AWS re:Invent 2025 - Keynote with CEO Matt Garman

Welcome everyone to the 14th annual re:Invent. It's so awesome to be here. We have over 60,000 people here with us in person and almost 2 million watching online, including a bunch of you that are joining us from Fortnite out there. It's where we're streaming the keynote for the first time. Welcome to everybody and thank you all for joining us. It is incredible to feel the energy as you walk through the halls here in Las Vegas and it matches a lot of what I've been seeing recently as I've been talking to you in recent months. It has been an unbelievable year. AWS has grown to be a $132 billion business accelerating 20% year-over-year. I want to put this a little bit in perspective. The amount we grew in the last year alone, it's about $22 billion. That absolute growth over the last 12 months is larger than the annual revenue of more than half of the Fortune 500. And that growth is coming from across the business. S3 continues to grow with customers storing more than 500 trillion objects, hundreds of exabytes of data, and every day averaging over 200 million requests a second. For the third year in a row, more than half of the CPU capacity that we've added to the AWS cloud comes from Graviton. We have millions of customers using our database services and Amazon Bedrock is now powering AI inference for more than 100,000 companies around the world. This year, we gave you the first building blocks for deploying and operating highly capable agents securely at enterprise scale with Bedrock Agent Core. And we're seeing incredibly strong momentum from Agent Core. In fact, just a few months since launch, Agent Core SDK has already been downloaded more than two million times. And we announced Ocelot, our first quantum computing chip prototype. Ocelot is a breakthrough in making quantum a reality, reducing the costs of implementing quantum error correction by more than 90%. Now, all of this starts with a foundation of secure, available, and resilient planet scale infrastructure that is frankly unmatched anywhere. AWS has by far the largest and most broadly deployed AI cloud infrastructure anywhere in the world. Our global network of data centers spans 38 regions, 120 availability zones, and we've already announced plans for three more regions. In fact, in the last year alone, we've added 3.8 gigawatts of data center capacity, more than anyone in the world. And we have the world's largest private network, which has increased 50% over the last 12 months to now be more than 9 million kilometers of terrestrial and subsea cable. That's enough optical cabling to reach from the Earth to the moon and back over 11 times. But at Amazon, everything starts with the customer. So, I want to start off by thanking all of you. Today, we have millions of customers running every imaginable use case. The largest enterprises in the world across every single industry and vertical running their businesses on us. You've transformed industries like financial services, healthcare, media, entertainment, telecommunications, even government agencies all around the world. And as you all know at AWS, security is priority one. For us, everything is built on that foundation. This is the reason why the US intelligence community has chosen AWS as its cloud of choice for more than a decade. It's why companies like NASDAQ have moved their trading markets to AWS and why Pfizer chose AWS as their core for their digital transformation. And since the beginning, we've known how important partners would be to our customer success. That's why we're so proud to have such a massive network of partners. Many of them are here with us this week. Thank you to all of our partners, SaaS providers, system integrators, and solution providers serving our massive set of customers all around the world. We couldn't do it without you. And while I personally appreciate all of our customers, I will tell you that I have a special affinity for our startup customers. More unicorn startups have been built on AWS than anywhere else. And it isn't even close. Thank you to all of you, the innovators out there. More than ever, every startup and AI startups in particular are flocking to AWS. 85% of the Forbes 2025 AI 50, 85% of the CNBC Disruptor 50 are all running on AWS. It's incredibly amazing and I'm personally amazed at what these founders are inventing and I thought you all might like to hear from some of them today. Let's hear from the first AudioShake who was the winner of last year's re:Invent Unicorn Tank pitch competition.

Very cool and thank you to the AudioShake team for sharing. Now none of what we do at AWS happens without builders and specifically developers. AWS has always been passionate about developers and this conference is and frankly it always has been a learning conference. It's a little bit different and it's dedicated to all of you out there. Thank you to every developer out there and the millions of additional AWS developers all around the world. A special call out to our AWS heroes over here. I see you. Thank you so much. Awesome. And thank you also to the million plus members of our user group community in 129 countries all around the world. So why do we do this? What motivates us? Why are we just as passionate today as we were 20 years ago when we first launched AWS? What drives us every day is giving you all the freedom to invent. This has been our motivation since when we started AWS, the very beginning. We wanted to make it possible for every developer or inventor in her dorm room or garage to access the technology infrastructure and capabilities so that they could build whatever they could imagine. 20 years ago, it just wasn't possible for developers or builders to get the servers or compute capacity that they needed without investing significant capital and time. Developers were spending way too much of their time procuring servers, managing infrastructure, and not enough of that time building. We've actually felt this ourselves inside of Amazon. We had a company full of builders who had these incredible ideas of how they could make our customers lives better, but they couldn't move as fast as they wanted. So we asked ourselves why not? Why couldn't developers focus on building instead of on infrastructure? Why couldn't we bring the time and the cost of experimentation down to zero? Why not make every idea possible? And we've spent the last two decades innovating towards those goals. Giving all of you the freedom to keep inventing is why we're here today. And right now we're witnessing an explosion of invention with AI. Every single customer experience, every single company, frankly, every single industry is in the process right now of being reinvented. And we're still in the early days of what AI is going to deliver. And the technology is iterating faster than anything any of us have ever witnessed before. It wasn't that long ago that we were all testing and experimenting with chatbots. And now it seems like there's something new every day. But when I speak to customers and many of you out there, you haven't yet seen the returns that match up to the promise of AI. The true value of AI has not yet been unlocked. But a lot of that is changing fast, too. AI assistants are starting to give way to AI agents that can perform tasks and automate on your behalf. This is where we're starting to see material business returns from your AI investments. I believe that the advent of AI agents has brought us to an inflection point in AI's trajectory. It's turning from a technical wonder into something that delivers real value. This change is going to have as much impact on your business as the internet or the cloud. I believe that in the future there's going to be billions of agents inside of every company and across every imaginable field. Already we see agents accelerating healthcare discoveries, improving customer service, making payroll processing more efficient, and agents are also starting to scale people's impact up by 10x in some cases so they have more time to invent more. Wouldn't it be awesome if everyone could see that level of impact? We think so. And that's why we ask the question, why not? Getting to a future of billions of agents where every organization is getting real world value and results from AI is going to require us to push the limits of what's possible with the infrastructure. We're going to have to invent new building blocks for agentic systems and applications. We want to reimagine every single process and the way that all of us work. At AWS, we've been innovating at all of the layers of the stack to give you all the freedom to invent what's next. We have a lot to share. Let's get started. First off, what are the components that you need to deliver agents that are going to truly deliver value for you? It starts with having the most scalable and powerful AI infrastructure to power everything. You have to have a highly scalable and secure cloud that delivers the absolute best performance for your AI workloads. And you're going to want it at the lowest possible cost across your model training and customization and inference. Now, that's quite easy to say, but to deliver that requires optimizing across every single layer of hardware and software, and that is something that only AWS does. It turns out there are no shortcuts. Now, when you think about AI infrastructure, one of the first things that comes to mind is GPUs. And AWS is by far the best place to run NVIDIA GPUs. We were actually the first to offer NVIDIA GPUs in the cloud and we've been collaborating together with NVIDIA for over 15 years. And what that means is that we've learned to operate GPUs at scale. In fact, if you talk to anyone who's run large GPU clusters at any other provider, they'll tell you that AWS is by far the most stable at running a GPU cluster. We're much better at avoiding node failures and definitely deliver the best reliability. And the reason for that is because we sweat the details. Minor things like debugging BIOS to prevent GPU reboots. If you go to other places, they just kind of accept that as how it works and it's just you go on about your business, not us. We'll investigate and root cause every single one of them. And then we collaborate with our partners at NVIDIA to make sure that we're making constant improvements. Nothing is too small for us to be focused on. Those details really matter and it's why we lead the industry in GPU reliability. It takes hard work and real engineering to make that happen. And we improve on those new dimensions with every generation. This year we launched our P6 generation of EC2 instances featuring the NVIDIA Blackwell processor, the P6E GB200 Ultra Server, which provides over 20x the compute compared to our previous P5N generation. These are ideal for customers that are working with really large AI models out there. And we're doing that again today. I'm excited to announce the new P6E GB300. These are powered by NVIDIA's latest GB300 NVL72 systems and we continue to bring you the best-in-class compute for the most demanding AI workloads. Our full stack approach to hardware and software plus operational rigor delivers you the absolute best performance and reliability for the biggest organizations in the world. This includes NVIDIA themselves by the way who run their large scale gen AI cluster Project Saba on AWS and many others like OpenAI who are actively running on AWS today. They're using clusters of EC2 Ultra Servers with hundreds of thousands of chips today GB200s and soon GB300s and they have the ability to scale to more than tens of millions of CPUs to manage their agentic workflows. All of this is to support their ChatGPT application which I'm sure many of you use as well as the training of their next generation models. Or take Humane. Humane is Saudi Arabia's newly created company that's responsible for driving AI innovation in the region. We recently announced a partnership together on this groundbreaking AI zone for the Kingdom of Saudi Arabia. This partnership will bring customers high performing infrastructure models, AI services like SageMaker and Bedrock all while helping meet the kingdom's standards for security, privacy and responsible AI standards. Now this type of work has sparked some interest in other large government organizations and public sector who are interested in a similar concept. And so we sat back and asked ourselves, could we deliver this type of AI zone to a broader set of customers? Maybe even something that could leverage customers existing data centers. And that's why today we're excited to announce AWS AI Factories. With this launch, we're enabling customers to deploy dedicated AI infrastructure for AWS in their own data centers for exclusive use for them. Effectively, AWS AI Factories operate like a private AWS region, letting customers leverage their own data center space and power capacity that they've already acquired. We also give them access to leading AWS AI infrastructure and services including the very latest Trainium Ultra Servers or NVIDIA GPUs and access to services like SageMaker and Bedrock. These AI Factories operate exclusively for each customer and it helps them with that separation maintaining the security and reliability that you get from AWS while also meeting stringent compliance and sovereignty requirements. We're super excited to see what these AI Factories unlock for customers. Now AI at AWS we have always been about choice and if you want to have the absolute best in AI infrastructure you need to have the best compute for AI training and inference and AWS is by far leading the way with the broadest set of options including our groundbreaking purpose-built AI processors. AWS Trainium is our custom AI chip designed to offer the best price performance for AI workloads. Now, customers love Trainium for what it achieves for the training workloads, but I'm going to have to pause and be a little bit vocally self-critical here. People often give us a little bit of a hard time about product naming in AWS. No, no, it's true. Well, it turns out Trainium is no exception. We named it Trainium because it's designed to be an awesome chip for AI training. And it is, but as it turns out, Trainium 2 is actually the best system in the world currently for inference. Customers often ask me how can I best take advantage of the benefits of Trainium. And what I tell them is you probably are already using it and you just didn't know it. In fact, if you look at all the inference that's running in Amazon Bedrock today, the majority is actually powered by Trainium already. And the performance advantage of Trainium is really noticeable. If you're using any of Claude's latest generation models in Bedrock, all of that traffic is running on Trainium, which is delivering the best end-to-end response times compared to any other major provider. And that's part of the reason why we've deployed over 1 million Trainium chips already to date. Now, we've gotten to a million chips in record speed. And that's because the whole process, we control the whole stack. We can optimize end to end how we roll it out. And it allows us to move even faster. In fact, we've been able to ramp the volumes of Trainium 2 in our data center 4x faster than the next fastest AI chip we've ever ramped and we're selling those as fast as we can make them. Trainium already represents a multi-billion dollar business today and continues to grow really rapidly. Now, what does it look like when all of this comes together in a system purpose-built around Trainium? It wasn't that long ago that people had this saying where they said that the data center was the new computer. Well, when you're training this next generation of models, it turns out that the data center campus is the new computer. One of the best models in the world today is Anthropic's Claude. We wanted to give you a little bit of a look behind the scenes at how a model like this is born. Made possible by Trainium in Project Rainer. Let's take a look.

Really cool to see the massive scale that we've got to with Trainium so quickly. So, what's next? Last year, we announced that we were already hard at work on our next chip, Trainium 3, designed to make AI workloads better, faster, and more cost-effective. Today, I'm excited to announce that Trainium 3 Ultra Servers are now generally available. Now, these Ultra Servers are our most advanced, containing the very first 3 nanometer AI chip in the AWS cloud. Trainium 3 offers the industry's best price performance for large scale AI training and inference. Now, we've talked about the incredible results that we've seen with Trainium 2 this past year, but Trainium 3 Ultra Servers bring another huge leap forward, 4.4x more compute, 3.9 times the memory bandwidth, and this one is super important. Five times more AI tokens per megawatt of power. And as a special surprise, I have a rack of our Ultra Servers on stage with me today. Our largest TRN3 Ultra Servers combine 144 total Trainium 3 chips acting together in a single scale-up domain connected by custom Neuron switches. This delivers a massive 362 FP8 petaflops of compute over 700 terabytes per second of aggregate bandwidth and all in a single compute instance. And our custom-built EFA networks support scaling these out to clusters of hundreds of thousands of chips. No one else can deliver this for you. It requires all of these system level pieces to be co-designed together. It requires multiple types of custom silicon. It requires scale up and scale out networking. It requires a detailed and integrated software stack and of course the industry's leading data centers. In a real world example of how performance improves, we ran through a number of open-source models, open weights models that we've optimized to run on Trainium 2 and we wanted to see how they'd run on Trainium 3. As one example, here's an inference benchmark for a popular open source GPT OSS 12B model from OpenAI and we ran it on both Trainium 2 and Trainium 3. As you see here with TRN3, we get remarkable efficiency gains over Trainium 2. You see over 5x higher output tokens per megawatt all while maintaining the same latency per user, what we call interactivity in this chart here. And this is just one example. We see similar results as we run this across a number of different models, which is fantastic. We're excited to see what TRN3 is going to unlock for customers, but we're also not stopping there. I want to give you a bit of a sneak peek as to what's coming around the corner. That's why I'm excited to announce that we're already hard at work on Trainium 4. We're well into designing it and we're excited about what we're seeing already. Trainium 4 is going to bring massive leaps across every single dimension. Compared to Trainium 3, Trainium 4 will deliver six times the FP4 compute performance, four times more memory bandwidth, two times more high memory bandwidth capacity to support the very largest models that you have in the world. Trainium continues to push the bounds of what's possible with AI infrastructure. So you all can be freed to push the bounds of your industry. Let's hear from a startup that's using AWS's massive AI infrastructure to transform computational biology.

Really amazing to see what this scale of compute is enabling for customers like Laya to be able to accomplish. It's just incredible invention that's happening at the infrastructure layer. But we also know that infrastructure is just a part of the story. We're seeing nearly every single application in the world being reinvented by AI. And we're moving to a future where inference is such an integral part of every single application that everyone builds. Now to be successful in that future, you need a secure, scalable, feature-rich inference platform that you all can build on. And that's why we developed Amazon Bedrock. Bedrock is a comprehensive platform that helps you fast-track your generative AI applications as you move from prototype into production. With Bedrock, you get a broad choice of all the latest models. You have the ability to customize these models for your individual use case and your performance needs. You get the tools to integrate them into your data and the capabilities to add guardrails as you need them. All of this with the security and integrations that make it easy to build on applications and data that you already have in AWS. And companies of every size and in every industry all around the world are using Bedrock. Customers like BMW, GoDaddy, Stripe, those are just a few. With more than twice as many customers building on Bedrock compared to just this time last year. Bedrock is seeing unprecedented momentum. But it's not just the number of customers that are using it. It's actually the volume of the usage that's quite astounding. In fact, some customers are processing a huge number of requests through Bedrock. Today, some of the largest scale AI applications in the world all run on this platform. I actually asked the team to check for me. We now have over 50 customers that have processed more than 1 trillion tokens each through Bedrock. Incredible scale and momentum. Now when you all start building a Gen AI application, the very first thing you likely decide on is which model are you going to use? Which is the one that gives you the best cost, the lowest latency, what's going to give you the best answers. And a lot of times actually the right answer is a mix of different models for your applications or your agents which is why we think model choice is so critical. We've never believed that there was going to be one model to rule them all but rather that there would be a ton of great models out there and it's why we've continued to rapidly build upon an already wide selection of models. We have open weights models and proprietary models, general purpose or specialized ones. We have really large ones and small models. And we've nearly doubled the number of models that we offer in Bedrock over the last year. Today, I'm pleased to announce that we're introducing a whole host of new open weights models. These models include ones like Google's Gemma, Miniax M2, and NVIDIA's Neotron. And today we have a couple of new models that are debuting to the world for the very first time. I'm excited to announce that today from Mistral AI and available immediately on Bedrock are two new sets of open weights models. The first is Mistral Large which is a big leap forward from their Large 2 model doubling the context window size and vastly increasing the number of model parameters by more than five times. We're also launching today Ministral 3, which is a cool set of three models that offer really great deployment flexibility for ultra-efficient edge devices or single GPU deployments or advanced local operations. It's going to be super fun to see how you all use all of these open weights models. Now, in addition to providing a huge selection of third-party models on Bedrock, last year we announced Amazon Nova, which is Amazon's family of foundation models, delivering the industry's best price performance for many workloads out there. Over the last year, we've actually extended the Nova family to support more use cases and deliver more possibilities for you that deliver real value. We've unlocked speech-to-speech use cases, an example, with Amazon Sonic. And just a few weeks ago, we launched the industry's best performing model for creating embeddings across multiple modalities with Nova's multimodal embeddings. And the momentum has been really fantastic. Nova has grown to be used by tens of thousands of customers today. Everyone from marketing giants like Dentsu to tech leaders like Infosys or Blue Origin or Robinhood to innovative startups like Ninja AI. And today we're making Nova even better. Announcing a new generation of Nova with Nova 2. Nova 2 delivers cost optimized low latency models with frontier level intelligence. The Nova 2 family includes Nova 2 Lite, which is our fast and cost-effective reasoning model suitable for a broad set of workloads, and Nova 2 Pro, which is our most intelligent reasoning model for highly complex workloads. We're also introducing Nova 2 Sonic, which is our next generation speech-to-speech model that enables real-time human-like conversational AI for all of your applications. Now, Nova 2 Lite delivers incredible price performance for many workloads that we actually see our customers wanting to deliver in production. Nova 2 Lite compares really favorably in industry benchmarks to models like Claude Haiku 4.5, GPT-5 Mini, and Gemini Flash 2.5. In particular, Nova 2 Lite excels at things like instruction following, tool calling, generating code, and extracting information from documents, often matching or exceeding the performance that we see from these comparable models at an industry-leading cost performance. We think that Nova 2 Lite is going to be a real workhorse and is going to be really popular for a wide variety of your use cases out there. Nova 2 Pro is our most intelligent reasoning model and it's going to be great for when you have those really complex workloads. In particular, we look at really important areas where you need your agents to be great and that's where Nova 2 Pro really shines where skills like instruction following and agentic tool use are critical. In fact, for those, Nova 2 is frequently coming out on top. If you look at Artificial Analysis benchmarks in those areas, Nova 2 Pro delivers better absolute results compared to leading models like GPT 5.1, Gemini 3 Pro, and Claude 4.5 Sonnet. And for applications that need voice capabilities, Nova 2 Sonic offers industry-leading conversational quality at awesome price performance with improved latency and significantly expanded language support. We think you're going to love it. Now I also have one more Nova model that I want to talk about that has a unique set of capabilities. Now today's models out there are actually quite good at reasoning across one type of modality. Say they're looking at an image or listening to audio or outputting text and then maybe outputting in a different modality. Say they're maybe reading text and then creating an image. But in the real world you have to understand multiple modalities at the same time. Take for example this keynote. If you wanted a model to try to understand all that was going on in the keynote today, you'd want to get the nuance and understanding of everything that we're saying, that means you'd have to listen to what I'm saying. You'd have to understand the contents of all these slides. You'd have to be able to watch the videos and understand what's going on and what we're showing. Now, let's say you want to take that same model and produce a summary output for your sales team where it had a summary of all the launches we announced today along with some images and marketing material. Now, if you wanted to do this, you could, right? This is possible today. But it means testing a wide variety of different models, stitching them all together, and trying to accomplish this outcome, which is totally doable, but is quite hard. It'd be easier if you had a single model that could do all of that. And that's why I'm excited to announce Nova 2 AMI. It's a unified model for multimodal reasoning and image generation. Nova 2 AMI is the industry's first reasoning model that supports text, image, video, and audio input and then supports text and image generation output. So that's four new industry-leading models for Amazon Nova and we're just getting started. Up next, let's hear from Gradial who's building some pretty cool capabilities with Nova and Bedrock.

Really amazing to see what Gradial has been able to do with Bedrock. Now, the models that are available today are really incredible and it continues to impress me what everyone out there is able to accomplish with them. But as these models are used to power more and more mission-critical line of business applications and your agentic workflows, it turns out that the AI's ability to understand your company's data is what really starts to deliver huge value for your company and for your customers. And I'll just pause here. I can't stress this strongly enough. Your data is unique. It's what differentiates you from the competition. And I see this over and over again. If your models have more specific knowledge about you and your data and your processes, you can do a lot more. Now, the wizardry here comes when you can deeply integrate a model with your unique data and IP. But in order to do this well, it's critical that you have your data in the cloud. So, what are the best ways to get these models to access your data? Clearly, third party models don't start with access to your data. They don't natively know about your business and frankly you wouldn't want them to since you wouldn't want your proprietary data embedded in those models so that everyone else could use it. It's why actually the isolation that we provide inside of Bedrock is so important to prevent your data from leaking back to the core models. Now the most common techniques that we see people successfully use today to combine your data with the models are things like using RAG or vector data to provide your chosen model with context at inference time. And these are quite effective to help your models more effectively navigate your massive set of data and return relevant results. Usually what we see though is this only goes so far. Almost every customer I talk to wishes that they could somehow teach the model to really understand their data, really understand their deep domain knowledge, the expertise. They want the model to know their expertise when it's making its decisions. Let's say for a second you work at a hardware company that's looking to accelerate R&D for new products. You'd optimally want a model that understands your past products, your manufacturing preferences, your historical success and failure rates, whatever process constraints you might have. And then you want something that could combine all of these to provide an intelligent guidance for your design engineers. And it turns out whatever your company is, you have this incredibly vast corpus of IP and data that would be super valuable if it was integrated into the model you use. Now the natural question is why not just train a custom model? It turns out there's only really two ways to do this today. You could build your own model from scratch, include your own data in that, but of course, this is super expensive. And frankly, you probably don't have all of the data that you would need to build the general intelligence in the model. And then even if you did, you may not have the in-house expertise to pre-train a frontier model anyway. So that's probably not very practical for most companies. What most people do instead is they start with an open weights model and you modify them. And there's lots of ability to customize there. You can tune weights with techniques like fine-tuning, reinforcement learning, and you can try to build something that really focuses on your use case. However, it turns out there's actually limits to how effective this is as well. It's really hard to teach a model a completely new domain that it wasn't already pre-trained on. And it turns out the more you customize, the more you add a bunch of data in post-training, these models tend to forget some of that interesting stuff that it learned earlier, the core reasoning. It's a little bit like humans trying to learn a new language. When you start when you're really young, it's actually relatively easy to pick up. But when you try to learn a new language later in life, it's actually much much harder. Model training is kind of like this, too. Now, there's been some pretty cool things done with the ability that you have to tune these open weights models, but you can only go so far. Today, you just don't have a great way to get a frontier model that deeply understands your data and your domain. But what if it was possible? What if you could integrate your data at the right time of training during the training of a frontier model and then create a proprietary model that was just for you? I think this is actually what customers really want. And so we asked ourselves, why not? And today I'm excited to announce Amazon Nova Forge. Nova Forge is a new service that introduces the concept of open training models. With Nova Forge, you get exclusive access to a variety of Nova training checkpoints and then you get the ability for you to blend in your own proprietary data together with an Amazon curated training data set at every stage of the model training. This allows you to produce a model that deeply understands your information all without forgetting the core information that the thing has been trained on. We call these resulting models Novellas. And then we allow you to easily upload your Novella and run it in Bedrock. Let me show you how this works. Let's say you're that hardware manufacturer that we discussed earlier. You have several hundreds of gigabytes of data, billions of tokens related to your past designs, your failure modes, your review notes, etc. And you decide that you're going to start from an 80% pre-trained Nova 2 Lite checkpoint. Using our provided tool set, you blend all of your data in with that Amazon curated training data set. And then you run the provided recipes to finish pre-training that model, but this time with all of your data included. This introduces your domain specific knowledge all without losing the important foundational capabilities of the model like reasoning. Nova Forge also provides the ability to use remote reward functions and reinforcement fine-tuning to further improve your model, letting you plug real world environments into the training loop. And because your baseline model already understands your business, these post-training techniques are actually much more effective. Once you're ready, you import this model, your Novella, into Bedrock and you run inference on it just like you would any other Bedrock model. Now, your industrial engineers can ask questions like, "What are the pros and cons of design A versus design B?" and get responses that are specific to your company's historical results, manufacturing constraints, and customer preferences. We've already been working with a few customers to test out Nova Forge, and they're already seeing transformative results from customizing Nova's open training models. Let's dive a little bit into the example with Reddit. Reddit uses Gen AI to moderate content for multiple different safety dimensions across their chats and searches. Fine-tuning existing models didn't get them the performance they needed. They even tried using multiple models for different safety dimensions, but it was super complex and even then they couldn't get the accuracy they wanted for the specific requirements of their community. With Forge, however, Reddit was able to integrate their own proprietary domain data during pre-training, enabling the model to develop integrated representations that naturally combined the general language understanding with their own community specific knowledge. For the first time, they were able to produce a model that met their accuracy and cost efficiency targets. And at the same time, it was much easier to deploy and operate. We think this idea of open training models is going to completely transform what companies can invent with AI. Now, here to share how Sony is transforming and reinventing their business on AWS, please welcome Chief Digital Officer and Corporate Executive Officer from Sony Group Corporation, John Codera.

Thank you so much, Gerasan. Such a great story. And just like Sony has seen, a key to their success over the long run is having the right foundation for that innovation. What really excites me about Sony's story is that because they have their data and their applications in the cloud in AWS, it's that much easier for them to deal with any uncertainty that comes their way. Not many companies have been able to transition as successfully as Sony from an electronics device company into a global digital media business. Such a major change and so cool to see how having the right technology platform in AWS has helped them on that journey. Now, when you have your data in the cloud, it turns out you can move more rapidly and you can adjust to any of those unexpected changes that come your way. Now, the world is not slowing down. In fact, if there's one thing that I think we can all count on, it's that more change is coming. Now, one of the biggest opportunities that is going to change everyone's business is agents. Agents are exciting because they can take action and they can get things done. They can reason dynamically and they create workflows to solve a job in the best way without you needing to pre-program them. These agents work in non-deterministic ways, which is part of what makes them so powerful. But it also means that the foundation and tools that got us to where we are building software aren't necessarily the ones that we need for agents. And that's why we launched Amazon Bedrock Agent Core delivering the most advanced agentic platform so that you all could build, deploy and operate agents securely at scale. We designed Agent Core at the same time to be comprehensive but also modular. It has a secure serverless runtime so that agents can run in complete session isolation. Agent Core memory enables agents to keep context handling both short and long-term memory so that they can learn and get better over time. We provide an Agent Core Gateway so agents can easily discover and securely connect to tools, data, and other agents. We have Agent Core Identity that provides you a way to do secure authentication and gives you controls over what tools and data your agents can access. Agent Core Observability gives you real-time visibility into the deployed agent workflows that you have. And we have a variety of foundational tools that allow your agents to securely execute real world workflows. Things like Code Interpreter that gives you access to a secure code execution environment or our Managed Browser Service which provides a managed environment that makes it easy for your agents to access the internet. Agent Core is truly unique in what it enables for building agents and it's significantly different than anything else out there. And we built Agent Core to be open and modular. So you can use it with a variety of frameworks, things like Crew AI or Llama Index or LangChain or AWS's Trans Agent. You can also use it with any model out there. Whether it's from the variety of models that we have in Bedrock or from models like OpenAI's GPT or Gemini models. You only have to use the building blocks that you need. We don't force you as builders to go down a single fixed path. We allow you to pick and choose which services you want to use for your own situation. Agent Core also then makes it easy to deploy your agents privately and securely inside of your Amazon VPC and then allows you to scale to thousands of sessions to support high-traffic use cases. It's also super fast and easy to deploy your agents. Agents can be deployed in under a minute with just drag and drop or a few lines of code. And this is part of why we're seeing so much momentum with Agent Core as our customers rapidly adopt it as the foundation for their agentic applications. We see it across industries from companies like regulated industries like Visa or National Australia Bank or Rio Tinto. We see it from ISVs like Palumi and ADP or startups like Cohere Health and Snorkel AI and the momentum is really accelerating. I'm going to talk about a few of them. Now, Dina Freeman is the CEO of NASDAQ, and she and her team are moving really fast to build agents that can do real work in core areas of their business. Now, before Agent Core, they were planning on dedicating a whole team to build the foundation infrastructure that they needed to reliably operate and build resilient agents to meet their very high standards. Agent Core, however, now frees them from this heavy lifting so they can just focus on building great agents. Bristol Myers Squibb built a new agent that's able to evaluate more than 10,000 compounds across multiple hypotheses in less than an hour. This is a process that used to take their researchers four to six weeks. The company's drug discovery agent uses Agent Core runtime for its ability to seamlessly and dynamically scale and to keep their sensitive data secure and isolated. We see ISVs like Workday who are building the software of the future on Agent Core. Agent Core's Code Interpreter delivered exactly what they needed, the essential features and security requirements and data protection that was needed to power their planning agent. This capability reduces the time spent on routine planning analysis by 30%, saving them nearly 100 hours of work every month. And you don't have to build your own agents either. Many companies are using AWS Marketplace as the trusted place to publish and procure pre-built agents, tools, solutions, and professional services. And this is where AWS partners can help you all move even faster. We're really excited what customers have been able to do with Agent Core, but we're far from done. One big challenge that we've seen when you're building agents is how do you get them to behave predictably and in line with your intents? What makes agents powerful is this ability to reason and act autonomously. But that also makes it hard for you to have complete confidence that your agents aren't going to stray way out of bounds. This is a little bit like raising a teenager. I currently have two awesome teenagers myself at home. Now, as your kids get older, you have to start giving them more autonomy and freedom so that they can learn or adulting as they like to call it. But you also want to put some ground rules in place to avoid major issues. Think about when your kids start driving. This is the current situation that I'm in. All of a sudden, the kids have all this autonomy. There's a ton of things that they can go and do by themselves, but you still kind of want to have those guardrails in place, like you have to be home by a certain time or you don't want to drive more than say 5 miles an hour over the speed limit, things like that. One way you can actually build trust in agents is by making sure that they have the right permissions to access your tools and your data. Agent Core Identity provides a great way to do this today. But while permissions on your tools that your agents can access is a good start, what you really want to be able to control is the specific actions that your agents can or cannot take with those tools. Have actions like what is the agent going to do with those tools? How can they use them? Who are the tools for? And today customers struggle with this. You can embed policies directly in your agent's code, but because agents generate and execute their own code on the fly, these safeguards are really best effort and can only provide you weak guarantees. And they're really difficult to audit. In practice, this means today you can't with certainty control what your agent does or does not do while also giving it the agency to go and complete these workflows on its own. As a result, most customers feel that they're blocked from being able to deploy agents to their most valuable critical use cases. And today, that's why we're announcing Policy in Agent Core. Policy provides you with real-time deterministic controls for how your agents interact with your enterprise tools and your data. Now you can set up these policies that can define which tools and data and agents, which tools that your agents can access but also how they access them. So whether they're APIs or Lambda functions or MCP servers or popular third-party services like Salesforce or Slack, you also can then define what actions they can perform and under what conditions. So what Agent Core does is it then evaluates every single agent action against this policy before ever granting access to your tools or your data. We'll walk through a simple example. Let's say you're in Agent Core policy and just using natural language, you define a policy. Say something like, I want you to block all refunds from customers when the reimbursement amount is greater than $1,000. Then under the hood, what happens is your prompt is converted to Cedar, which is a popular open source language that's powered by our automated reasoning work across authorization and our verifiable systems inside of AWS. Once established, these policies are then deployed to your Agent Core Gateway and they're evaluated in milliseconds, which ensures that all of your actions are checked instantly and consistently to keep your agent workflows fast and responsive. And the design of where this sits is actually super important because this policy enforcement is outside of your agent's application code. The policy evaluation actually sits in between your agent and all of your data and your APIs and tools. So you can predictably control their behavior. Going back to our example in the refund policy, if every agent action is checked against your policies before the agent is able to access the tools. So let's pretend a situation happens where a refund is over the limit that you've defined. The agent is blocked from then issuing that refund. Now that you have these clear policies in place, organizations can much more deeply trust the agents that they're building and deploying knowing that they'll stay within the boundaries that you've defined. Now of course you all need agents to do more than just follow the explicit rules that you define. You have to know that they're behaving in the right way. Trust but verify is a phrase that we've kind of co-opted at Amazon as a mental model for how you manage at scale. At AWS, we give our teams incredible autonomy. I trust our teams to go and invent for customers and execute on that mission. But I also have mechanisms that allow me to dive deep and inspect when things are on track. I want to check that our strategic initiatives that we've identified are in fact getting done in the way that we've intended. If I go back to our teenagers, I generally trust that they're following the rules, but I can still check my Ring camera to ensure that they got home on time. And I can always check the status of my Life360 app to ensure that they're within the bounds of where I expect. The same thing applies to agents. To gain confidence, you want visibility into how they're acting. Now, customers love what they're getting with Agent Core Observability. You get real-time visibility into all your operational metrics. You can see your agent response times. You can see the computational power that's being used and your error rates and which tools and functions are being accessed. That's all great. But in addition to how agents are performing operationally, there's other things that you actually want to know. You want to know things like, are your agents making the right decisions? Are they using the best tool for the job?

Are their answers correct and appropriate? Are they even on brand? These are things that are super hard to measure today. It usually actually requires you to have a data scientist. The data scientist is going to build some complex data pipeline. They're going to select a model that's going to try to judge the outputs of their agents. They have to build the infrastructure to serve these evaluations and then manage quotas and throttling. And each time you want to roll out a new agent or you want to upgrade to a new version of a model that you're using, you have to do all of this work all over again. But unlike traditional software, actually testing and pre-pro here even then is really hard. You only know how your agents are going to react and respond when you have them out there in the real world. That means you have to continuously monitor and evaluate your agent behavior in real time and then quickly react if you see them doing something that you don't like. We think we can make this a lot better. Today, I'm excited to announce agent core evaluations. Evaluations is a new agent core service that helps developers continuously inspect the quality of their agent based on real world behavior. Evaluations can help you analyze agent behavior for specific criteria like the ones I mentioned, correctfulness, helpfulness, harmfulness, and they come with 13 pre-built evaluators for common quality dimensions. Of course, you can always create your own custom scoring system with your own preferred prompts and models as well. And you can easily evaluate agents in this testing phase to correct any issues before you end up deploying them broadly. So now if you're going to upgrade to a newer version of a model as an example, you run your evaluations to evaluate your agent and you want to make sure that it maintains the same level of helpfulness for example that you have in your current release. You can actually also use evaluations in production to catch any of those hard to find quality degradations really quickly. You'll see your results in CloudWatch right alongside your agent core observability insights. And agent core evaluations automates what used to take specialized expertise and a bunch of infrastructure heavy lifting into something that everyone can access and allows you to continually improve the quality of your agents. We're quite excited about it. All right, so this is agent core, the agentic platform that's powering the next wave of agents. We're helping you move quickly to get your agents into production without compromising or making any sacrifices, which is what we're all about. We want you to move fast so you have the broadest set of capabilities to build for your own customers. Today, we're really excited that we've added two new powerful capabilities in policy and evaluations. And I'm really excited to see how this unlocks some real powerful production use cases. Now to tell us more about how they're building agents of their own to transform their business and how they're using AWS as a key part of their agentic transformation, please welcome Shantanu Narayen, CEO and chair of Adobe.