest sample, then you will cost-optimize it, latency-optimize it, and you'll start fine-tuning it for your specific use cases for evals. Right? So that's sort of the loop that you are constantly going through. And that's the idea behind Foundry is to just streamline all of that. Fantastic momentum again in India. When I look at customers who are already deploying this, using it, in fact, lots of good feedback I'm gaining from lots of people. People are pushing on even these multi-agent type of deployments. So we're learning a lot quite frankly from some of the ambitious things that you all are doing. And we are definitely very, very grounded in how we are going to progress on that roadmap. I think next year we'll not be talking as much about models, but we're going to be talking about model orchestration, model evals, and how you're able to deploy these model-forward applications. That I think is going to be the big shift across the industry. And so today, I mean, like, you know, when I saw the folks at Bank of Baroda, they showed me three things, right? They showed three agents. They built a self-service agent, they built for the new customers, they built a relationship manager agent. They also then built an agent for their own employees. I had a chance to see this fantastic startup, ClearTax. ClearTax, you know, I guess I shouldn't be able to do my taxes on that, which is just simple on a WhatsApp, you go submit your receipts, I guess, and then you get refunds. I love that part in particular. And then I had a chance to see the ICICI Lombard folks. And one thing I had not realized, in healthcare in India, for example, you don't have standardized claims forms. So each one is a novel different entry and so someone has to sort of go read it. That's where I think you can improve. Like you think about anytime you improve healthcare efficiency, that improves the economy, right? Because you then have assurance of your insurance being taken care of. Say MakeMyTrip, I had a chance to meet with the team and they're doing some phenomenal high-ambition work of being able to take one of the most sophisticated, you know, industries and verticals around travel, and really whether it's hotels or whether it's air or whether it's other transportation, how do you really have a multi-agent framework that really they can deploy. But one of the other things is it's not just about the broad big companies that are or startups that are doing it, but it's sort of the diffusion rate of this technology in India is what's exciting. So to that end, I just wanted to roll out, I just wanted you to see the video from The Cooperative Baramati. Go ahead and roll out the video.
In India, we are producing some of the major crops in large quantity like sugarcane, wheat, rice, pulses, and cotton. But when we compare our yield with the other developed countries, it is too low. Soil erosion is the major problem in India today because of overuse of pesticides and fertilizer. L.B. Baramati is working for the agriculture community for more than five decades. This trust was started to uplift the farmers who were very poor, had no resources. Through this, we have changed the life of millions of farmers. AgriPilot.ai allows farmers to avoid guesswork and use science to give the ground truth to make the right decisions and make them successful. In this project, we have selected 1,000 progressive farmers from across Maharashtra. We have installed weather stations, sensors in the soil, and provided satellite support to all the farmers. We are collecting real-time data from the soil every day. AgriPilot.ai uses Microsoft Azure Data Manager for Agriculture for making the decisions. It's running more than 20-plus algorithms that can lead to accurate results based on the historical pattern. Azure OpenAI allows the farmer to ask the questions in local languages, which can be delivered on their WhatsApp. They are precisely using irrigation, fertigation, and pesticide practices in their farm.
I mean, and to me it sort of really connects all the dots, right? It sort of connects the dots between even all the technologies we have been building, right? From Azure IoT, the data connectivity back to a data plane, and then to be able to use something like Azure AI. But ultimately, to empower a farmer to be able to do their farming with higher yield, that I think in some sense speaks to the power of this technology and what we can do with it. Now, the last layer I want to talk about, if you have infra, you have data, you have the AI app server, is tools. Now, I always go back, you know, Microsoft started as a tools company. We continue to be super passionate about our tooling. With GitHub, it's fantastic to see what's happening. Right? In the... we have now 17 million members of GitHub in India. It's the second-largest community next to the United States. In fact, it's projected to be the largest, I guess I forget the year, I think 2028. Yeah, so I can't wait, three years from now, 2028 is when the crossover will happen, where India will have more developers on GitHub. It's just exciting to see that. Now, we also have contributions from India to AI projects that are just second to the United States. And so it's fantastic to see the active involvement of the developer community out of India in making progress on all the open-source projects on GitHub when it comes to AI. It just speaks to again the talent there is and the energy there is in this community. Now, we're continuing to make great progress on GitHub Copilot. One of the features that I was like looking forward to, we now have, which is the multi-file edits. Right? First we started with continuations, we then had chat, we brought continuations plus chat together, and then brought it to multi-file. That's fantastic, so that I can do repo-level edits. Now, the other thing we are also doing is bringing a free tier. In fact, we just launched it in December, and I think in India it's sort of like the place where it's really taken off. We're very, very excited about bringing to VS Code GitHub Copilot free tier and seeing that now broadly getting distributed. The feature or the product area, you know, back I forget now, 2019, now maybe not 2019, 2020 I think is when I first saw GitHub Copilot. That's when my own conviction on what LLMs can do completely changed when I started seeing it. Similarly, when I first saw GitHub Copilot Workspace is when I felt like time had come for us to make the next leap beyond chat to real agents. Because that's what it is, right? GitHub Copilot Workspace is the first agentic sort of piece where you now can take a GitHub issue, create a spec which you can edit, you can create a plan, you can then edit the plan, and then you can see it execute across the full repo. And to show all of this, I wanted to invite up my colleague Karen up on stage.
