WEBVTT 00:30.338 --> 00:31.381 This is going to be 00:31.381 --> 00:34.967 the most promising technology in human history. 00:35.134 --> 00:40.473 AI is capturing the essence of cognition. In 2024 and 2025, 00:40.598 --> 00:44.477 coding became possible, and now we're starting to find that agents 00:44.477 --> 00:48.940 that do things end to end in other areas of knowledge work are possible. 00:49.065 --> 00:53.486 And even more recently, we're starting to see that fundamental discoveries 00:53.486 --> 00:58.491 in areas like biology and mathematics and physics are becoming possible. 00:58.699 --> 01:03.204 If humans can use AI to cure cancer, an evil 01:03.204 --> 01:07.083 human can use AI to make a bioweapon or do a cyber attack, 01:07.542 --> 01:11.003 we cannot succeed if we don't address these risks. 01:11.087 --> 01:15.925 And we are determined to get this right, because this is why I and this is why 01:15.967 --> 01:19.512 many people who work at Anthropic, this is why we work on this technology. 01:26.936 --> 01:29.856 Intelligence is now something humans can build. 01:29.856 --> 01:32.817 The implications of this are profound. 01:32.817 --> 01:36.654 We believe it will make every person far more capable and expand the world's 01:36.654 --> 01:41.325 capacity for cognitive work, transforming every industry and every discipline. 01:41.325 --> 01:45.246 This enormous economic impact begins with frontier intelligence, 01:45.496 --> 01:48.916 with what Claude can do. In 2024, 01:49.000 --> 01:53.504 Claude could answer a question, retrieve information, or summarize a document. 01:53.546 --> 01:55.423 By the end of 2025, 01:55.423 --> 01:58.176 Claude could do deep research and multi-step projects 01:58.176 --> 02:00.595 like refactoring an entire code base. 02:00.595 --> 02:06.100 The type of task that required hours or even days of human work. Today, 02:06.309 --> 02:09.687 Claude can take on complex, open ended projects 02:09.687 --> 02:13.483 that once took months or years of research, analysis 02:13.483 --> 02:17.570 and execution by many people and complete them in days. 02:17.904 --> 02:21.949 In the near future, we see potential for frontier AI to drive 02:21.949 --> 02:26.829 some of humanity's greatest breakthroughs and open entirely new markets. 02:27.163 --> 02:31.209 Our success thus far emerges from five core advantages. 02:31.542 --> 02:34.962 These are our mission and focus, talent and culture, 02:35.671 --> 02:38.049 frontier research and safety work, 02:38.049 --> 02:41.010 our approach to compute and our enterprise strategy. 02:41.010 --> 02:43.346 All of this begins with our mission. 02:43.346 --> 02:46.432 The responsible development and maintenance of advanced 02:46.432 --> 02:49.435 AI for the long term benefit of humanity. 02:49.685 --> 02:53.689 It's an organizing force and key to our focused execution. 02:53.773 --> 02:58.903 We manage a universe of finite resources: compute, talent, and research capacity. 02:59.153 --> 03:03.115 The mission is how we create clarity for the team and deploy those resources 03:03.115 --> 03:07.745 to maximum effect. At challenging moments that call for hard choices, 03:08.079 --> 03:10.414 our mission sharpens our decision making. 03:10.414 --> 03:14.377 It led us to deploy models like Mythos exclusively to cyber defenders. 03:14.544 --> 03:18.256 It also led us to commit to keep Claude free of advertising. 03:18.589 --> 03:23.010 This mission and focus continue to attract the very best talent in the industry. 03:23.302 --> 03:25.221 That's been true since our founding. 03:25.221 --> 03:28.266 Our seven co-founders include the originators of ideas 03:28.266 --> 03:31.477 that now make up the foundation of the AI industry. 03:31.727 --> 03:36.607 This concentration of top talent extends to all levels of the organization, 03:36.816 --> 03:40.820 and we lead the industry in attracting and retaining the brightest 03:40.820 --> 03:41.821 minds in the field. 03:48.869 --> 03:51.080 Why do I work at Anthropic? 03:51.080 --> 03:53.874 Why do I? Why do I work at Anthropic? Why do I work at Anthropic? 03:53.874 --> 03:56.752 Why do I work at Anthropic? 03:56.752 --> 03:59.463 Because I'm an optimist. 03:59.463 --> 04:01.799 I want a safer, better world for my family 04:01.799 --> 04:04.802 and for humanity. 04:05.678 --> 04:06.387 Three years 04:06.387 --> 04:09.432 ago, my older brother died of a brain tumor 04:09.432 --> 04:12.810 at the age of 42, which is the age that I am now. 04:13.227 --> 04:17.815 And I believe deeply that through our mission and this technology, 04:17.815 --> 04:21.152 there will be a situation where people like my brother 04:21.193 --> 04:24.155 don't have to die. 04:24.155 --> 04:26.574 Somebody has to carry the weight of getting this right. 04:26.574 --> 04:29.577 And I really didn't want to watch from the sidelines. 04:30.911 --> 04:34.373 People here ask, is this the right thing to do? 04:34.624 --> 04:36.876 Before they ask, can we do it? 04:38.711 --> 04:39.503 It's a rare 04:39.503 --> 04:44.091 combination of a group of people who care deeply about making AI go well, 04:44.091 --> 04:47.511 and also care deeply about competence, 04:47.678 --> 04:50.640 coherence, integrity, qualities 04:50.640 --> 04:53.934 that actually make an organization effective. 04:54.518 --> 04:58.564 These benefits and these risks, this light and this shade 04:58.731 --> 05:01.734 was precisely why I chose to come work at Anthropic. 05:06.614 --> 05:08.658 The work we do to make our models safe 05:08.658 --> 05:12.161 and reliable is the same work that makes them more capable. 05:12.328 --> 05:16.707 We care about useful intelligence, how effectively a model handles 05:16.707 --> 05:20.586 a real world task, and the value this delivers to customers. 05:20.795 --> 05:24.590 We're relentlessly focused on answering the question, what can Claude 05:24.632 --> 05:27.635 do for me that other models can't? 05:27.718 --> 05:31.055 Compute is the critical infrastructure that enables this work. 05:31.430 --> 05:34.934 We've built a differentiated advantage here, not just in how we build 05:34.934 --> 05:36.143 our compute capacity, 05:36.143 --> 05:38.354 but in how we make the most of this capacity 05:38.354 --> 05:41.023 through the discovery of compute multipliers. 05:41.023 --> 05:42.024 We are the leading 05:42.024 --> 05:45.653 AI lab that has been focused on building for the enterprise from day one. 05:45.778 --> 05:48.739 We train our frontier models to address knowledge work. 05:48.739 --> 05:52.952 As a result, our customers get the best of both capability and value. 05:53.577 --> 05:56.205 These customers have enabled our historic growth 05:56.205 --> 06:00.418 and they span industries across financial services, technology, 06:00.668 --> 06:04.839 pharma, healthcare, government, manufacturing, and more. 06:05.297 --> 06:08.801 And this is also where our public benefit mission and safety focus 06:08.801 --> 06:12.012 find natural alignment. Enterprises care that the frontier 06:12.012 --> 06:15.433 is safe, trustworthy and useful as much as we do. 06:15.599 --> 06:18.144 As frontier intelligence grows more capable, 06:18.144 --> 06:22.314 our dedicated focus on safety research becomes that much more important. 06:22.565 --> 06:26.402 It's critically important to our customers and necessary for realizing 06:26.444 --> 06:28.279 AI's full potential. 06:28.279 --> 06:31.449 And Anthropic is fully committed to doing this work with excellence. 06:31.615 --> 06:34.493 Our public benefit mission is what brought us to this point, 06:34.493 --> 06:37.788 and it will continue to shape our work, including at the governance level. 06:37.997 --> 06:41.959 In addition to our executive leadership and our board of directors, Anthropic 06:41.959 --> 06:45.212 is advised by the Long-Term Benefit Trust, an independent, 06:45.212 --> 06:48.424 financially disinterested body that selects a majority of our board. 06:48.799 --> 06:52.386 The LTBT's role is to help Anthropic plan for and fulfill 06:52.428 --> 06:54.430 our public benefit mission as we grow. 06:54.430 --> 06:58.642 We believe AI's future should be shaped and led by democratic societies. 06:58.851 --> 07:01.395 We want to be good partners in helping governments 07:01.395 --> 07:04.648 adapt to the changes that come with AI's rising capability. 07:04.899 --> 07:08.194 We do this at the executive level, through the work of our policy team 07:08.235 --> 07:11.238 and through efforts like the Anthropic Institute, which shares 07:11.238 --> 07:14.450 research on how frontier AI shapes how we live our lives. 07:15.034 --> 07:18.913 We are still early in what we believe will be the biggest transformation of the 07:18.913 --> 07:23.292 human experience in the modern age, not just the profound economic potential, 07:23.584 --> 07:27.713 but in all the ways in which we believe AI can help humanity flourish. 07:28.255 --> 07:30.800 From curing diseases to accelerating growth, 07:30.800 --> 07:33.761 to helping people live more fulfilled and abundant lives. 07:36.138 --> 07:38.599 So if you think about our Claude model family, 07:38.599 --> 07:42.311 it's designed to deliver value across all the goals you might have. 07:42.478 --> 07:45.606 Whether you're optimizing for cost or speed 07:45.606 --> 07:48.609 or intelligence. At the top end, 07:48.734 --> 07:51.737 Mythos and Fable push the frontier forward. 07:51.779 --> 07:55.324 We see customers use it to harden cyber infrastructure 07:55.366 --> 07:58.786 or make new scientific discoveries. At the other end, 07:58.911 --> 08:02.957 we have models that are extremely capable and highly cost effective, 08:03.123 --> 08:07.670 and businesses can rely on them at scale for well-defined, routine tasks. 08:07.962 --> 08:10.840 A key thing we think about is how everyone 08:10.840 --> 08:13.801 can get the benefit of that whole spectrum of models. 08:13.843 --> 08:17.680 Because even as model capability increases exponentially, 08:17.888 --> 08:21.600 what most people do with the models doesn't change that fast. 08:22.184 --> 08:25.187 We think about three principles as we build. 08:25.479 --> 08:28.440 The first is meeting customers where they are. 08:29.108 --> 08:31.235 There's a few ways to use Claude. 08:31.235 --> 08:34.280 First is through our own applications, like Claude Code 08:34.280 --> 08:36.323 or the Claude desktop or mobile app. 08:36.323 --> 08:40.244 But Claude is also available on enterprise applications like Word and Excel. 08:40.661 --> 08:43.664 We go deep into verticals and try to understand 08:43.873 --> 08:48.752 what does an expert in this domain need to be successful immediately? 08:49.211 --> 08:52.882 For instance, if you're a lawyer, we have a Claude for Legal bundle, 08:52.882 --> 08:55.885 which connects directly to the tools you need. 08:56.051 --> 08:59.889 This is something we think about for every domain. 09:00.264 --> 09:02.933 We have bundles for healthcare, financial 09:02.933 --> 09:05.936 services, life sciences, and more. 09:06.020 --> 09:10.691 Our second principle is that we build models and products side by side. 09:10.816 --> 09:14.778 We develop products for tomorrow's models, not just today's. 09:15.195 --> 09:18.908 Claude Code is a great illustration of why that matters. 09:19.325 --> 09:22.828 A couple of years ago, we could see that our prerelease 09:22.828 --> 09:25.831 models were getting better at understanding code. 09:25.873 --> 09:28.959 Couldn't quite do it all yet, but we could see the trend line. 09:29.335 --> 09:32.212 We built a product around that thinking 09:32.212 --> 09:35.090 and Claude Code was launched just a few months later. 09:35.090 --> 09:39.053 That introduced people to an entirely new way of coding, 09:39.386 --> 09:42.056 and of understanding what AI can do. 09:42.056 --> 09:47.186 This flywheel between model and product is a durable advantage in how we build. 09:47.394 --> 09:48.854 The model gets better. 09:48.854 --> 09:51.482 We develop new features powered by those advancements. 09:51.482 --> 09:53.651 Then we train the models to get even better 09:53.651 --> 09:57.279 at what people are trying to do, which reinforces the whole system. 09:57.655 --> 10:00.199 And we do that in coding and beyond. 10:00.199 --> 10:01.951 Our third principle is that we 10:01.951 --> 10:05.371 close the gap between the frontier and customers day to day. 10:05.871 --> 10:09.959 Most people still think about their AI tools as a replacement for search 10:09.959 --> 10:13.170 or a simple Q&A tool, and it can be those things. 10:13.504 --> 10:17.049 But that understates the power of what people can really do 10:17.091 --> 10:18.759 with this intelligence. 10:18.759 --> 10:23.639 This is where product becomes a meaningful competitive differentiator. 10:23.681 --> 10:27.768 Without a strong product, the difference between growing intelligence 10:27.768 --> 10:30.813 and where users are just keeps compounding. 10:31.063 --> 10:35.192 We think the future will involve making Claude feel as simple 10:35.192 --> 10:40.030 and natural as possible, so people everywhere can just treat Claude 10:40.072 --> 10:44.618 the same way you treat a colleague and get high quality output every time. 10:44.910 --> 10:50.374 We think these three principles are how we bring frontier intelligence to everyone. 10:50.708 --> 10:54.712 This is what we mean when we say we want to give people superpowers. 10:54.920 --> 10:59.842 Claude starts by helping you do what you need to better and faster. 11:00.009 --> 11:01.218 Then it takes on 11:01.218 --> 11:05.055 the things you never would have tried because they were just too hard. 11:05.264 --> 11:09.852 And in the future, that means all of us doing what used to be impossible 11:10.227 --> 11:13.439 so we can all live better, fuller lives. 11:16.942 --> 11:19.236 When we discovered scaling laws in AI, 11:19.236 --> 11:22.906 we realized that AI capabilities could improve very fast. 11:23.157 --> 11:26.994 That's why we founded Anthropic to ensure that if AI became 11:26.994 --> 11:30.789 a transformational technology, it would be safe and beneficial. 11:30.873 --> 11:35.252 Our fundamental bet is that AI would make tremendous progress this decade. 11:35.377 --> 11:38.464 These transformational capabilities are why research 11:38.464 --> 11:41.467 and safety must advance together. 11:41.467 --> 11:44.803 It's the only way we'll realize the benefits of the technology. 11:45.012 --> 11:49.850 We assess our models across three dimensions: capability, efficiency, 11:50.100 --> 11:51.477 and safety. 11:51.477 --> 11:55.189 We optimize our models to be able to plan and complete complex work 11:55.230 --> 11:58.525 accurately, across hours and even days. 11:59.068 --> 12:01.737 A good example is Opus 5.5. 12:01.737 --> 12:05.949 One of our users was able to complete a 680K-line code migration 12:05.949 --> 12:06.950 in less than a day. 12:06.950 --> 12:10.287 This would have taken an engineering team weeks just a few years ago. 12:10.454 --> 12:14.708 We think our models are the very best at this long horizon work where accuracy 12:14.708 --> 12:15.959 really matters. 12:15.959 --> 12:18.962 The capability begins in the model training process. 12:19.129 --> 12:23.342 In the pre-training phase, a model learns from the totality of human knowledge. 12:23.425 --> 12:24.051 Text. 12:24.051 --> 12:27.054 Images, scientific papers, code. 12:27.304 --> 12:29.807 This phase is where it comes to understand 12:29.807 --> 12:32.810 fundamental human concepts and ideas. 12:33.185 --> 12:35.687 The second phase is reinforcement learning. 12:35.687 --> 12:38.690 This is where Claude learns to be helpful, honest, and harmless. 12:38.857 --> 12:41.819 It's where we draw out useful capabilities 12:41.819 --> 12:44.738 in coding, knowledge work, and science. 12:44.738 --> 12:48.200 As models have moved from conversational assistance to autonomous agents, 12:48.242 --> 12:49.827 reinforcement learning has evolved 12:49.827 --> 12:52.913 into one of the largest areas of research investment. 12:53.080 --> 12:55.457 One that really matters to our customers. 12:55.457 --> 12:58.127 In 2021, I wrote an internal memo 12:58.127 --> 13:03.215 outlining how our most likely product might be an aligned AI programmer. 13:03.340 --> 13:07.010 I wrote a lot of other things too, but this one turned out to be pretty accurate. 13:07.094 --> 13:10.347 Claude's training process is integral to Anthropic's mission. 13:10.514 --> 13:13.433 We need to understand how to train frontier intelligence 13:13.433 --> 13:14.768 and how to make it safe. 13:14.768 --> 13:17.271 And this is how we deliver the benefits of AI. 13:24.361 --> 13:25.487 We think that AI 13:25.487 --> 13:29.616 will help us reduce the time to do a clinical trial by 20 to 30%. 13:29.658 --> 13:33.787 If you get a drug to market 2 or 3 years sooner, those are either cures 13:33.871 --> 13:37.666 or treatments that patients that actually don't have solutions can have sooner. 13:37.749 --> 13:40.127 So we think it's actually a huge win for patients. 13:40.127 --> 13:43.130 And it's also a huge win for BMS. 13:43.380 --> 13:45.340 We chose Anthropic for a couple of reasons. 13:45.340 --> 13:48.177 One is they had the most mature enterprise capabilities. 13:48.177 --> 13:51.180 So it's one thing to have a chatbot that you can use. 13:51.180 --> 13:54.099 It's a whole different thing to actually put it in an environment 13:54.099 --> 13:56.143 where you're giving it access to some of the company's 13:56.143 --> 14:00.397 most confidential information, and you're able to do observability, 14:00.731 --> 14:04.902 manage the security, manage what the fine tune controls of what you give users. 14:04.902 --> 14:07.362 So there's like the enterprise controls that you'd often 14:07.362 --> 14:10.365 see that IT organizations really value a lot. 14:10.449 --> 14:13.452 We rolled it out across all 30,000 of our employees. 14:13.744 --> 14:14.953 This is the fastest, 14:14.953 --> 14:19.249 most well adopted AI tool in the history of what we rolled out. 14:19.333 --> 14:22.377 So now you can ask a very, very complicated question 14:22.544 --> 14:23.545 and it will go through. 14:23.545 --> 14:27.132 It will figure out which database or databases, because it can actually 14:27.132 --> 14:30.135 cross those databases, can be queried. 14:30.219 --> 14:31.553 And then it will produce a report 14:31.553 --> 14:34.890 that actually has the context of the epidemiology built right into it. 14:35.015 --> 14:39.353 So that takes you from what would have typically taken 1 to 2 months 14:39.353 --> 14:44.066 to get a simple question answered, to being able to get it in 4 to 5 minutes. 14:44.399 --> 14:48.654 And we're now finally at a stage both with high throughput computation, 14:48.654 --> 14:51.740 but also large language models where you're able to scale up 14:51.740 --> 14:55.202 your computation, your analytics to match the better scale of biology 14:55.202 --> 14:58.413 that can help the way that you inform and change the standard of care. 15:02.000 --> 15:02.751 Consider all the 15:02.751 --> 15:05.754 reasons a frontier AI lab needs to use compute efficiently. 15:05.796 --> 15:09.258 We use it at every stage of the research and model training process. 15:09.466 --> 15:13.470 It's how we serve models to customers and how we use Claude at scale internally. 15:13.679 --> 15:16.598 By training what we believe to be the most efficient models, 15:16.598 --> 15:19.476 we can do more of all these things with the same amount of compute. 15:19.476 --> 15:21.478 It's also how we drive down costs, 15:21.478 --> 15:24.731 both for customers as well as our own underlying costs. 15:24.982 --> 15:28.318 We're accelerating these efficiency gains through our research process. 15:28.694 --> 15:32.114 Our pace of discovery is speeding up, and a lot of that acceleration 15:32.114 --> 15:33.949 comes from Claude itself. 15:33.949 --> 15:36.743 When you deploy frontier models, safety is critical. 15:36.743 --> 15:38.161 This idea brought the co-founders 15:38.161 --> 15:41.373 together, and it's been at the foundation of Anthropic from the beginning. 15:41.623 --> 15:45.043 Excellence in safety is how you manage organizational and systemic risk. 15:45.168 --> 15:47.296 It's how you safely enable frontier capabilities 15:47.296 --> 15:50.966 in fields like coding and medicine, and it's valued by our customers. 15:51.174 --> 15:54.052 It gives them the trust they need to point AI at important work. 15:54.052 --> 15:56.513 We build safety in at every stage. 15:56.513 --> 15:59.224 Through alignment training, we train models to follow user intent 15:59.224 --> 16:01.059 and avoid taking harmful actions. 16:01.059 --> 16:04.062 And we build in the right safeguards aimed at preventing misuse. 16:04.479 --> 16:06.648 We also pioneer fields like interpretability, 16:06.648 --> 16:09.651 which gives us insight into what's happening inside the model's brain. 16:09.776 --> 16:10.902 Within Anthropic, 16:10.902 --> 16:14.531 we increasingly turn frontier models back on improving our own processes. 16:14.656 --> 16:16.700 We call this closing the loop, 16:16.700 --> 16:19.494 and it drives forward all three of our research priorities. 16:19.494 --> 16:22.205 Claude is increasingly our equal partner here. 16:22.205 --> 16:25.208 The better Claude gets, the more it helps us with all of our work. 16:25.250 --> 16:26.668 It improves our security. 16:26.668 --> 16:28.712 It makes training and inference more efficient. 16:28.712 --> 16:30.255 It drives safety work. 16:30.255 --> 16:33.050 It helps us try out more research ideas more quickly, 16:33.050 --> 16:35.469 and it makes the whole organization more effective. 16:35.469 --> 16:38.138 Our most important work is still ahead 16:38.138 --> 16:40.724 if we keep research and safety advancing together. 16:40.724 --> 16:43.602 We believe that we can help bring about some of the most significant 16:43.602 --> 16:45.103 breakthroughs in human history. 16:51.610 --> 16:54.863 Take a number and double it and double that. 16:54.863 --> 16:56.782 Then double that. 16:56.782 --> 17:00.577 That curve is really at the heart of what we do at Anthropic. 17:01.870 --> 17:05.832 We noticed early on that performance seemed to scale very smoothly 17:05.832 --> 17:08.835 with compute. 17:10.420 --> 17:11.129 Our bet 17:11.129 --> 17:14.466 was there is a whole internet's worth of context. 17:15.008 --> 17:18.970 What if the limiting factor of model size just went away? 17:20.639 --> 17:23.975 So in 2022 or 23, we had chatbots. 17:25.435 --> 17:27.771 You could ask them for a history of the samurai, and 17:27.771 --> 17:30.774 it would come back with a little essay. 17:31.691 --> 17:33.026 In 2024, 17:33.026 --> 17:37.656 coding becomes possible, but Claude still struggles on hard problems. 17:38.782 --> 17:43.495 In 2025, the models start to be able to do surprising things. 17:43.912 --> 17:46.039 Coding gets much better. 17:46.039 --> 17:49.709 Claude can even play video games quite badly. 17:50.919 --> 17:52.838 This same year, it found a flaw 17:52.838 --> 17:56.716 in OpenBSD that had gone unnoticed for 27 years. 17:59.636 --> 18:02.639 Suddenly it can do multi-step processes. 18:03.223 --> 18:05.976 Knowledge work. 18:05.976 --> 18:10.230 It builds a controller for laser locks at a quantum computer company, 18:10.480 --> 18:14.734 a task that used to take a team of people with PhDs a week. 18:15.068 --> 18:17.821 And it does this without human intervention 18:17.821 --> 18:20.824 overnight. 18:20.866 --> 18:21.825 We start to see 18:21.825 --> 18:24.828 Claude making fundamental discoveries in math, 18:25.036 --> 18:28.039 working at the frontier of life science. 18:28.540 --> 18:31.543 This acceleration can be disorienting, 18:32.210 --> 18:35.755 but there's still this sense that we're just setting out. 18:35.839 --> 18:39.259 That exponential curve stretches up ahead of us, 18:39.759 --> 18:42.762 and it's only just begun. 18:46.600 --> 18:48.602 Achieving Anthropic's mission requires 18:48.602 --> 18:52.189 staying at the frontier and staying at the frontier depends on compute. 18:52.230 --> 18:54.691 Our approach here goes back to research that I did 18:54.691 --> 18:57.486 with several of my other co-founders on scaling laws. 18:57.486 --> 18:59.404 We found that with the right recipe, 18:59.404 --> 19:03.200 more compute reliably and predictably yields more intelligence. 19:03.241 --> 19:06.870 Later, we found that it was possible to improve this recipe over time 19:06.953 --> 19:10.540 to achieve more intelligence with the same amount of physical compute. 19:10.665 --> 19:14.419 So by the time we started Anthropic in 2021, we knew two things: 19:14.628 --> 19:18.507 one, compute is the foundation of the business, and two, getting the most 19:18.507 --> 19:22.552 out of every unit of compute is key to making increasingly intelligent models. 19:22.594 --> 19:26.097 That's why we're so focused on efficiency, and it's why we believe a dollar 19:26.097 --> 19:30.185 of compute at Anthropic goes further than a dollar at any other frontier lab. 19:30.268 --> 19:32.521 We've achieved that advantage in two ways. 19:32.521 --> 19:33.939 I'll start with the first. 19:33.939 --> 19:37.359 Our proprietary approach to bringing up and deploying silicon. 19:37.651 --> 19:41.613 We train and serve Claude now across all four major chip families. 19:41.821 --> 19:45.492 That's only possible because early on, we made an unconventional bet 19:45.534 --> 19:47.577 on a multi chip strategy. 19:47.577 --> 19:51.206 At the time, many people thought that running AI models on different chips 19:51.206 --> 19:52.749 wasn't worth the effort. 19:52.749 --> 19:56.044 Each chip needs its own tailored software stack and inference 19:56.044 --> 19:59.047 stack, which is expensive and difficult to build. 19:59.089 --> 20:02.551 We've now been doing it for years, and we've gotten incredibly good at it. 20:02.842 --> 20:06.846 This comes with two key advantages. In a supply-constrained compute market, 20:06.846 --> 20:08.473 our multi-silicon strategy 20:08.473 --> 20:12.310 allows us to focus on the best price performance regardless of the provider. 20:12.435 --> 20:15.438 We can match each chip to the job that it's best at. 20:15.689 --> 20:18.441 Some are better at training, some are better at inference. 20:18.441 --> 20:22.445 Our multi-silicon chip strategy means we can treat all of our chips across 20:22.445 --> 20:27.409 providers, architectures and locations as one fungible pool of compute. 20:27.784 --> 20:30.787 This has also made us really good at bringing up new chips. 20:30.787 --> 20:32.247 As an example, this year 20:32.247 --> 20:36.293 we started working with AMD GPUs for the first time. In the past, 20:36.334 --> 20:40.088 bringing on a new chip platform took dozens of engineers months of work. 20:40.505 --> 20:42.716 This time we had Mythos try it out. 20:42.716 --> 20:46.303 It got Claude running on AMD in just three days with basically 20:46.303 --> 20:47.846 no human intervention. 20:47.846 --> 20:50.974 Claude ported or wrote almost 100% of the kernels. 20:51.266 --> 20:54.603 We also have a distinct approach to the data centers that house our chips. 20:54.936 --> 20:57.772 Historically, we bought compute as a cloud service. 20:57.772 --> 21:00.442 We've now added a second approach where we lease chips 21:00.442 --> 21:03.445 and run them in purpose built data centers designed for us. 21:03.570 --> 21:05.905 We call this Anthropic Direct. 21:05.905 --> 21:08.199 That approach has a couple advantages for us. 21:08.199 --> 21:11.119 It helps to lower the total cost of ownership of compute, 21:11.119 --> 21:14.873 and allows us to separate the commitment to data centers from the chips. 21:15.290 --> 21:18.627 The second driver of our compute advantage is our rate of discovery 21:18.627 --> 21:20.086 of compute multipliers. 21:20.086 --> 21:24.090 We focus on total effective compute, which is the multiplicative product 21:24.090 --> 21:27.844 of physical capacity, times the compute multipliers that we discover. 21:28.094 --> 21:32.307 Each one we discover makes our pool of total effective compute bigger. 21:32.474 --> 21:33.308 For example, 21:33.308 --> 21:37.020 say you need 100 units of compute to train a model of a certain intelligence. 21:37.270 --> 21:40.815 A 4x compute multiplier means you can use 25 units of compute 21:40.815 --> 21:42.442 to train the exact same model. 21:42.442 --> 21:46.613 Or you can use those 100 physical units to train an even more intelligent model 21:46.696 --> 21:49.699 that would have previously required 400 units. 21:49.699 --> 21:53.536 Compute multipliers create a competitive advantage for us that's hard to replicate. 21:53.536 --> 21:54.788 They build on each other. 21:54.788 --> 21:55.997 Each new multiplier we 21:55.997 --> 21:59.918 discover has a positive effect on the multipliers before it. They stack. 21:59.959 --> 22:01.586 We believe that this can accelerate 22:01.586 --> 22:04.589 even further as we approach recursive self-improvement. 22:04.631 --> 22:08.176 Claude will be able to discover compute multipliers entirely by itself, 22:08.385 --> 22:11.888 allowing the number of compute multipliers to grow faster and faster. 22:12.180 --> 22:14.557 We believe this combination of smart investments, 22:14.557 --> 22:17.560 breakthrough research, and the improving capabilities of Claude 22:17.602 --> 22:20.980 have put Anthropic in a leading position among the frontier labs. 22:26.319 --> 22:27.237 AI is 22:27.237 --> 22:31.199 a breakthrough innovation, but only reinvention 22:31.199 --> 22:34.869 is going to get the value that AI promises. 22:35.078 --> 22:38.665 At Accenture, our job is to be the bridge between technology 22:38.665 --> 22:41.668 and real outcomes for our clients. 22:42.127 --> 22:45.463 I love the question about, you know, are we using Claude 22:45.463 --> 22:49.759 because we have tens of thousands of engineers using it every day? 22:49.884 --> 22:53.221 We use it across our entire software development lifecycle, 22:53.304 --> 22:54.723 and it frees up time. 22:54.723 --> 22:56.558 It makes us go faster and 22:56.558 --> 23:00.770 allows our engineers to spend more time focusing on solving our clients' 23:00.770 --> 23:04.566 challenges, leading decision making, and setting strategy. 23:05.316 --> 23:10.196 We defend thousands of our clients from cyber attacks, 23:10.196 --> 23:15.076 and we've paired Claude Code and Claude's agent software development kit with 23:15.285 --> 23:19.998 Cyber.AI, our cutting-edge platform that we use to help protect our clients. 23:20.039 --> 23:21.958 Just one result of using this 23:21.958 --> 23:25.670 is that we can now do scans that used to take days in under an hour. 23:25.837 --> 23:30.842 We are focused on our clients' success and safety, and our mindset 23:30.884 --> 23:35.805 is that because of our partnership, our clients can get to value faster. 23:40.685 --> 23:43.062 The early bet we made was that intelligence 23:43.062 --> 23:46.107 was the highest value asset that a business could get from AI. 23:46.524 --> 23:49.903 We believe frontier intelligence solves the toughest problems, 23:49.944 --> 23:53.615 and for that reason it drives the biggest impact. 23:53.907 --> 23:56.576 AI capability doesn't improve gradually. 23:56.576 --> 24:00.622 When models cross a certain threshold, there is a step change 24:00.747 --> 24:04.083 and an entire category of work is done differently. 24:04.834 --> 24:07.712 Coding is the example we're most familiar with. 24:07.712 --> 24:10.006 The speed of the change is remarkable. 24:10.006 --> 24:13.259 We went from autocomplete a few lines of code 24:13.343 --> 24:17.222 to fleets of agents working around the clock in just a year. 24:17.347 --> 24:20.350 For a company like Rakuten, this means they're shipping 24:20.350 --> 24:24.229 new features 4 to 5 times faster than their competitors. 24:24.687 --> 24:30.193 After coding came cyber, and with Mythos and Project Glasswing, Claude has helped 24:30.193 --> 24:34.739 cyber defenders identify and patch vulnerabilities at a societal scale. 24:35.198 --> 24:37.408 Consider a company like Mozilla. 24:37.408 --> 24:41.120 In the past, they used to fix over 20 security bugs per release, 24:41.371 --> 24:46.167 and with Mythos Preview, they identified and fixed 271. 24:46.584 --> 24:48.962 Knowledge work is on a similar curve. 24:48.962 --> 24:52.173 Scientific research, investment analysis, underwriting 24:52.173 --> 24:55.385 all great examples of what we're seeing across industries. 24:55.718 --> 25:00.807 Claude helped an insurer like AIG cut down time to quote by 55%. 25:00.932 --> 25:03.017 In each domain, we see an improvement. 25:03.017 --> 25:07.856 Then a step change, and then a whole category of work moves fast. 25:13.152 --> 25:15.280 It's hard to imagine 25:15.280 --> 25:18.241 the speed with which that Anthropic 25:18.241 --> 25:21.244 has become embedded inside of AIG. 25:21.536 --> 25:23.621 It's helped us create better underwriters. 25:23.621 --> 25:27.917 We were able to use their skill set more frequently because we're able 25:27.917 --> 25:30.920 to pull the submission information, get it in a format where they can 25:30.962 --> 25:31.921 then underwrite. 25:31.921 --> 25:36.092 We've seen a significant uptick in the speed with which they can make decisions, 25:36.175 --> 25:37.385 which really means 25:37.385 --> 25:41.014 we're able to get to more submissions in a more consistent way. 25:41.180 --> 25:44.183 And for clients, we're able to respond faster. 25:45.393 --> 25:47.478 I think Anthropic and Claude together 25:47.478 --> 25:51.524 have a very similar value proposition that we try and bring in terms of safety 25:51.524 --> 25:55.403 and resilience and technical capability, and we communicate pretty closely 25:55.403 --> 26:00.158 with the regulators on how we're using AI and how the human element 26:00.158 --> 26:02.869 and the human oversight is obviously very important to us 26:02.869 --> 26:05.580 and to our regulators and is embedded in the process. 26:08.625 --> 26:10.418 Everything Anthropic does, 26:10.418 --> 26:13.463 from research to product to go-to-market, 26:13.504 --> 26:16.799 it all comes together in how we serve the customer. 26:17.091 --> 26:22.764 Today, we have deep engagements across the Fortune 500 and all of the Fortune 10. 26:23.097 --> 26:27.727 More than 50% of the Fortune 500 spends more than $100K with us. 26:28.019 --> 26:31.648 And that's true for more than 30% of the Global 2000. 26:32.190 --> 26:35.109 With all of these customers, we believe we are still 26:35.109 --> 26:38.279 in the early stages of a multi-decade transformation. 26:38.655 --> 26:41.866 On average, spending from these Fortune 500 customers 26:41.866 --> 26:44.911 is less than 0.1% of their total expenses. 26:45.203 --> 26:49.958 Even among the top ten, it's still less than 1% of their total expense line. 26:50.333 --> 26:51.000 Let's zoom 26:51.000 --> 26:55.546 in on the journey of a typical customer, and it starts with a single use case. 26:55.838 --> 26:57.256 Often that's Claude Code 26:57.256 --> 27:01.094 for developers and engineers, and it soon expands to adjacent roles. 27:01.177 --> 27:04.639 Knowledge workers start using Claude for their highest value work. 27:04.639 --> 27:07.600 Then teams automate more business processes 27:07.600 --> 27:10.353 and build AI into core infrastructure. 27:10.353 --> 27:15.066 And ultimately they embed Claude into entirely new products and services. 27:15.233 --> 27:18.903 The process happens quickly, and that's why we see an unprecedented 27:18.903 --> 27:22.782 net revenue retention rate of approximately 900%. 27:23.074 --> 27:25.827 What excites me is that the most interesting 27:25.827 --> 27:28.830 and transformative use cases are still ahead of us. 27:28.997 --> 27:33.084 Everything we've seen in coding and cyber, we're now starting to see across 27:33.084 --> 27:37.672 virtually every function, every company and every industry. 27:37.714 --> 27:41.676 And they want to partner with Anthropic because they trust we will help them 27:41.718 --> 27:45.221 see the future first, that we will keep them on the frontier, 27:45.304 --> 27:47.598 and that we will help them transform and win. 27:53.938 --> 27:54.647 Something 27:54.647 --> 27:59.986 exciting happens when the discoveries unlocked by Claude move into our world, 28:00.403 --> 28:03.364 or even beyond it. 28:03.364 --> 28:05.074 In December 2025, 28:05.074 --> 28:08.286 Claude helped plan the Perseverance rover's route on Mars. 28:09.412 --> 28:12.457 It wrote the code in the rover's own markup language. 28:13.207 --> 28:16.502 Engineers checked everything before the commands went out. 28:17.045 --> 28:19.422 400m driven. 28:19.422 --> 28:22.425 Mission accomplished. 28:22.884 --> 28:25.720 Claude's ability to process vast amounts of data 28:25.720 --> 28:28.931 came to the fore again in the summer of 2026. 28:29.515 --> 28:33.311 About 950 Claude agents spent 21 hours 28:33.436 --> 28:36.773 sifting 1.9 billion protein clusters. 28:36.981 --> 28:41.694 One spotted an unusual pattern: a system with a CRISPR-like repeat array. 28:42.153 --> 28:45.406 Our scientists were able to check Claude's findings in the lab. 28:45.656 --> 28:49.077 The new array had never been published before. A genuine 28:49.077 --> 28:52.080 discovery. 28:54.457 --> 28:55.208 The ability 28:55.208 --> 28:59.879 to quickly understand new information really matters where resources are thin. 29:02.048 --> 29:03.132 In September 29:03.132 --> 29:07.553 2026, teams responding to an Ebola outbreak in eastern Congo 29:07.678 --> 29:12.517 were able to process each region's daily reports in under an hour, 29:13.226 --> 29:15.770 creating a better picture of the epidemic 29:15.770 --> 29:18.773 and freeing up time for patient care. 29:20.066 --> 29:21.359 The power to affect 29:21.359 --> 29:24.320 the real world makes safety even more important. 29:24.904 --> 29:27.573 It's why Claude's full biology capabilities 29:27.573 --> 29:30.576 go only to organizations we've verified. 29:32.537 --> 29:35.540 And why Claude has a constitution. 29:36.040 --> 29:40.336 It equips it to make judgment calls in situations it has never encountered 29:40.336 --> 29:42.380 before. 29:42.380 --> 29:47.135 And it's why we put so much energy into interpretability and alignment. 29:47.343 --> 29:51.681 To keep Claude accountable in situations where people will rely on it 29:52.807 --> 29:53.808 again and again. 29:53.808 --> 29:56.435 We're finding safety doesn't hold us back. 29:56.435 --> 29:59.480 In fact, it's because we can trust Claude's work 29:59.564 --> 30:03.109 that we can take its discoveries and apply them in the real world 30:04.235 --> 30:07.238 in ways that will benefit everyone. 30:11.325 --> 30:14.370 We believe the transformation of the global economy through 30:14.412 --> 30:17.540 AI is the greatest economic opportunity in history. 30:17.665 --> 30:20.376 And Anthropic is uniquely positioned to capture it. 30:20.376 --> 30:23.462 We have leadership at the frontier of AI, where we expect value 30:23.504 --> 30:26.507 will differentially accrue, and the strength of our business model 30:26.549 --> 30:29.510 allows us to keep investing to advance the frontier 30:29.510 --> 30:33.639 while delivering more value to customers and superior returns for our shareholders. 30:33.639 --> 30:36.642 The clearest evidence of that is our financial results. 30:36.684 --> 30:39.896 In Q3, we earned nearly $17 billion 30:39.896 --> 30:43.065 of revenue, up 12 times year over year. And year to date, 30:43.107 --> 30:46.027 we've delivered a compute margin of 43%, 30:46.027 --> 30:48.070 gross margin of 85%, 30:48.070 --> 30:51.949 adjusted operating margin of 12% and positive free cash flow. 30:52.491 --> 30:55.328 These results come from our frontier leadership, 30:55.328 --> 30:58.456 and they give us the capacity to invest heavily to extend it. 30:58.456 --> 31:02.418 So while our profitability is strong, we are in an aggressive investment 31:02.418 --> 31:05.546 mode as we focus on building this business for the long term. 31:05.755 --> 31:09.133 There are five elements that differentiate Anthropic's business model. 31:09.175 --> 31:13.221 First, we build models that deliver the most useful intelligence 31:13.346 --> 31:14.680 as we've pushed the frontier. 31:14.680 --> 31:17.516 Our customers have found more valuable work for Claude. 31:17.516 --> 31:19.477 And our revenue growth has followed. 31:19.477 --> 31:21.479 We call this model led growth. 31:21.479 --> 31:23.981 The trajectory of our business demonstrates this. 31:23.981 --> 31:28.694 Quarterly revenue grew from $376 million in Q1 2025 31:29.070 --> 31:32.490 to nearly $17 billion in Q3 of 2026. 31:32.865 --> 31:36.410 This tracks significant increases in the capabilities of our models 31:36.410 --> 31:40.414 in coding workflows and real world tasks. 31:40.748 --> 31:44.543 Being at the frontier unlocks new use cases and capabilities. 31:44.585 --> 31:48.673 It also moves us towards Pareto dominance across most customer workloads, 31:48.798 --> 31:52.969 enabling us to deliver the most capable models at a given cost. 31:53.094 --> 31:56.055 That is because smarter models are more efficient. 31:56.055 --> 32:00.476 They can complete tasks at lower effort levels, which use fewer tokens. 32:00.643 --> 32:04.814 We also cascade the frontier intelligence from our larger models 32:04.897 --> 32:06.482 into our smaller models. 32:06.482 --> 32:10.695 So we address the full range of cost, capability and speed. 32:11.070 --> 32:13.781 Each generation has pushed that further. 32:13.781 --> 32:16.701 Our recent Claude 5.5 family of models are a 32:16.701 --> 32:20.162 significant step up in capability from the Claude 5 family 32:20.204 --> 32:22.123 we released just a few months prior. 32:22.123 --> 32:25.334 They're also significantly cheaper on a cost per task basis. 32:25.459 --> 32:28.337 That is the power of building at the frontier. 32:28.337 --> 32:32.008 We believe we deliver the highest value intelligence at the lowest cost, 32:32.133 --> 32:34.969 which means we can serve the entire market efficiently. 32:34.969 --> 32:38.931 While we lead across most benchmarks, the true measure of our success 32:39.432 --> 32:40.599 is our customers. 32:40.599 --> 32:43.602 And that leads me to our second differentiator. 32:44.020 --> 32:46.981 We build the leading platform and products to make 32:46.981 --> 32:49.984 frontier intelligence accessible to our customers. 32:50.151 --> 32:53.654 Our platform solutions are the foundation of our business, 32:53.988 --> 32:57.700 and there is significant opportunity for customers to build products 32:57.700 --> 33:01.412 and reimagine processes on top of our APIs and SDK. 33:01.829 --> 33:06.042 At the same time, our own products have become a key driver of diffusion. 33:06.208 --> 33:07.168 For example, 33:07.168 --> 33:11.714 Claude Code was fundamentally a bet on the exponential in model progress. 33:11.839 --> 33:15.760 Getting it into the hands of our customers put us in a position to capitalize 33:15.760 --> 33:19.221 on agentic coding capabilities as they dramatically improved. 33:19.430 --> 33:23.309 We followed that same playbook with Claude Cowork, and its adoption 33:23.309 --> 33:26.312 has exceeded that of Claude Code at the equivalent point, 33:26.312 --> 33:28.522 following its own research preview last year. 33:28.522 --> 33:33.319 Together, our products and platform enable us to serve the needs of our customers. 33:33.486 --> 33:37.656 Today, the majority of enterprise customers use both our platform 33:37.698 --> 33:42.536 and at least one of our products, up from 30% in the fourth quarter of 2025. 33:42.703 --> 33:47.375 Our product velocity is also accelerating as we expand our internal use of Claude. 33:47.458 --> 33:51.295 The average daily volume of code contributed per engineer is up 33:51.295 --> 33:54.423 45x relative to the 2024 average, 33:54.423 --> 33:58.469 and that's showing up in our launches. In the first nine months of 2026, 33:58.594 --> 34:02.056 we shipped 250 models, products and features. 34:02.348 --> 34:06.310 And Claude isn't just accelerating our engineering teams across the company. 34:06.435 --> 34:10.272 Our employees' use of Claude grew more than 250 times 34:10.314 --> 34:13.317 between January and September of 2026. 34:13.359 --> 34:17.238 Together, these help explain why our revenue per employee has grown 34:17.238 --> 34:22.076 more than 4.5x from $1.4 million in Q3 of 2025 34:22.201 --> 34:25.955 to over $6.6 million as of Q3 2026. 34:25.996 --> 34:30.668 Anthropic's third differentiator is that we are extremely efficient users of compute. 34:30.751 --> 34:34.755 We treat our compute as one fungible pool that serves customers, 34:34.839 --> 34:38.092 trains our models and powers our own use of Claude. 34:38.217 --> 34:41.387 We are focused on getting the most out of every unit of compute. 34:42.138 --> 34:45.558 This gives us a cost to serve and a reinvestment advantage. 34:45.850 --> 34:49.186 As Tom described, there are two drivers of this efficiency. 34:49.728 --> 34:53.107 The first is our multi-silicon strategy, which lets us 34:53.107 --> 34:57.111 buy the most price-performant chips and run them as a single fungible pool. 34:57.194 --> 35:01.824 The second is compounding compute multipliers, which have scaled our total 35:01.824 --> 35:05.744 effective compute significantly faster than our physical capacity. 35:06.287 --> 35:09.915 You can see this in our financial results over the last six quarters. 35:09.999 --> 35:14.086 We grew revenue by more than 45x, while computing infrastructure costs 35:14.086 --> 35:17.423 grew only 10x, a more than 4x improvement 35:17.423 --> 35:20.426 in revenue per dollar of compute. And year to date, 35:20.551 --> 35:23.679 our gross margin has improved to over 85%, 35:23.929 --> 35:27.266 up from 29% in the same period of 2025. 35:27.349 --> 35:31.270 We think of these efficiency gains as capital that we get to allocate. 35:31.479 --> 35:35.357 In some cases, we pass them along to customers by lowering prices, 35:35.399 --> 35:38.527 which drives diffusion of Claude across those customers. 35:38.736 --> 35:42.072 In other cases, we retain the margin and reinvest it 35:42.072 --> 35:45.784 in additional compute, which is what makes this advantage compound. 35:45.910 --> 35:50.456 The next differentiator is our relentless focus on building for enterprises. 35:50.748 --> 35:54.710 We had the early insight that enterprises would be the largest and best customers 35:54.710 --> 35:55.669 for AI. 35:55.669 --> 35:59.298 We have focused on serving that market since our commercial launch 35:59.298 --> 36:02.885 in the first quarter of 2023, and have seen rapid adoption 36:02.885 --> 36:05.888 since. As of Q3 2026, 36:05.930 --> 36:09.725 we had more than 11,900 enterprise customers, up 36:09.725 --> 36:11.810 nearly 8x from the start of the year. 36:11.810 --> 36:15.105 As the first frontier lab available on all three major clouds, 36:15.397 --> 36:18.359 we allow enterprises to deploy Claude within the environments 36:18.359 --> 36:19.610 they already operate in, 36:19.610 --> 36:22.530 and we maintain the direct relationship with the end customer, 36:22.530 --> 36:24.406 including control over pricing. 36:24.406 --> 36:27.535 We believe these partnerships enable us to be far more efficient 36:27.535 --> 36:30.162 in our go-to-market efforts than the rest of the industry. 36:30.162 --> 36:33.874 In fact, our entire non-compute expense base, including sales 36:33.874 --> 36:37.628 and marketing personnel and other operating expenses, represents 36:37.628 --> 36:40.589 a lower percentage of revenue than the median enterprise 36:40.589 --> 36:43.592 software company spends on sales and marketing alone. 36:43.592 --> 36:47.054 We expect to invest meaningfully in go-to-market to support our growth, 36:47.388 --> 36:51.267 while remaining highly efficient in how we sell to and grow with our customers. 36:51.350 --> 36:54.520 Our final differentiator is our expanding market opportunity. 36:55.229 --> 36:58.566 We believe every company, across every industry 36:58.774 --> 37:01.318 will be transformed by powerful AI. 37:01.318 --> 37:03.737 It started with coding and is spreading to knowledge 37:03.737 --> 37:07.908 work, a $40 trillion market that we have just begun to penetrate. 37:08.325 --> 37:12.830 We count many of the world's largest and most complex businesses as our customers, 37:13.038 --> 37:18.002 including over 50% of the Fortune 500 and over 30% of the Global 2000. 37:18.043 --> 37:23.465 And as Paul described, our Fortune 500 customers spend less than 0.1% 37:23.549 --> 37:27.761 of their total expenses on Claude. As the frontier of AI advances, 37:28.012 --> 37:31.181 we expect it will open up entirely new lines of business. 37:31.348 --> 37:34.643 This has started already with cybersecurity, and we expect it 37:34.643 --> 37:37.980 will extend to life sciences and robotics in the coming years. 37:38.230 --> 37:42.526 Given the size of the opportunity we see and the strength of our business, 37:43.152 --> 37:46.739 we are investing aggressively in three areas: compute, 37:47.323 --> 37:50.326 customer value and safety and alignment. 37:50.492 --> 37:53.495 Compute is a critical enabler of our growth. 37:53.537 --> 37:57.708 It powers our R&D and enables us to meet expanding customer demand. 37:57.958 --> 38:01.378 We plan to make significant investments in compute over the coming years, 38:01.462 --> 38:05.090 given the compelling returns we see. That includes both short-term 38:05.090 --> 38:09.595 and multi-year cloud deals as well as data centers and chips we lease directly. 38:09.637 --> 38:12.931 We call this last piece Anthropic Direct, and we expect 38:12.931 --> 38:16.226 it will make up a growing share of our compute base over the coming years. 38:16.685 --> 38:19.688 Our second investment area is customer value. 38:19.813 --> 38:22.858 We are maniacally focused on ensuring customers 38:22.858 --> 38:25.861 get the strongest return from their spend on Claude. 38:25.986 --> 38:30.783 We do this through developing the most capable, aligned and useful models 38:30.949 --> 38:34.578 and through reducing customer costs as we make our models more efficient. 38:34.662 --> 38:39.124 For example, our Opus 5.5 model in September delivered substantial gains 38:39.124 --> 38:42.795 in model capability and 40% lower customer costs. 38:42.878 --> 38:46.590 This reduction was margin neutral to us because Opus 5.5 38:46.632 --> 38:48.592 is significantly cheaper to serve. 38:48.592 --> 38:52.221 While we sacrifice some near-term revenue when we reduce customer costs, 38:52.429 --> 38:55.808 in our experience, we quickly see a tailwind to revenue 38:55.891 --> 39:00.020 as the lower cost of intelligence drives greater usage and diffusion of Claude. 39:00.396 --> 39:03.232 The third area is safety and alignment. 39:03.232 --> 39:06.068 Safety is core to how we operate at Anthropic, 39:06.068 --> 39:09.113 and we've been investing here since the inception of the company. 39:09.196 --> 39:11.615 As model capabilities advance rapidly, 39:11.615 --> 39:14.743 we're investing even more in people, processes 39:14.743 --> 39:17.746 and compute to support our safety and alignment work. 39:17.830 --> 39:22.710 Taken together, we expect our focus on safety to remain a commercial advantage 39:22.876 --> 39:26.463 because customers and developers want models and products they can trust. 39:26.588 --> 39:30.968 The investments we're making in compute, customer value and safety and alignment 39:30.968 --> 39:35.097 will strengthen our business and broaden our opportunity over the long term. 39:35.347 --> 39:40.102 We believe Anthropic can deliver adjusted operating margins of more than 35%, 39:40.144 --> 39:43.689 with continued efficiency gains, a greater share of Anthropic 39:43.689 --> 39:47.359 Direct capacity, and more of our compute shifting towards inference. 39:47.401 --> 39:50.404 We see a path for compute and infrastructure expenses 39:50.404 --> 39:53.782 to trend towards approximately 45 to 50% of revenue. 39:54.366 --> 39:57.619 Second, we expect sales, marketing and partnership costs 39:57.619 --> 40:00.038 to trend towards 7 to 10% of revenue, 40:00.038 --> 40:03.208 as partner fees become a smaller share of revenue as we scale. 40:03.625 --> 40:07.296 And third, we expect personnel and other operating expenses 40:07.296 --> 40:10.466 combined to trend towards 7 to 10% of revenue, 40:10.549 --> 40:14.803 as Claude makes our teams more productive and we grow headcount efficiently. 40:14.845 --> 40:18.098 We believe we have the best financial model: leading compute 40:18.098 --> 40:21.643 efficiency, the greatest concentration of talent, and a culture 40:21.643 --> 40:25.522 that keeps us focused on where we can add the most value to enterprises. 40:25.773 --> 40:30.194 AI's transformation of the global economy has the potential to be the largest 40:30.194 --> 40:34.031 economic opportunity in history, and we are at the foothills of it today. 40:38.035 --> 40:39.411 You know, if today we look 40:39.411 --> 40:43.832 at things like cancer, Alzheimer's, heart disease, 40:44.249 --> 40:48.879 mental illness, these will be relics of the past 40:49.087 --> 40:52.049 in, I think, less than a decade. 40:52.049 --> 40:55.636 And I believe that the company which gets the risks right 40:55.677 --> 40:59.848 will also be the company that has the license to operate 40:59.848 --> 41:04.853 that society trusts and society believes can build the benefits in the right way. 41:05.437 --> 41:09.733 I think if we do this right, we can create an age of miracles 41:09.733 --> 41:14.321 where humanity is able to solve problems 41:14.488 --> 41:17.699 that have eluded us for millennia. 41:17.950 --> 41:21.495 We have to make sure that all the risks, 41:21.703 --> 41:24.706 all the ways that things can go wrong with this technology, 41:24.706 --> 41:27.876 we very carefully address those risks 41:28.085 --> 41:32.005 as we bring the benefits, and as we have a viable business at the same time. 41:32.089 --> 41:36.218 And I invite you to join us on the next phase of this journey.