In this keynote, Sonya Huang frames the current surge in artificial intelligence as the "AI race," a transformative technology wave built on six decades of prior advances in semiconductors, networking, the Internet, the application economy, and mobile computing. She argues that AI is the ultimate wave because it combines massive compute, vast data, and ubiquitous devices, enabling neural networks to "eat software"—automating code, self‑driving cars, customer‑support agents, and even software engineering itself. Huang highlights AI’s shift from a productivity boost to a customer‑experience revolution, citing ChatGPT, Claude, and voice agents as examples that outperform traditional search and call‑center models. She estimates a $10 trillion economic opportunity, far exceeding the $600 billion software market, by moving from SaaS workflows to end‑to‑end AI‑driven work. She then presents a layered "AI stack" metaphor: compute (raw silicon), foundation models (the "brain in a jar"), digital embodiment (models that can write and execute), and physical embodiment (robotics). She details the compute arms race, noting hyperscalers’ capex jump from $25 B to $300 B annually, and explains how scaling models yields emergent capabilities, with reinforcement‑learning‑enhanced LLMs (OpenAI’s "Strawberry" team) marrying breadth and depth. Tool‑use, agentic frameworks, and multi‑agent systems are unlocking digital embodiment, while advances in hardware, simulation, and reinforcement learning are reviving robotics, exemplified by Tesla Optimus. Huang contrasts two approaches in the application layer: foundation‑model companies leveraging massive technology hammers and vertical specialists building domain‑specific solutions (e.g., Harvey for legal, Open Evidence for medicine). She describes this landscape as a fractal, with endless niche opportunities. Finally, she forecasts four near‑term breakthroughs—continual learning, video/world models, robotics, and AI‑driven scientific discovery—arguing they will further accelerate the AI race. Overall, the talk blends historical context, technical insight, market sizing, and forward‑looking predictions to persuade investors and builders that we are at a pivotal moment in a $10 trillion AI-driven transformation.
Chapters
Sam. Good morning everyone. Thank you for hosting me today in your beautiful country, this beautiful city. I'm delighted to share some observations on what's happening in AI from a venture capitalist point of view. And so from the 30,000 foot point of view, what are we seeing? I frame this talk as the AI race because what we're seeing play out in front of us is very much the biggest race of our lifetimes from a technology perspective. So let's get started. Technology comes in waves. And if you look at the previous waves of technology, it started with the semiconductor wave of the 1960s, giving us the power of raw compute. Beyond that, systems and networks made those computers useful, connected them. In the 90s, the Internet revolution completely changed the way that we exchange and access information. In the 2000s, the application economy shaped the way that we as consumers get to experience the power of technology. And of course, the 2010s gave us the mobile wave. And it is the culmination of 60 decades of technology progress that has given us the necessary and critical conditions for AI to take place. And AI is not just another wave, it is the ultimate wave. This is the technology pinnacle that computer scientists have been dreaming of since the 1960s. But there's a reason it's only arriving now. Because it took six decades of hard work in order to have the raw compute, the GPU clusters that are capable of training models of this size. The Internet, that's given us trillions and trillions of tokens worth of training data. And that's before you even look at YouTube data. The mobile phone that's given everybody a supercomputer in the palm of their hands. It's the collective progress of all of these waves of technology that's led us to this point in time that computer scientists have been dreaming of for decades. And so I think we should all feel incredibly fortunate to be alive and building at this point in time. About a decade ago, Mark Andreessen coined the very catchy term neural nets. Sorry, Mark and Jason coined the very catchy term neural net software is eating the world. And he was right. Every business became a digital business. But I think what's happening now is something even more profound. Neural nets are eating software. So instead of computer scientists, programmers, programming machine with code manually, what's happening is neural nets, machines are figuring out how to code themselves. We see this in self driving, where millions of lines of handwritten code of C are being replaced with end to end neural nets. We see this happening in customer support, where instead of having these intricate decision trees and IVR systems that have been crafted end to end neural nets are now powering customer support agents. And then finally we see this in software engineering itself, where 30 million professional software engineers have dedicated themselves to the craft of how to power of how to program a computer. And now, thanks to Anthropic Sonnet, many other models, the neural nets are now giving the power of coding into everybody's hands. And so I think what's happening here is neural nets are not just a side project, an academic fascination, but they very much will be the drivers of the economy going forward. Several years ago, Jensen Huang said something provocative at one of our Sequoia events. Every pixel will be generated, not rendered. And I think this extends not just to pixels in the visual universe, but every job, every workflow will be generated. I don't think we've all properly grasped yet what does that mean for the world? What does it mean when video games and TV shows can be generated on the fly? What does it mean when you can personalize a child's education, the way that they learn and what their personal passions are? I think it's exhilarating to think about what that future holds. And I think we're just scratching the surface right now. Okay, let's do a quick mindset check with every technology wave. There's also naysayers who can't wait to wait to claim this thing was overblown. We saw it happen with the Internet, we're seeing it happen right now. This MIT study came out that 95% of enterprises are not getting value out of their AI pilots. And there's some truth to this. Like if you are a CEO of a company and you are decreeing top down, we're going to become AI native and you buy a few AI tools and call it a day, you are going to be disappointed because AI requires an entire mindset shift. But this isn't a top down revolution. And so as with any other pyrotechnology, if you look at the people on the ground floor who are actually building with the technology, developers, tinkerers, enthusiasts, and you see that they're actually getting to understand the quirks of these models. Where they fail, where they succeed. If you see what's happening at the ground floor, I think there's reason to be incredibly excited because I think what's happening is AI is driving a fundamentally better customer experience. ChatGPT is just a better customer experience than traditional search. Waymo is a better user experience than getting a traditional ride hailing vehicle, even customer support. This is one of the most fascinating things, 10 years ago, you call customer support, you get put through to the answering machine and you just say, talk to human, talk to human, talk to human. And now what we're seeing is that these voice agents, we're going to hear from MADI later in this conference, voice is becoming so expressive and digital. Customer support agents are capable of being so knowledgeable, so empathetic, so humorous, that people are actually preferring them to traditional customer support agents. So AI is not just a productivity revolution, it's a customer experience revolution. And we think the size of this opportunity is massive. If you look at what happened in software, software was a $350 billion market a decade ago that grew to $650 billion, of which $400 billion a year is SaaS. But SaaS is fundamentally workflows on top of databases. You still need a human to execute the work. What's different about AI is that the neural net is capable of doing the work end to end. And so that means that you're not only playing for the $600 billion software opportunity here, it's the $10 trillion economy that comes into view. Okay, section two, let's talk about what's happening for the race for AGI. From the technology layer out. And I think of AI as a layer cake. At the very bottom you have compute. This is the raw silicon that makes all of this possible. Above that, you have the foundation models. I call this affectionately the brain and jar because it is super intelligent. It's the core fabric of what makes this all work. But it only has read access, it can't write or execute. Digital embodiment adds write and execute access. So these models can actually do work on our behalves in the digital world. And finally, the final layer, physical embodiment. We're just on the precipice of a robotics revolution renaissance as well. Let's hop through each of these layers in a little bit more depth. Starting with the race for compute. This is the single biggest technology build outs we've seen in history. The hyperscalers have gone from spending $25 billion a year on capex to north of $300 billion this year. And this is the fuel for everything that's happening in the model layer. This is the reason we have these models that are capable of being such incredible reasoners, such incredible generations of video. It's because the hyperscalers are playing a game that they cannot afford to lose, pouring compute into the ground. And the rest of the ecosystem gets to benefit. Similar to what we've seen in prior waves. And why does this matter? Because AI is still fundamentally compute constrained. So who here has played with a new Sora video launch? Okay, nice. Encourage you guys all to try. It's incredible. And it's a direct result of OpenAI having the access to a lot more compute, because the reality is the technology already exists. We can build incredible things, but there's not enough compute out there to power all the things that we as consumers want to build. And so as more compute gets put into the ground, this AI customer experience frontier gets to be experienced much faster. Okay, let's move up layer in the stack to the model layer. I think the most important breakthrough of the last decade was what happened in the large language model race and specifically increasing returns to scale. So typically in almost every industry, the law of physics is diminishing marginal returns. And what researchers found when scaling GPT2, which is on the order of a billion parameters, to GPT3, which on the order of 100 billion or so parameters, was that emergent capabilities started to emerge. And so you throw more computer at the problem, and the models were now capable of Q and A, of understanding language, of telling jokes. And so this kicked off a gigantic race to scale because these problems that had eluded natural language researchers for decades. How do we get the machine capable of understanding human language? We're coming out of the machine purely as a function of scale. And we're seeing this happen in other modalities like the video market as well, where the more computes and data you throw at a video model, the more capable it becomes of actually understanding and rendering real physics. I think the most important AI breakthrough of the last couple years belongs to OpenAI's Strawberry team. And so if you look at historically the major AI paradigms, large language models are incredibly broad and generalist, but they're thin. They only think on the order of a few seconds before answering. Reinforcement learning. By contrast, if you look at systems like AlphaGo, they're incredibly deep, they offer super intelligence. The best machines can beat the best human go players, but they're incredibly narrow by definition. And the biggest breakthrough that the OpenAI Strawberry team had and that other reasoning teams have had through the industry is figuring out how to bridge large language models with reinforcement learning together. So you get the breadth and the generality of what's happening in the LLM ecosystem combined with the depth of reinforcement learning. And today reinforcement learning is just the cherry on top of the pre training cake. Imagine what happens when that paradigm starts to invert. Let's move Up a layer. You have this brain in a jar, but it can sit there pontificating and theorizing. But one of the most exciting breakthroughs in terms of actually unleashing AI as a productivity game is allowing these models to use tools on our computers to make things happen in the digital ecosystem. And thanks to things like mcp, agentic frameworks, different harnesses, multi agent systems that are now being reinforcement learned and seem extremely promising, we're starting to bring this raw layer of intelligence out into the digital world. And finally for the last layer, it's physical embodiment. And robotics just went through about a decade of, you know, we're in the peak of the valley of despair right now. But I think if you look at the foundation model companies, they're all very clearly making investments in robotics now, because I think they see the same thing that we do, which is that you now have all this intelligence bottled up in the digital world. But if you actually want to change the world, you have to move into the physical economy. 85% of GDP is in the physical world. You need to close the loop on these systems. You need video data in order to train the models better. And thanks to a lot of innovation that's happening in this layer of the stack, from hardware to learning and simulation to reinforcement learning, neural nets that are now making their way end to end neural nets are making their way into robotics. We actually have hope in robotics for the first time in a long time. And I think Tesla Optimus in particular deserves a lot of credit for inspiring and invigorating the entire robotics industry to take flight. Who here seen this image? Okay, this is the image OpenAI released when it won the International Math Olympiad this year. And this was an oh, shit moment for me, because if you had asked anyone in the industry a couple decades ago, sorry, a couple years ago, most folks would never have guessed we would be here today. We're AI can solve math competitions and it's not about winning a math medal. It's the fact that these systems can actually reason. People used to think LLMs were just stochastic parrots. You can't stochastic parrot your way to winning the International Math Olympiad. So I think this was an incredible demonstration of what's happened from this technology race. Because what happens when you start thinking on the order of an hour means you can solve the International Math Olympiad. But what happens as you scale up that compute for further? What happens when you can think on the order of 10 hours, which anthropic just released? What happens when you can think on the order of hundreds or thousands of hours or millennia. I think scaling, inference, time, compute and reasoning is what people believe is the most promising path towards achieving scientific breakthroughs. And lest you think this is all about nerds nerding out to win math Olympiad medals, I promise you it's not. This is affecting the real economy. OpenAI put out this evaluation GDP Val if you compare human experts to the models at a variety of tasks across the knowledge economy, we are getting very close to the 50% mark where the model's on parity with industry experts. Let's talk about the race for the application layer. There's two killer apps in AI already. ChatGPT, 700 million weekly active users, more than $10 billion in revenue. Front Door for AI Claude, more than $5 billion in revenue. Coding revolution. If you look at how the race for the application layer is playing out, it's a race between the foundation models that we just talked about in the prior section, who are moving up the stack from the model layer and the application company. So they're moving from the customer back and each has its own benefits. The technology players have magnificent technology hammers. They have account distribution, they have brands. But the people building from the customer back have deep understanding of customer pain points, which means they can take an opinionated point of view on the ideal workflow to solve these problems. And so in reality, the way we think it's going to play out is a lot of the application layer will be won by the foundation model companies. They have the benefit of these magnificent technology hammers and the benefit of being able to vertically integrate and reinforcement learn in a when getting their user data to improve the underlying models. But there's also this entire universe of application layer companies to be built. And so I very much think of the application race in AI as a fractal. You have things like OpenAI consumer anthropic for developers at the root of this fractal. But then you have companies like Harvey that are taking all this magic and making it useful for lawyers. In Harvey's case, getting legal data sets, working on legal reasoning and workflows. You have companies like Open Evidence that is used by more than half of the physicians in the United States and just scored 100% on the US medical licensing exam. And so the reason I call this a fractal is as you get to the frontier of the application space, the fractal just keeps multiplying. There's so many opportunities, so many different workflows, different customer pain points, different companies to Build. Finally, just to know them, what's ahead. Here are four predictions for what we should expect to see in the next 12 months. And just so you all know, I revise this and throw it out every day or so because AI is moving so quickly. But one, continual learning. This is one of the biggest handicaps of today's models and all the foundation research companies are working on it. Two, Video and world models. We saw this recently with Meta Vibes and OpenAI Sora. I think we're about to see a video revolution. 3. Robotics. If you look how quickly the rate of progress has been in companies like Tesla, Optimus Figure, Physical intelligence, skilled. It seems very clear that we're reaching a breakthrough moment for robotics. And finally, science. Some of the top physicists I know chose to leave their fields and work in AI because they thought that working on AI was actually the fastest way to solve the greatest unsolved problems in physics. And I think that's we're going to see that happen. We're going to see it happen in physics, we're going to see that happen in biology, we're going to see that happen in chemistry, material science. So these are some of the pillars that make me incredibly excited for what's ahead. So, quick wrap up. This is the biggest technology race of our lifetimes. We think it's a $10 trillion opportunity. All the critical ingredients are finally here and I think we should all feel incredibly lucky that we're alive at this point in time. The race for AGI from the technology out is firing across every layer of the stack, from compute to foundation models to digital embodiment, now even physical embodiments. The race for the application layer we should think of as a fractal. There's so much to build. And finally, what's ahead? Continual learning, video models, robotics, science, and we'll see what else surprises us. Thank you very much for the time today.
