Oracle AI World 2026 Keynote: Now That AI Works, What Are We Going to Do With It?

[00:00:00] *sad music* [00:00:30] *sad music* [00:00:50] *sad music* [00:00:54] There's always been a gap between what we can imagine [00:00:58] and what we can make real. [00:01:00] Every major advance in technology has narrowed that gap [00:01:04] because ideas were never the problem. [00:01:08] It was everything between the idea and the outcome. [00:01:12] The expertise, the complexity, the cost, the time, the tools. [00:01:18] But today, that gap is smaller than ever before. [00:01:22] We can learn faster. [00:01:24] Build better. [00:01:26] Bigger. [00:01:28] Safer. [00:01:30] Try things that once felt out of reach. [00:01:32] And spend more of our time on what people bring best. [00:01:36] Curiosity. [00:01:38] Creativity. [00:01:40] Ambition. [00:01:42] Solutions. [00:01:44] So what happens when technology can keep pace with more of our imagination? [00:01:50] We're about to find out. [00:01:54] Please welcome to the stage, Clay McWark. [00:02:00] Today, I was here, uh, in this room, uh, listening to a very interesting set of musical acts, uh, before, before Mike showed up. [00:02:10] And I was very thankful that I didn't have Mike's job yesterday, of starting first. Um, but, uh, during that process, I was sat in the front row and I realized two things. Um, the first is that I realized that all of the music being played was the music of my youth. [00:02:26] And just how old I am getting. Um, um, um, the second thing I realized is that I was like, this music is way too loud. And I was like, I am becoming my parents. Uh, so you might be wondering, well, how does that work? [00:02:52] Uh, so you might be wondering, well, how's that relevant to you? You're thinking, why do I need to sit here and listen to a random tech executive have their midlife crisis on stage at nine o'clock in the morning? Touche. Reasonable. [00:03:08] But if you cannot come to Vegas, the city of love and vulnerability and share with 10,000 of your closest friends, your biggest worries, where else can you do it? [00:03:21] Okay, moving on. Uh, look, we find ourselves, uh, constantly surrounded by AI. Uh, we're talking about it. There's AI content everywhere. And when you're in the middle of something this big, I think it's very easy to lose perspective of what's going on around you. [00:03:40] Right. It was just two years ago at Oracle that we made the decision to change the name of this conference, which previously was Cloud World, and we made the choice to rename it AI World. And I think that choice was the right one. At the same time that we made the change of the name, we also made a claim that AI changes everything. [00:04:00] Now, my goal for today is to show you just how true that statement is. Now, I'm going to do that by presenting evidence across every aspect of our business. Along the way, I'll cover many significant products that just would not exist without AI. And I can tell you that every single one of the things we're going to talk about today would not have happened nearly so quickly without AI helping us build them. [00:04:24] So, first, AI touches how all work is done at Oracle. Across software development, operations, finance, HR, legal, every single function of how Oracle works today is changed significantly by AI. [00:04:41] AI also changes the products and the services that we deliver. This is true across applications, our database, cloud infrastructure, obviously. [00:04:51] Now, to convince you of this a bit, I want to take a look at some concrete examples about the size of this impact that's happening across Oracle. [00:05:01] Now, the average time for us to resolve a support request has reduced from three and a half hours to just 16 minutes. [00:05:09] The time it takes for us to go from a product idea all the way to production has been reduced from more than nine months down to just six weeks. [00:05:19] We now discover and remediate vulnerabilities in software and service configurations ten times faster than we did just six months ago. [00:05:27] And today, I can tell you categorically across the board, the vast, vast majority of software that's written at Oracle is being written by AI, not humans. [00:05:38] If we look back to 2023, traditional compute that was being installed was much bigger than the AI compute being installed. [00:05:46] Now, since then, there's obviously been a massive increase in total compute, both traditional and AI accelerated. [00:05:52] But today, the AI compute that we install dwarfs the traditional compute that we're installing. [00:05:57] Now, we've covered some small examples and I want to dive into some bigger areas. [00:06:04] You know, you may think that's great, not convinced yet. [00:06:07] I want to talk about media, security, infrastructure, and then the data ecosystem. [00:06:12] For each of these, we'll see some new announcements only made possible by AI. [00:06:16] Let's take a look at media first. [00:06:19] So, many of you might be thinking, I didn't realize that Oracle was participating in the media space. [00:06:26] And traditionally, that has not been an area we focused on. [00:06:30] That's now changing. [00:06:32] And it's changing because AI is fundamentally changing how media is created, how media is delivered, and how media is experienced. [00:06:41] Now, to deliver on this vision, we've created a product called Oracle Media Fabric, which is a set of applications that work together. [00:06:48] It includes AI media models and Creative Studio to deliver new creativity tools. [00:06:53] Video Edge is our new delivery technology. [00:06:56] And our adaptive media platform changes how audiences engage with content. [00:07:01] We'll dive into each of those. [00:07:03] First, our media models form the foundation of our Creative Studio. [00:07:08] Now, these models start with trusted data. [00:07:10] It's incredibly important that you have confidence that when you use these models, the data that comes out of it can be used for your purposes. [00:07:17] It supports both still image and video generation. [00:07:20] And also critically important is that you can customize these models with your own content. [00:07:26] And that resulting model is exclusive to you. [00:07:29] Oracle Creative Studio reimagines the entire production process. [00:07:33] The workflows are redesigned around AI. [00:07:36] You use the models to enhance every aspect of the creative process. [00:07:40] Now, not only does Creative Studio obviously integrate with our AI media models, but it integrates with other leading models from other providers. [00:07:48] So to see how that works, let's take a quick look. [00:07:51] Generative AI is expanding who can create and the scale of what they can make. [00:08:00] Individual creators can bring bigger ideas to life. [00:08:03] Studios can develop more stories faster. [00:08:06] Oracle Creative Studio connects that ambition to production. [00:08:10] You get best-in-class AI models, including the Oracle AI media models for images and video. [00:08:16] Creative Studio helps ensure data provenance, IP rights management, collaboration, and production intelligence across your tools and systems while helping control costs and quality. [00:08:26] Connect stories, characters, scenes, assets, and more. [00:08:31] Discover what you have and what it could become. [00:08:34] Turn a script into a storyboard. [00:08:36] Animate it. [00:08:37] Refine the action before production. [00:08:39] Bring performers from a green screen set into any world you wish, such as a sailboat, a rooftop, a space station, or a farmer's market. [00:08:47] Then re-imagine the same scene with a different cast. [00:08:50] From studio productions to the next TikTok masterpiece, creators stay in control. [00:08:56] More ideas explored. [00:08:57] More stories brought to life. [00:08:59] Oracle Creative Studio gives more people the power to make what they imagine. [00:09:04] Now, in case you were worried, zero grandmothers were injured in the creation of that video, all thanks to the power of AI. [00:09:18] Okay, Oracle Video Edge is a different type of content delivery system. [00:09:22] This is no longer just a cache for your static content. [00:09:25] It also contains edge compute for content optimization and content generation. [00:09:29] This is extremely important for live events and for dynamic experiences. [00:09:34] The result is higher quality streaming and a new set of generation capabilities. [00:09:38] Let's take a look at how this works. [00:09:44] You've been there. [00:09:45] You're streaming the big event. [00:09:47] It all comes down to this moment. [00:09:50] And then the dreaded spinner. [00:09:52] Oracle Video Edge is designed to help keep you in the action. [00:09:56] It orchestrates video delivery across clouds, content delivery networks, and internet providers. [00:10:02] Optimized by logic running close to viewers at the edge. [00:10:06] With transport based on the emerging open standard media over quick, [00:10:10] a delayed packet doesn't have to hold up every frame behind it. [00:10:14] That means playback starts quickly. [00:10:16] When bandwidth gets tight, the system can prioritize audio and the video frames that matter most. [00:10:22] Helping you hear the commentary and get back to a clear picture as close to live as possible. [00:10:27] Oracle Video Edge brings a whole new generation of edge compute close to viewers. [00:10:32] Enabling completely new kinds of streaming experiences. [00:10:36] For the teams delivering the big event, Oracle takes care of the complexity of getting it right. [00:10:41] For you, it's content delivery built around what you're here for. [00:10:45] Enjoying the moment. [00:10:47] And then our adaptive media platform combines Creative Studio and our Video Edge capabilities together. [00:10:57] It enables experiences that were previously impossible. [00:11:00] Suddenly, we can all experience the same live event, but the actual delivery is customized for each of us. [00:11:07] We're used to the level of control and personalization that we get online. [00:11:10] An adaptive media platform brings those same capabilities to traditional media. [00:11:15] Let's see how that works. [00:11:19] The Oracle adaptive media platform is transforming the consumer experience, making it more relevant and more personalized. [00:11:27] It brings together the latest media technology and AI, connecting content creation, personalization, and delivery. [00:11:34] That means ads tailored to each audience and live viewing experiences that feel personal to every viewer. [00:11:41] During a live broadcast, relevance means customizing ads for different viewers in real time. [00:11:46] Start with one original ad. [00:11:48] With the Oracle adaptive media platform, teams identify what to change. [00:11:53] The product, environment, message, or call to action for a particular audience. [00:11:58] Choose from Oracle AI media models and other models to get the right result, generating ad variations for different viewers during the broadcast. [00:12:06] That means high quality ads at a fraction of the production cost. [00:12:10] Deliver the right ad at scale through Oracle Video Edge or distribute it through traditional ad networks using existing workflows. [00:12:18] With Oracle adaptive media, your streaming experience can be exactly what you want it to be. [00:12:23] You can join a watch party with your favorite celebrities and share the moment with fellow fans. [00:12:28] Behind the scenes, AI suggests trivia and other experiences tied to what's happening in the event. [00:12:34] Producers select an experience and prepare it for you. [00:12:37] The same feed can deliver a personalized ad with your name, your offer, and a QR code. [00:12:43] The Oracle adaptive media platform, making the consumer experience more relevant, more personalized, and more engaging. [00:12:55] Okay. [00:12:56] So Oracle media fabric offers something for everybody in the industry. [00:12:59] Creators can suddenly create higher quality content more quickly and at a much lower cost. [00:13:04] Broadcasters deliver a more engaging, higher quality live experience. [00:13:08] And then consumers get new tools to participate and the ability to customize their shared experiences. [00:13:14] Now, this is in limited availability today with some key customers, but we're launching this early next year. [00:13:20] Now, security is a space that is rapidly evolving. [00:13:26] And that's true across many different dimensions. [00:13:28] AI is just driving a significant shift in how we secure our digital systems. [00:13:33] And so to talk more about that, I'm going to bring to the stage Rob Clark, [00:13:37] Bruce Vice President of Security Architecture. [00:13:40] Come on out, Rob. [00:13:50] Rob, now the first thing you've got to do is adjust your pillows. [00:13:53] If you don't adjust your pillows properly, this whole thing is going to fall apart. [00:13:56] Okay. [00:13:57] So security, it's evolving rapidly. [00:14:06] I feel like AI has been changing a lot and security has had to adapt right along with it. [00:14:13] We've gone through multiple transitions. [00:14:15] Why don't we start kind of back at the beginning when really we were just using AI to answer questions for us. [00:14:20] How did that have an impact on security? [00:14:24] I think one of the most interesting things on that journey was learning that new technologies can bring back old problems. [00:14:34] Initially, LLMs were put in support type roles where they were given access to internal documents and training. [00:14:42] The same sorts of things that your support staff would have access to. [00:14:46] And pretty quickly, bad actors realized they could convince these models to disclose secrets and discount codes. [00:14:53] Say things that you as a company would never say to your customers. [00:14:59] And this was really an old security lesson, which was don't trust user input. [00:15:05] And don't give frontline systems access to sensitive information. [00:15:09] That makes sense. [00:15:10] And then models continue to evolve. [00:15:12] And suddenly, we're using these new AI, you know, amazing AI models to write software. [00:15:17] That just fundamentally changes the risk profile. [00:15:19] So what do we do in that scenario? [00:15:21] Well, again, it sort of came down to trust. [00:15:23] But now we realize that we can't trust code, even from developers that we know and trust that have been through our training. [00:15:30] And we actually learned this with an early wake-up call where some AI included a malicious package in something that was destined for production. [00:15:38] This actually targeted our build systems seeking to steal keys and secrets and that sort of material. [00:15:45] We caught it with our supply chain security systems. [00:15:48] But it was a really stark reminder that even your build and test systems now can't trust the code that's coming in, even if it's from a developer that you trust. [00:15:57] Makes complete sense. [00:15:59] And then we kind of move forward to the next step where reasoning models were suddenly much more capable. [00:16:04] And we thought, I know what we should do. [00:16:06] We should get those same models access to all the tools. [00:16:09] Left unchecked, seems like it's going to create a lot of problems. [00:16:13] What do we do about that? [00:16:14] Well, we knew by now that trust was key, right? [00:16:17] Right. [00:16:18] So we wanted to make sure we didn't have to trust these things at all. [00:16:21] We built a strong sandbox environment with virtualization, instrumentation, really detailed logging. [00:16:27] It was really very secure. [00:16:29] That sounds amazing. [00:16:30] How did that work out? [00:16:31] Not very well at all. [00:16:32] Our developers really didn't like using it. [00:16:34] It introduced way too much friction and took away almost all of the benefits. [00:16:38] Okay. [00:16:39] So basically it was choose between giving access to a bunch of tools and it going, do whatever it wants, or you could have it safe and secure and useless. [00:16:48] That was where we got to at that point. [00:16:51] Okay. [00:16:52] So if we don't like either of those options, what's the actual solution? [00:16:55] So we partnered with our AI vendors on new types of enterprise controls. [00:17:01] This meant that there was no extra stacks to install. [00:17:04] Things were better supported. [00:17:06] We got enterprise visibility and control across the entire real estate. [00:17:10] And each time a new model or feature was brought to us by one of these vendors, as long as it was supporting the controls we'd agreed with them, we could roll it out across the entire company and it led to much happier developers. [00:17:23] Okay. [00:17:24] So now we have happier developers, which is always important. [00:17:27] It feels like we've transitioned to a world where AI can suddenly discover vulnerabilities and exploit them. [00:17:34] For a lot of people that can be very alarming. [00:17:36] Where do we go from here? [00:17:38] I think it's actually going to be a difficult couple of years, but it's important to remember that we as builders have the advantage. [00:17:45] So we use the same models that can get used by some of these bad actors, but often we get much earlier access. [00:17:52] We have access to the source code and the architecture documents and all the things that will guide those tools to be more effective than they would be outside. [00:17:59] Okay. [00:18:00] So we've got some techniques, but how do we actually use those to help ourselves and to then help customers? [00:18:04] So within Oracle, we built a machine to constantly find and fix issues. [00:18:10] It's a loop, a virtuous cycle of hypothesizing based on source code, validating in the real world against our test machines, building patches. [00:18:19] And this runs 24/7 across our entire stack. [00:18:23] And one of the ways it's really advantaged us is we've found several issues in cloud that we were able to patch using technologies like Oracle Case Spice, creating zero downtime for our customers. [00:18:35] Okay, so suddenly the rate of change is an order of magnitude more than it was previously. [00:18:40] How can a typical company actually keep up with this extremely high rate of change? [00:18:47] It is difficult. [00:18:48] I think the lesson is to move to cloud where you can and shorten your pre-patch validation where you can't. [00:18:55] There is now more patching than ever before. [00:18:59] Companies are learning how to use agents to accelerate their testing and validation for use on-prem. [00:19:04] And they're prioritizing moving to cloud where obviously this patching is done for you. [00:19:10] We're investing to make it easier for customers to migrate to cloud and working on some new security tools that span cloud and on-prem and really help customers prioritize what to fix when. [00:19:22] Okay, so great. [00:19:24] What do you think comes next? [00:19:25] Where do we actually go from here, Rob? [00:19:27] I'm actually quite optimistic about the future. [00:19:31] As much as it's going to be hard, I think we are entering into a new era of higher quality software. [00:19:37] You know, these bugs, they're not new. [00:19:40] The AI didn't put them there. [00:19:42] And soon I think they'll be gone forever. [00:19:44] You know, we build these tools, this machine that I mentioned earlier for finding and eradicating them. [00:19:48] Those same mechanisms we use across our entire software development lifecycle to make sure they can never be reintroduced. [00:19:54] So I actually believe that the future is fewer patches and more secure software. [00:19:59] Okay. [00:20:00] Thank you, Rob. [00:20:01] I appreciate you coming out. [00:20:02] It's clear that AI requires us to change how we approach security. [00:20:14] A world of agents requires new ways to authenticate and authorize. [00:20:18] And this is true across both agents as well as how we think about human beings, how they authenticate and authorize. [00:20:26] So we live in a world where AI is increasingly capable of impersonating humans. [00:20:32] Many times we actually need confidence that the person we think is doing something is actually the person that we think is doing something. [00:20:40] When you think about something like a human in the loop approval, it only works if there's a real human in the loop. [00:20:48] So to help with that, we've added biometrics capabilities. [00:20:52] These are things like facial recognition, voice recognition, fingerprint recognition, and other methods to Oracle identity and access management. [00:20:59] The use cases for this are quite broad. [00:21:02] They span helping you to verify your employees, confirming sensitive actions, as well as approving agent tasks. [00:21:10] And this only becomes more relevant as more and more work is done by agents. [00:21:15] Now, agents take on whatever identity that we give them. [00:21:20] I'm sure all of us have some friends. [00:21:23] I have some of my own. [00:21:24] Not that I have friends. [00:21:26] But I have some friends that they give agents access to their laptop and their email. [00:21:34] And suddenly they find themselves, you know, they talk about it's really cool. [00:21:38] But then later on, you know, they're arguing with their AI agent about did I send the drunken email at night? [00:21:45] Did you send the drunken email at night? [00:21:46] There's confusion. [00:21:47] They don't know how they got there. [00:21:48] I think overall it's a bad idea. [00:21:50] It becomes very confusing about who did what. [00:21:53] Instead, I think we need a new identity system that really operates at agent scale. [00:21:59] What that means is that you need the ability to create an enormous number of identities very quickly. [00:22:04] You need to be able to also delete them. [00:22:06] To support that, we've added agent identity into Oracle Identity and Access Management. [00:22:12] Now, once agents have an identity, we also need ways to control their actions. [00:22:18] You need ways to start with a little tiny amount of access and then step up that access [00:22:22] when it's necessary. [00:22:23] You need access to be minimal and then also be short for a small period of time. [00:22:28] And you need the ability to specify that some actions require special approval, sometimes specific human approval. [00:22:35] Combining biometrics, agentic identity, and just-in-time access is really the only way to ensure agents act responsibly. [00:22:43] Now, if we move beyond identity, we also need new tools for finding and fixing vulnerabilities. [00:22:49] As Rob shared with us, agents can discover as well as fix these vulnerabilities. [00:22:55] To help with this, we're launching a new product called Oracle AI Security Mesh that helps customers use cyber models [00:23:01] to improve the security across all other systems. [00:23:03] We've focused on lessons that we've learned at Oracle, securing our products and our cloud systems. [00:23:09] AI Security Mesh provides a platform for analyzing all of your systems, discovering vulnerabilities, fixing them, and then releasing that to production. [00:23:16] And for that, let's take a look at how it does that. [00:23:23] Oracle AI Security Mesh enables any company to use advanced AI models to find vulnerabilities and deploy fixes before attackers can exploit them. [00:23:32] All with no AI expertise required. [00:23:35] Here's an example: Northstar Manufacturing wants to reduce risk across their enterprise. [00:23:40] Security Mesh discovers Northstar's infrastructure and applications across its on-premises and cloud environments. [00:23:47] It identifies business-critical systems that might be a target. [00:23:51] From there, specialist AI agents use advanced cybersecurity models, source code, and the Mesh's knowledge of Northstar's environment to test potential attack paths. [00:24:01] Here, we can see them find serious problems. [00:24:04] Someone who isn't signed in can read from the ordering API, and a configuration issue in that service gives it access to other systems, letting the attacker reach sensitive data. [00:24:13] A separate agent then validates the complete attack path to confirm it's real, and packages those findings and a recommended fix into a report. [00:24:21] A coding agent uses the report to fix the API. [00:24:25] The Mesh updates the application's cloud configuration so it can reach only the services and data it needs. [00:24:32] After both changes are deployed, we confirm the API block works. [00:24:41] And because the system learns from each finding and works with you to fix the issues, future security testing and recommended fixes become smarter. [00:24:54] Let's hope that AI security mesh was not responsible for that video. [00:24:59] All right. [00:25:00] AI is enabling us to make much more secure systems. [00:25:03] Better authentication, better authorization, and lockdown access have always really been best practices, but now we can actually deliver on them. [00:25:11] And finding vulnerabilities yourself and being able to generate fixes quickly is far preferable to the alternative. [00:25:18] I think most people would agree that AI is rapidly changing infrastructure. [00:25:23] You know, kind of behind the scenes, software that we're using to run it, the way in which we do operations, all of that is changing. [00:25:30] But most clearly, it's changing what we actually build. [00:25:36] And so what I want to do is take a few minutes to look at just how much has changed recently in infrastructure because of AI. [00:25:43] So, NVIDIA has been the clear thought leader in AI accelerators for quite some time. [00:25:50] From the launch of P100 in 2017 to H100 in 2023, everything got bigger. [00:25:56] Power density increased, network scale increased, cluster sizes went up, right? [00:26:01] You move forward suddenly to 2024, and AMD decided to participate with the launch of MI300. [00:26:10] Grace Blackwell ushered in a new type of system in 2025. [00:26:13] Suddenly, NVIDIA has created true rack scale compute, fully integrated within Vlink, right? [00:26:19] You have an AI optimized system across all layers, from power design, through networking, CPU, and across the GPU itself. [00:26:27] At this conference a year ago, we launched Acceleron, which is really all about providing higher performance and larger scale AI networks. [00:26:36] And one of the key pieces of that technology is multi-planar back-end networks that provide better reliability and higher performance. [00:26:44] Now, that brings us to the present day. [00:26:47] So, I'm very excited to say that we have begun delivering our first Verirubin systems to customers. [00:26:54] Verirubin is by far the most powerful system that we've ever deployed, and we see an immense amount of demand for those systems. [00:27:00] Google is another hardware provider that's been in this game for a long time. [00:27:05] They've been building their own hardware for many, many years, and we're excited to bring TPUs to OCI, starting with TPU V8. [00:27:14] We're also adding many other accelerators. [00:27:16] AMD's 455 is their latest rack scale offering. [00:27:20] Positron Atlas provides a specialized inferencing option focused on power efficiency. [00:27:24] You know, we obviously spend a lot of time with OpenAI, and in conjunction, they've done a great job with Broadcom and their Jalapeno chip, which is really co-designed around their AI models and custom silicon. [00:27:37] And then Cerebris and Sambanova both provide extremely low latency inferencing options. [00:27:44] I'm also excited to talk about some new things that we're launching with Acceleron this year, right? [00:27:49] We've taken that same technology that we put into our backend network, and we're really bringing it to the front end. [00:27:54] So suddenly, we're offering multi-planar front-end networks. [00:27:57] What that gives is it provides better resiliency, increased performance, but it also, thankfully, gives you RDMA over your front-end network, which is critically important when you add together our new storage offerings. [00:28:11] So first, we have a great new partnership with NetApp, where we're bringing a first-party NetApp offering into OCI. [00:28:17] NetApp's been a long provider for both enterprise and AI-focused storage offerings, but we also have a new partner program for AI accelerated storage for companies like Weka, Vast, and DDN that is integrated into Acceleron. [00:28:30] And what that means is that you can now talk to your storage via Acceleron using RDMA across all of these different AI accelerators. [00:28:39] Now, across Oracle, we've long invested in multiple types of regions, from the smallest dedicated region, which is as small as only three racks, to the largest AI regions that span millions of square feet and campuses that are thousands of acres. [00:28:55] AI is changing where we build regions and also how we build them. [00:29:00] To help me kind of explain some of the interesting changes that are going on, I'd like to welcome Pradeep Vincent, Executive Vice President of Compute, to the stage. [00:29:08] I hope you were watching, Pradeep. [00:29:17] Remember, you've got to start with the pillows. [00:29:18] Rob did great, and the pillows did all the work for us. [00:29:22] Yes. [00:29:23] They can't even see the pillows. [00:29:25] That's the best part. [00:29:26] Okay. [00:29:27] Pradeep. [00:29:28] AI has different requirements, right? [00:29:32] Suddenly, when it comes to infrastructure, we're building much larger sites. [00:29:35] Why is that? [00:29:36] And what areas does it have the most impact on? [00:29:40] Yeah. [00:29:41] So if you think about AI infrastructure, we really need to think about, you know, the entire AI infrastructure as one big computer. [00:29:50] That's really the key difference. [00:29:51] So with the AI infrastructure, we put a lot of high-performance computers, and we put it in a very small, you know, physical space, if you will, and connect them with extremely powerful network. [00:30:03] And that essentially drives us to think about power, cooling, and networking all in very different ways. [00:30:10] Okay. [00:30:11] So, right, I just talked a little bit about Acceleron. [00:30:15] We launched it last year. [00:30:16] We've added some great new capabilities. [00:30:18] Why is Acceleron so core to our overall AI strategy? [00:30:22] Well, it's all about networking. [00:30:23] So when you have, you know, an AI cluster, particularly for training, you know, the way the training algorithms work, if there's a single GPU that's actually stuck behind a network packet that's delayed, the entire cluster slows down. [00:30:39] As an example, if there's a one-second delay for one GPU, let's say, you know, 100,000 GPU cluster, the entire training run now sees 28 GPU hours of delay, just for one second for one GPU. [00:30:53] That is a really big traffic jam, if you think about it. [00:30:55] Okay. [00:30:56] So, I don't even want to think about what a longer delay would look like. [00:31:00] The GPU hours wasted add up very quickly. [00:31:03] How does Acceleron make this better? [00:31:05] Yeah. [00:31:06] Yeah. [00:31:07] So, if you think about networking, you know, it's always been the case, there's always redundant paths in a network, right? [00:31:12] So, when there's a congestion or some failures, you know, we always have redundant links in the network so that the packets can take a different route. [00:31:20] But with Acceleron, we take that to the next level. [00:31:22] We not only have redundant links, now we actually have redundant networks itself. [00:31:27] So, all GPUs are connected to, you know, multiple network planes, and each network plane is actually independent fully. [00:31:35] So, if there's any problem with any of the network planes, you know, the NICs or the GPUs on the NICs, they essentially decide to read out the packets to a completely different network. [00:31:45] Okay. [00:31:48] Yeah, I still think multi-planar networking is very, very cool. [00:31:51] But moving on to a new topic, one area that I think is underappreciated in AI infrastructure is power generation and conditioning. [00:31:59] I think I remember a few years ago, I feel like the both of us also did not appreciate some of the complexities in that space. [00:32:07] What's so different about AI workloads when it comes to power infrastructure? [00:32:11] Yeah, so if you think about how power is used in AI infrastructure, you know, there's really two ways to, you know, think about the impact. [00:32:19] The sum of it is actually, you know, not even related to AI, it's any large workload on a power grid has some challenges. [00:32:27] And then there's also challenges, you know, specific to AI. [00:32:30] Now, if you think about a grid, you know, you have, you know, power generators and then you have power consumers. [00:32:36] And, you know, at the grid, there's always a little bit of a tug of war going on, but except the goal of this tug of war is the opposite. [00:32:43] We want to keep both sides stable. [00:32:45] We don't want to pull the other side down. [00:32:47] So, you know, when you have, let's say, you know, power generation or power plant that, you know, goes offline because of some problem, the grid usually sees some disturbance. [00:32:56] Now, you would have seen that in the form of perhaps light dimming in your house, and then it eventually, you know, comes back up. [00:33:04] Now, as a large consumer of a load, if we overreact, then what happens is, you know, that grid suddenly from a low load condition, it swings back to a high load condition, and then it, you know, kind of destabilizes the grid. [00:33:17] So, when we actually remove the load from the grid, it's very important that we actually do it in a responsible fashion, in a slow methodical fashion, in coordination with the grid, and not, you know, abruptly remove the load, if you will. [00:33:29] Got it. So, kind of, just as a large consumer of electricity, this is what you have to do, but that's not really, I mean, it's true that AI workloads are very large, but we've had AI, like, large data centers for a while, and in that sense, it's no different. [00:33:42] So, what is it about these AI workloads that then add even more complexity when it comes to being a good power consumer? [00:33:50] Yeah, so, you know, the AI does have its own implications. So, let's take the example, you know, imagine you have a space heater, and the space heater consumes one kilowatt of power, if you will. [00:34:01] Imagine you actually keep turning the space heater on and off repeatedly, over and over again. Well, that's okay for a grid, because the one kilowatt space heater in your home is not that big compared to the, you know, grid itself. [00:34:14] So, a grid can handle that. Now, imagine you have one million space heaters across one million homes, and imagine you found a way to, you know, in a coordinated fashion, turn all the one million space heaters on exactly at the same millisecond, and then turn it off. [00:34:30] And then turn it on, and over and over again. Well, that's actually challenging for the grid, and they would actually notice. [00:34:37] So, yeah, I think if that happened, they would come, they'd find a way to come to all one million homes and tell you to please stop. [00:34:43] Exactly. [00:34:44] So, traditional IT loads are not coordinated in this way, and even though big, not coordinated. How do we actually solve this for coordinated workloads? [00:34:55] Yeah. So, we use a combination of, you know, large energy stores with a powerful and sophisticated power electronics, if you will, to control the energy stores. But energy stores can be a battery. [00:35:10] Sometimes it's also a battery combined with supercapacitors, where we need to have extremely high-frequency charge and discharge cycle. [00:35:19] So, imagine, you know, the GPU cluster, let's say, you know, 100,000 GPU cluster, suddenly ramps up on the load. [00:35:25] What we do is, you know, the energy system, now, instead of, you know, loading the, transferring the entire load to the grid, it actually provides some of the energy. [00:35:34] Right? And then the power electronics, you know, make sure that the ramp, as seen by the load, as seen by the grid, is actually fairly smooth and slow and steady, so that the grid can actually take it. [00:35:44] And same thing, when the cluster suddenly, you know, shuts down the load, because the training ramp job is done, or maybe there's a crash, or whatever it may be, the energy store now charges, starts charging, you know, on its own. [00:35:59] So, it essentially takes some load, and then the power electronics ensures that the load, as seen by the grid, is gentle overall. [00:36:05] Yeah, so what you're saying is, we just need to create, it's an electrical equivalent of a giant shock absorber for these. [00:36:11] Exactly, exactly. [00:36:12] Okay. Well, thank you, Pradeep. I really appreciate your time. Thanks for coming out. [00:36:15] Thanks. [00:36:16] So, every aspect of infrastructure, from power generation and cooling, to power conditioning, how and where we build these data centers, the size of the data centers, the network and the storage that we put inside of them, [00:36:33] and then, of course, the accelerators that we're putting inside these data centers, all of that is changing at a very rapid pace, because of AI. [00:36:41] Now, moving on to data, AI is changing the entire data ecosystem. [00:36:50] If we go back a few years, you know, it really started with structured data, and most people put their structured data in a relational database. [00:36:58] And then, over time, you know, we moved into the era of data warehouses, and people used more relational databases to handle analytical workloads. [00:37:06] And then we moved to the age of big data, where suddenly you could put all of your data, all your log files, everything unstructured, you just put it out there into a big data store, and then you would analyze over that. [00:37:16] AI needs access to all of that data, but it needs more than that. [00:37:20] It needs everything within your company, it needs the information that's in your wiki, the information that's in your messaging systems, it's in your email, everything your employees need access to, AI needs access to as well. [00:37:30] So, to help with that, we've given the AI data platform the ability to analyze enterprise data and then generate AI ontologies. [00:37:40] Previously, business context was encoded inside individual applications. [00:37:45] Instead, these ontologies provide AI the business context, which is critical for agents to answer complex questions, for them to write software, and for them to make decisions for you. [00:37:55] Along with this, we're also launching AI data platform innovation services, which is a set of forward deployed engineers that are there to help you implement these systems faster. [00:38:05] Now, Snowflake has been investing very heavily into AI for many years, across products like Cowork, Cocoa, and Cortex AI. [00:38:14] And we're bringing the Snowflake platform to OCI, so Snowflake customers can run Snowflake alongside their Oracle workloads. [00:38:21] Together, our goal is to ensure that customers get the most out of all of their enterprise data. [00:38:26] AI is also changing who or what accesses data. [00:38:31] Traditionally, that's been people with defined identities or software written by people. [00:38:36] Suddenly, it's agent-led with software written by AI agents or agents accessing data directly. [00:38:42] So, at Oracle, we've had a long history of adding every data capability to the Oracle database. [00:38:48] Most recently, we've created the DeepTrust AI architecture, which is really about taking data controls and pushing them down into the database to each individual table, each row, each actual column. [00:39:01] Suddenly, now, policies are enforced inside the database. [00:39:05] And the result is it provides much better protection, regardless of who's accessing the data. [00:39:10] Now, AI is also changing how data is delivered. [00:39:16] We've been working on this for quite a long time. [00:39:19] And so, to have a bit more dive into kind of what's changing around how we think that people will be experiencing data, I'd like to welcome Chris Rice to the stage. [00:39:28] Don't walk too fast, Chris. You're going to beat me. Don't worry if you go. [00:39:36] No. We'll move past the pillows. [00:39:39] Okay. Chris, I think you've been working on Apex for a very long time. [00:39:46] My guess is a lot of people here are not super familiar with what Apex is or why we built it or how it helps define a consistent foundation for applications. [00:39:55] Can you tell us a bit more about how that works? [00:39:58] Sure. So, Application Express Apex has always been governed to try to help the business user build apps. [00:40:07] So, what that means is we've tried to make it so that the business user can go to a nice web UI. [00:40:12] They can point and click. They can build the apps. [00:40:14] They can focus on the business problems and not necessarily the technology or frameworks or languages that are needed underneath. [00:40:21] And then after that web UI, there's actually a runtime that operates the entire application. [00:40:27] So, it's in a safe governed place. [00:40:30] Okay. So, the result of that, right, is that the business user can stay closer to the objective. [00:40:36] How does this apply in the world of AI coding agents? [00:40:40] Well, this is where it got really fun actually. [00:40:42] So, what we did is we took a step back and looked at what Apex provided and realized that the power was the runtime. [00:40:48] So, we took the definition of the application and we actually distilled it into a new language. [00:40:54] So, a language at the level of the application. [00:40:56] So, that then we can take that language and give it to the agents and the agents can stay within that governed safe runtime to actually turn around and build the applications. [00:41:05] Okay. I think we've got a video to take a look at how that works. [00:41:08] Yes. [00:41:09] This is an application for onboarding employees that was built using natural language to explain the desired functionality and then generated with Apex Lang, which is the definition of the application, as well as the database structures required. [00:41:26] Oracle Apex is a proven, trusted runtime running millions of enterprise applications for over 20 years. [00:41:34] This app is obviously hosting employee information and Apex knows that and has applied the right restrictions so that data is only accessible to other employees with the right permissions to see that data. [00:41:46] All without the application builder thinking about it. [00:41:48] This example shows how using Apex Lang and the Apex runtime empowers every enterprise to allow their employees to safely write applications in a way that minimizes technical debt, ensures applications are secure, limits access based on roles, and provides an Oracle maintained stack. [00:42:06] Okay. But Chris, I mean, AI can already generate applications. Why do I need this? [00:42:16] So we've all seen the applications that have been generated. [00:42:19] So one of the powerful things about this runtime is if you let the agent do something, it might build something local, it might use a file locally, it probably won't do integration with your identity yet. [00:42:32] Now, the minute you have that useful application that someone else needs to access, you have to integrate with identity. [00:42:38] Your IT government might kick in. Maybe there's some PII information in there. Maybe there's some salary information in there. [00:42:44] All of that is enforced at the application layer inside the database and that identity is pushed all the way down into the database so that it's in a nice, safe, governed place that enables all the corporate governance and monitoring metrics that need to be in place. [00:43:01] Okay. Fair. But I mean, like, why is Apex the right technology to standardize on versus like saying, hey, we're just going to pick some web frameworks and say, focus on Python or something like that? [00:43:16] So what would happen is you could do that and then you would end up actually instructing Python to do a lot of the things that this runtime provides. [00:43:24] So say your identity posture changes in your company and you vibe coded or generated one application. No problem. You update that one application. Now, say you enabled your enterprise and you have a thousand of these applications. Now, instead of changing one definition of the application at the runtime and let the runtime enforce it, you're actually now retrofitting a thousand Python programs, a thousand containers or even more. [00:43:48] Okay. So I feel like we've done a lot already, created Apex for a long time. You have Apex link, et cetera. What do you think comes next? And how should we think about this stuff beyond Apex? [00:44:00] I think the lesson we learned is that with using AI and code generation to build applications in general, and as you mentioned earlier on the email, you need to give the AI the ability to do the work that has been tasked for at the highest level of abstraction that it needs only. [00:44:18] You can't give it the lower level abstraction. If you're writing an operating system, maybe it needs a very low programming language. If you're writing a business application, you do not need that and you can stay at a layer like application. [00:44:30] Okay. All right. Chris, thanks for coming on. I appreciate the time. Thank you. So how do we take this to the next level? Well, I think it starts by enabling everyone to write applications. [00:44:46] And Oracle App Composer is a new product that we've created that does just that. It combines the ability that you get from things like codecs to write code with a platform that makes it easy to govern, control, and secure those applications within your enterprise. [00:45:02] And so now let's take a look at how Oracle App Composer works. [00:45:05] Oracle App Composer turns business ideas into useful apps. Consider Nina, who used codecs to build a marketing launch calendar for superior data systems. [00:45:20] She asks App Composer to deploy the app and share it with the entire marketing organization. [00:45:26] An IT approved profile defines libraries, security controls, and data rules. [00:45:34] App Composer checks the app and deploys it to marketing production in Oracle Cloud. [00:45:39] Nina's app has an internal company URL and company sign-on. She shares it with her team. [00:45:47] I am here with Clay McGork, the CEO. [00:45:51] Nina's marketing coworker, Leo, opens launch calendar to link a SharePoint document and Nina can see his contribution. [00:45:58] Clay, thank you so much as always. [00:46:00] Alex is an IT administrator at Nina's company, and he manages app profiles and deployment environments with company sign-on and monitoring. [00:46:09] Just because Alex manages the apps, that doesn't mean he is granted access to the business data they include. [00:46:16] AI Insights help Alex and other IT administrators spot overlapping apps and bring their owners together for review. [00:46:23] It flags unused apps still deployed, so administrators can review them for retirement. [00:46:28] And when an active app has no current owner, IT can assign an owner. [00:46:34] Oracle App Composer: Turn business expertise into enterprise applications. [00:46:41] Okay. [00:46:46] I was like, it's either going to be more clapping or less clapping. [00:46:52] We can't have this intermediate clapping. [00:46:53] That's not helpful. [00:46:54] Oracle AI Workspace is another product that combines all of your Oracle data. [00:47:00] stuff in your Oracle databases, OCI, Fusion, NetSuite, your industry applications, with all of your non-Oracle data. [00:47:07] Stuff in Slack, Office 365, your Wiki, et cetera. [00:47:10] And it provides a unified experience to answer questions and then drive workflows across your entire enterprise. [00:47:16] Now, we're very excited about this product. [00:47:18] We've built it in deep partnership with OpenAI. [00:47:20] Let me show you how this one works. [00:47:22] Oracle AI Workspace is an AI experience connecting your business through one interface. [00:47:34] I've received an alert from Workspace. [00:47:36] A supplier delay threatens 70 control units for a key customer. [00:47:41] Northstar is opening a new distribution center in two weeks. [00:47:45] They need this order by Friday to stay on schedule. [00:47:48] Across Oracle applications, data, and third-party tools, Workspace automatically finds a plan to save the order while meeting our business requirements. [00:47:57] No custom wiring needed. [00:47:59] Workspace coordinates Oracle AI agents to identify a qualified replacement product using specifications from our company's Oracle AI database. [00:48:09] It also brings in the customers receiving requirements from our meeting transcripts [00:48:16] and warehouse availability from our team chat. [00:48:21] With this context, Workspace builds a recovery plan. [00:48:24] It uses Fusion agentic applications and Fusion Claw, an AI engine for complex tasks, to plan a cost-effective delivery. [00:48:37] Workspace checks funding and proposes order changes and Fusion. [00:48:41] Everything changes without the required approvals. [00:48:44] Once approved, Oracle executes the plan and we are able to save Northstar's opening while maintaining our margins. [00:48:51] Before, we would spend hours coordinating a recovery and still miss critical details in chat and email. [00:48:58] From here, I can use Oracle App Composer to turn this workflow into an app and deploy it in OCI. [00:49:11] That gives my team a repeatable way to keep customer orders on track in the future. [00:49:24] Well, it's good to know that as an audience that we're learning as we go. [00:49:28] Something else we're very excited about, we want to make it easy for new businesses to adopt and make use of these new technologies. [00:49:36] To help with that, we've bundled together NetSuite Quick Start, which is our latest new financial software focused on startups, [00:49:42] along with Oracle credits for both OCI and OpenAI credits. [00:49:46] Combined, you get the full advantage of Oracle and OpenAI technologies. [00:49:50] Now, next, I want to kind of discuss more about how we've actually been able to build these incredible new technologies in partnership with OpenAI. [00:49:59] And for that, please help me welcome Sam Altman to the stage. [00:50:15] Thanks for coming, Sam. [00:50:16] Thanks for having me. [00:50:17] Well, look, you guys have been doing a lot. [00:50:24] The pace of innovation has clearly been incredible. [00:50:27] And I mean, I can tell you from my perspective, the rate of change of the models only seems to keep accelerating. [00:50:34] You launched GPT-6 in August. How's that been going? [00:50:38] The rate of change has been one of the biggest surprises to me this year. [00:50:42] I mean, we've talked about how fast it's been going for four years. [00:50:45] You know, we'd launched GPT almost four years ago. [00:50:50] And four years ago, our models couldn't solve grade school math. [00:50:52] And now they're like knocking down the hardest unproven math problems we know of. [00:50:57] But the last, say, six months have felt like a very accelerated version, even on that timeframe. [00:51:04] And GPT-6 was one of those moments where I think the models took a step forward all at once, both in terms of their ability to discover new knowledge, to do very hard things, but also for people's ability to work with the models and get substantial pieces of work done at a rate that they never could before. [00:51:24] So, adoption of GPT-6 has been our strongest adoption by far of any new model. [00:51:29] And watching what people are doing with this across, you know, coding, kind of any version of enterprise software you'd like, security, science, like industry after industry has really been quite amazing to see. [00:51:42] Yeah, I, you know, I can tell you from, obviously, within Oracle, we're a big open AI user. [00:51:49] We made the decision to roll ChatGPT Desktop out to everybody. [00:51:53] But what's been surprising to me as well is if you look at, say, Codex, right, and the integration of who now is a developer, like, you know, the vast majority of code, more than 70% of our software these days is written by AI, and that's only accelerating. [00:52:10] Yeah. [00:52:11] But in addition to that, like, the number of people I go and I talk to who I don't think they could, you know, could not imagine themselves being a developer just a few years ago, are out creating entirely new applications. [00:52:23] And the joy that they have about being able to actually solve their problems and create has been incredible. [00:52:29] I mean, it's a big part of the reason why we focused on building App Composer was we had so many people building these new applications, and we ended up secure, you know, controlled way to actually deploy them. [00:52:39] How are you seeing AI change, like how people think about actually solving their problems, and who can do that for themselves? [00:52:46] You know, you mentioned that Oracle has rolled this out to everybody at Oracle, and that's been great to see you all have been some of the most aggressive adopters of this. [00:52:58] As we've had this big push into enterprise, there have been a bunch of surprises, but one of them has been watching that when a company does do this all the way across the board rollout, just how much people that did not think of themselves as creating software, in the broad sense of like getting a computer to do things that they want to do, how much that sort of spreads quickly throughout a company. [00:53:21] And so you see people, they may not describe it as programming, in fact, I kind of think they wouldn't, but the fact that now when they need something, they make an app to just do it, and the way that they interact with their own computer, the way that they create artifacts for other people to interact with for whatever their job function is, that has really changed quite a lot. [00:53:46] And, you know, like at this point it's become very trite, but this comment of people saying that I feel twice as productive at the end of 2026 as I did at the beginning of 2025, and that I enjoy my job much more because I'm doing kind of less of the, you know, less of the schlep. [00:54:06] Yeah, I think, you know, part of what's happening, right, is that, you know, I used to, I used to, I think of myself as having once been a software developer, although it feels like it's been too long now. [00:54:17] And what's interesting is that there's this big separation, it feels like, between that creative thing that got me hooked on it, right? [00:54:24] I remember being a kid and going, I'm going to learn how to like, I, the first time I started programming, it was like, I can have an idea and I can make it real. [00:54:30] But that creative aspect is very separate from how do I actually instruct this computer to do what I want. [00:54:37] And I feel like what's happening right now is that by having AI automate the kind of manual aspects of this activity, it unlocks that creativity to people that, for whatever reason, weren't necessarily inclined that way naturally, right? [00:54:54] One of the most, I strongly agree with that, it unlocks it in people that maybe weren't as inclined, and for people who are more inclined, it just gives you like, much more ability. [00:55:06] My experience, my subjective experience now, sitting in front of the computer using the latest tools, is some, like, it's like I am managing a factory inside of my computer. [00:55:16] And I can get things that I don't enjoy doing to be automated. [00:55:22] I don't feel like I'm stumbling or waiting or can't figure something out. [00:55:27] I feel like I can just sort of like move from one idea to another pretty continuously and without getting blocked. [00:55:33] And even if there's something that I've got like a vague sense for, but I'm not quite sure how to do it, and normally I would like give up on it and not try the activation energy would just be too high. [00:55:44] I am willing to explore these unlikely but sometimes big payoff creative directions. [00:55:49] And it feels like, I mean, it kind of feels like the cool thing we were all promised in sci-fi, where we're effectively just talking to a computer about our ideas and it's all happening and we get to be the creative force behind it. [00:56:01] Oh, I mean, it's the same experience. [00:56:04] I remember the joy I had when I discovered programming and I've been able to help a few younger people with that. [00:56:09] I see that joy now in people who are 30 years old and who are building complete websites from scratch to achieve their goal who just didn't think it was possible before. [00:56:17] We've got to go on, go ahead. [00:56:20] Well, look, we are on a time limit, Sam. [00:56:22] Okay, I want to talk about security. [00:56:25] It's changing very rapidly. [00:56:28] The models that you guys are creating clearly can both discover vulnerabilities and help us fix them. [00:56:34] That's a big part of why we've worked together and kind of created AI security mesh because it was like we saw a need from customers where they had access to the models. [00:56:42] But okay, how do I use it? [00:56:43] How do I point it at things? [00:56:44] How do I then validate the issues? [00:56:47] How do I then go out and deploy them? [00:56:50] We all have a lot more to do in the security space. [00:56:53] How should companies be thinking about adopting AI to protect themselves and their data? [00:56:59] So, if I could recommend like only one thing to a company right now to do with AI in the very short term, even more than like become super productive with Codex, it would be to really understand the security threat landscape from AI and mitigate against that. [00:57:17] And I realized that this is probably somewhat annoying for me to say. [00:57:21] And it's kind of like me like they're like, well, you're creating the problem. [00:57:24] Now you're telling us we have to like spend or work on the solution. [00:57:27] And that's where I think Rob did a good job. [00:57:29] Sorry to interrupt. [00:57:30] I don't AI is not creating the problem. [00:57:32] AI is not creating these bugs. [00:57:34] AI is helping us discover all of the security issues that we've have developed over the past 50 years. [00:57:39] I'm glad that we're able to now discover them with AI rather than waiting for some unknown hacker to let us know about them. [00:57:46] Super strongly agree. [00:57:48] I think the only way through this is to go find all these issues and fix them. [00:57:52] In fact, I think we'll need new ways to fix them. [00:57:54] It's maybe too hard to patch every bug one by one across the entire supply chain that we're all dependent on. [00:58:01] And I think we'll have to like look at new kinds of services that are constantly monitoring traffic and not letting these things be exploited. [00:58:08] But I find myself shocked at just how many vulnerabilities these AI systems are finding and what it's going to take. [00:58:18] Just the incredible, the Herculean effort that it's going to take to get ahead of us. [00:58:23] Yeah. [00:58:24] Well, I think I agree with you. [00:58:25] I think that's actually where AI is going to help us as well. [00:58:27] I think that we'll be able to, we can't do it manually. [00:58:29] We're going to automate it and we're going to use new technology to help us solve that. [00:58:33] Okay. [00:58:34] Last question. [00:58:35] So we could not have had this conversation a year ago. [00:58:38] Right? [00:58:39] The technology didn't exist. [00:58:40] The experience didn't exist. [00:58:41] So everything's moving faster as we discussed. [00:58:44] When you think about what comes next across both software and hardware, like what are you most excited about? [00:58:49] And what do you think everybody should be considering? [00:58:50] Many things here, but one of them that I can't resist talking about is how much AI is now helping us discover new science, invent new technology, actually like do this work of innovate and create better things. [00:59:09] We have seen tremendous examples with math problems that I mentioned earlier. [00:59:14] We're seeing that in lots of science. [00:59:15] We're seeing that in some of the most valuable technological problems we have. [00:59:20] AI is really good at chip design. [00:59:22] In fact, we made a chip called jalapeno that we're deploying in OCI and we're thrilled to be doing. [00:59:28] And we did this chip, which I think is the best AI chip in the world for inference, faster than would have ever been imaginable without AI to do a lot of this work. [00:59:39] So we have, we have seen in a small number of years, AI move from a chat bot to coding agents, to the ability to, you know, directly convert AI into innovation, like AI compute into innovation. [00:59:54] And whether that's make a chip or cure disease, it means we're all going to get much better stuff, higher quality services, lower cost services. [01:00:01] And I think what this is going to enable, we're still like, even though we now intellectually accept that it's happening, we have not emotionally felt that yet. [01:00:11] And the models all of a sudden are here. [01:00:13] No, I completely agree. [01:00:15] I think that the rate of change continues to increase. [01:00:18] Well, Sam, look, I want to say thank you for coming. [01:00:22] Thank you for the great partnership with you and everyone at OpenAI. [01:00:25] It's been great and I'm really excited for what comes next. [01:00:26] Thank you very much. [01:00:27] Thank you for having me. [01:00:31] What I'm realizing is that everything else I said was much less interesting than watching Sam walk away. [01:00:41] So I'm going to remember that. [01:00:45] I know we're, we have all your email addresses. [01:00:48] Okay. [01:00:50] So we started this conversation with a thesis that AI changes everything. [01:00:55] We went through an analysis of a bunch of evidence. [01:00:58] We talked about some small things. [01:01:00] We talked about some large things. [01:01:01] We talked about multiple verticals. [01:01:03] All of them showing, from my perspective, how categorically true it is that AI is changing everything. [01:01:10] Hopefully it is now clear to you just how impactful and how enabling AI is going to be across all of our industries. [01:01:18] You know, as we just talked about with Sam, AI is accelerating everything we can do. [01:01:23] And I personally cannot wait to see what happens next. [01:01:26] So with that, I want to say thank you all for being here and have a great rest of the AI world. [01:01:30] Thank you.