video_OCP2026_Google_主題演講_智慧基礎設施_20261015

Hi folks. It's a pleasure to be back again here at Open Compute. So I'm really happy to see the amazing turnout as well. So my talk today is titled intelligent infrastructure for the intelligence era. And I've structured it into a few broad sections. So in the first half of the talk, I'm going to talk about infrastructure for intelligence. And I want to talk about systems for AI. I'll present some of the amazing collaborations we have had with OCP in designing systems for AI. And I'll also talk about some exciting announcements that we have as well. In the second part of the talk, I want to talk about intelligent infrastructure for the intelligence era. And I want to talk about AI for systems. I want to talk about how we have an opportunity to use AI to transform how we design systems. And I want to talk about how we can get together as a community to accelerate and amplify that. So with that as the outline, let's jump right in. I think you've heard this from all the prior speakers as well. We live in extraordinary times. Compared to prior revolutions, whether it's the industrial revolution or the information revolution, we are in the beginning of an intelligence revolution. One where AI is profoundly transforming all aspects of our life, health, education, entertainment, productivity, software, engineering. And at Google we are seeing this as well. Whether it's our consumer applications reaching billions of users or Google Cloud where we reach enterprises, small and medium businesses, vertical industries, and so on. AI is profoundly transforming everything that we touch. What we are also seeing is that these extraordinary advances in AI need corresponding extraordinary advances in the computing infrastructure that powers them. At last year's keynote, I had this metaphor where I said, "If AI researchers are the ones discovering new worlds, our space explorers discovering new worlds, all of us all of us here in open source open compute community here, the system designers are the ones building the rocket engines to make that possible." And we have certainly built some pretty amazing rockets the past years. At Google, we have fundamentally transformed how we think about computing for AI. In particular, looking at vertical integration across the AI stack. Optimizing for AI has led to some pretty radical innovations compared to how we have traditionally designed hyper scale systems. For example, we have introduced custom silicon accelerators, tensor processing units or TPUs now in our eighth generation. We have introduced custom arm CPUs and SOCs for agent take computing. We have introduced new memory technologies, high bandwidth memory. We have introduced new power and cooling, vertical power, liquid cooling. We have introduced new networking technologies, very sophisticated ICI torus networks, MEMS based optical circuit switching. We have even introduced new kinds of system designs, very very complicated gigawatt data center AI factories, high densification designs and so on. And underpinning all of these advances has been one common theme, the notion of vertically integrated cross stack cross disciplinary innovation. All the way from energy, data center, hardware, software, services to the models. So, we start off with our custom tensor processing units TPUs. We think about integrated compute storage networking. We pair that with open source frameworks like tensor flow, pytorch, jacks. We then co-design with the cloud stack, kubernetes, flex scheduling, and then we have this cloud hypercomputer, and then we continue to co-design this cloud hypercomputer with our Gemini models, with our agent gardens, all the way up to applications and services. So, from chips to systems, systems to platforms, platforms to ecosystems, this vertical integration has served us really well. Getting us orders of magnitude improvement in performance, 10 to 100x uh performance, power, cost. But, last year, we wanted to do more. I called on the community to go beyond ability to think about agility as well, to think about agile and fungible data centers for the AI era. And the community delivered in spades. I'm very pleased to announce that over the past year, we have worked with multiple of you, multiple OCP partners, to systematically think about the key principles behind designing agile and fungible data centers. One of my favorite examples is the open data center specification, where working with uh Meta, uh AMD, and Nvidia, we have converged on what does the data center fungible data center of the future look like. Things like server hall, aisle widths, uh IT facilities, interconnects, and so on. We've also worked with more than 50-plus OCP partners to converge on the blueprint for what agile and fungible data centers look like. Agile power, fungible power, cooling, uh thinking about uh common standards, common telemetry, and so on. So, I want to showcase three big announcements that we are talking about at Open Compute later this uh in the next 2 days. The first one is on power. We first introduced Mount Diablo that brought plus-minus uh 400-V DC to the uh data center. On top of that, we've now contributed LVDC that makes data centers DC-native all the way from uh 0 to 800 volts. And then on top of that, we have solid-state transformers that make it uh DC-native all the way from the grid to the chip. You heard a lot about that in the previous talk as well. On uh cooling, and can we go back to the previous slide, please? On cooling, we basically uh talked a lot about uh Can we go back to the previous slide? Okay. So, on cooling, we contributed uh Project Deschutes, which was basically our uh multi-megawatt uh open CDUs. And uh our hope with uh Deschutes was that we would have a lot of uh community engagements, and I've been very pleased to see the amazing uptake in the number of suppliers uh who have adopted Deschutes, and you'll see a lot of them in the show floor today. We also have Project Brazos, which brings liquid cooling goodness to air-cooled brownfield data centers. And then on networking, again, we have some pretty exciting developments there as well. Uh a few years back, Google contributed Falcon uh to Open Compute, and Falcon is our hardware transport protocol. It has a whole bunch of really nice features like multipath load balancing, congestion control, line rate encryption, and and our hope really with contributing Falcon was that others could build on that and and do more stuff. And today, Nvidia has done exactly that. From BlueField to ConnectX, we now have wire format intercompatibility with Google Mix. All three of these examples are fantastic examples of how we can come together as an Open Compute community and do some really amazing collaborations. I also want to talk about two other announcements that are particularly relevant to challenges that we face very specific to AI data centers. First, on power, Google has announced uh multiple collaborations with grids and utilities on demand-side flexibility, where we can unlock valuable power capacity by doing load management. And we have with the US utilities, we have a gigawatt of contracts that we have gotten into. And I'm very pleased to announce that Meta, Microsoft, Nvidia, and Google, we have worked together to contribute a specification that defines the interfaces and power variation specifications so that data centers and utility providers can collaborate together to think about AI data centers and AI workloads. And I'm really excited to see how we can build on that. On water, Google has announced some responsible water sustainability practices all the way from positive replenishment, thinking about funding local municipal infrastructure, thinking about context-specific water cooling. For example, when you have at-risk watersheds and so on, thinking about alternative water sources, waste water reclamation. And perhaps most importantly, transparency in how we think about water spending, having site-specific annual data that we can disclose as well. I hope again we can build on all of that and have a community conversation on this important topic. Of course, these are a few examples and we have a lot more where over the past year working with all of you, with all of OCP, we have systematically addressed many, many different aspects of what agile and fungible data centers for the AI era mean. And we also continue to innovate on all the other work streams in open compute as well. Google and open compute have a very rich history of collaborations and we have continued to accelerate and amplify our work together now totaling close to about 120-ish contributions altogether. Now, all of this is super nice and pretty amazing progress, but we have a lot more amazing opportunities ahead. Especially when we think about the explosive growth in AI workloads still ahead, we need to continue to innovate. And so today, again, I want to have a call to the community to go beyond agile and fungible data centers to start thinking about intelligent infrastructure for the intelligent era. To go beyond just thinking about systems for AI to thinking about AI for systems. And to illustrate why I think this is so exciting, I want to give you four vignettes how how we have been using that at Google. The first example I want to talk about is using AI to transform systems hardware. And here the showcase example that I want to use is Arc agent. So Arc agent is an agentic framework that uses an evolutionary AI algorithm. In our case, we used Google's Alpha wall and to think about designing completely new computer architectures. So using Arc agent, we were able to design cache replacement policies that were not only superior to human-designed state-of-the-art policies, but not only championship-winning, leaderboard-topping contributions, but we were able to make those contributions three to five times faster engineering velocity than traditional approaches to designing systems. And that's just one example. We have seen numerous examples where EA can be used as a force multiplier for hardware design. Everything from chip design, RTL synthesis, place and route, verification to system design, PCB design, validation, testing, co-design, simulation. The opportunities are really significant here and I think there's a lot we can do here. The second vignette I want to talk about is to use AI to transform system software. And here the example I want to use is Mantis, an open-source framework that we have contributed and it's basically a multi-agentic review harness that can systematically and pervasively do vulnerability scanning and bug fixing. And again, the results are pretty amazing. Not only do we get pretty strong accuracies in how bug fixing works, we can also do that at fast enough latencies that every single piece of code that we check into production has continuous pervasive scanning and bug fixing. Another example, Shikoku, that we have also contributed to open source, does thousands of open source contributions, scanning and bug fixing, and it does that at 97% accuracy. We have also used AI to transform software in other ways. Take the toy aloud, automatically translate TensorFlow to Jax, translate from one ISA to another, x86 to arm, arm to risk five, translate from C++ to rust. We have used AI for workload scheduling, for congestion management, for performance optimization, for power management, for resilience, for reliability. Again, we are just scratching the surface of what is possible, and there's a lot of opportunities here as well. The third example, continuing through the vertical stack that I talked about, is around AI to transform data center design. And here I want to talk about Prometheus, again an open source contribution where we have combined physics-based climate modeling with AI ensembles to predict extreme weather events. And what Prometheus does is it solves the incredibly hard problem of how do I balance upfront costs of infrastructure, how do I provision cooling very early on for the 15 years of data center staff, with the downstream costs of downtime, where you have risky costs when you don't provision enough cooling. And not only does using AI allow us to find this balance in a much more optimal manner, AI also unlocks new optimizations. We can have multi-epocal prediction, decadal, annual, weekly, with corresponding optimizations. We can do site-specific optimizations. Later in the conference, you'll also hear other talks where we talk about how we use AI for sustainability to process huge amounts of complex information to be able to look at life cycle analysis. We can use AI in the data center for power and cooling, for dense designs. And then the last vignette I want to talk about uh is something that uh is not discussed as much. It is about using AI to transform data center operations and capacity delivery. So, again, we have talked in the past about how robotics can be incredibly helpful in data centers to reduce toil, uh automate drudgery, and really amplify human productivity. But, I also want to talk about how we can use AI to improve uh capacity delivery. How can we use AI to automate uh manufacturing? How can we use AI to do uh better supply chain management, supply-demand management, deployment, capacity planning? Again, an area that we don't spend as much time thinking about, but AI can profoundly transform how we think about data center delivery and operations as well. Hopefully, these examples gave you a sense of why I'm excited about how Google is using AI to transform system design. But, going back to the vertical stack that I talked about, we have used AI pretty much at every layer of the stack and shown some pretty amazing benefits. I often like to joke that uh everything tastes better with a sprinkle of AI, and and and I think you will see that as well. So, this is all nice, but in some respects, we are at least only at the very early stages of this technology inflection. I often find it helpful to think about this framework as we think about the AI transformation journey, and as we think about how we can work together, and I call it IUR. And it's it's obviously a good uh framework for collaboration, but it's also a memory device for us to think about how AI transformation journeys work. And I stands for impressive, U stands for useful, S stands for sustainable. And often AI journeys begin with impressive, where we try to think about an innovative new application for AI, and then it goes to useful, where we think about how do we optimize for the common case, how do we optimize for corner cases, how do we make it generalizable. And then it goes to sustainable. How can we make it cost-efficient, scalable, secure, safe, and so on. So, with this framework as context, I want to talk about some call to action for us as a broad community. The first one, of course, is responsible AI. So, AI is not a panacea. And we need to be very intentional about thinking about the three H's, hate, harm, and hallucination. And we want to be very responsible about thinking about security as the default, privacy as a first-class design consideration. How do we think about information quality? How do we think about explainable AI? And this is a place where community and partnerships are really important, not just across Open Compute, but across industry, academia, research, government, civil society, to really think about what are common AI principles we can agree on, what are responsible AI practices, what are fair AI practices that we should look at. Second, I've talked about systems for AI, I talked about AI for systems. We also need to be thinking about systems for AI for systems as well. We need to be thinking about data collection. How do we design the system to be scalable, modular, latency? How do we think about the flywheel for self-improvement? Again, something that OCP is really good at is designing systems, and I think we have a lot of opportunities here. Third, I want to also think about how we can think about AI as a full-stack redesign. We need to be thinking about the end-to-end AI infused workflow. And too often, we spend time optimizing one part of the stack only to realize some other part is holding us back. Like the picture on the left talks about software development life cycles, and you could spend a lot of time increasing the velocity of code development only to find that testing or validation is a bottleneck. Or you could think about using AI to develop hardware accelerators really fast, but if you don't use AI to also develop the corresponding software stack, you're not going to get the benefits. And again, this is a place where OCP offers a really nice opportunity for collaboration because we can come across multiple companies, multiple disciplines to really think about the problem holistically. And then last, I want to leave with an observation around how AI transformation journeys work. How many people here know what this picture represents? Okay, a few hands up. So, this is a Ford Model T, and uh and the key point really is that the very early cars looked a lot like uh horse-drawn carriages. And this is how technology transformation usually work. You start with basically taking the existing approach and move it to the new technology, and you can see this across multiple things. Like if you look at um digitization of media, we started with ripping CDs and uploading it, and then we did searchable songs and single songs and new business model streaming, and then we finally ended up with uh YouTube shots or TikTok videos. A completely digital native format that we wouldn't have even imagined a few years back. And so, the AI for systems journey is also going to go through similar uh transformational paths. We're going to go to from lift and shift, we're going to then go to AI augmented to AI forward to AI native. And AI native is where the true magic is going to happen, where you're going to have truly transformational opportunities that change how we think about design and uh value to our customers, and I'm very excited to see how we can all work together on this. So, I will end with how I began. We live in extraordinary times. Extraordinary advances in AI powered by extraordinary advances in infrastructure. One example where we have extraordinary infrastructure deliver societal change is around AI and sciences. Whether it's biology or material design or computational science or or new robotics and so on, we have truly transformed science with recent advances in infrastructure and AI. And in particular, I was very lucky to work on a recent launch of Alpha Genome on Google Cloud. And Alpha Genome is really notable because it's been able to have about 9 billion variants of single letter transformations to the human genome. And interestingly enough, in its very first clinical application, Alpha Genome was able to do something that traditional medicine could not. It was able to find with great precision, identify the genetic marker for a very rare unsolved case of pediatric epilepsy. And you can read more about that in the paper that I have linked, but that particular young child and his family who had given up hope now have a completely new future. And so going back to the metaphor that I started earlier, maybe we shouldn't be thinking about building rockets for space age, maybe the metaphor is that if AI is helping us decode the secrets of life, all of us here in Open Compute are the ones building the high precision microscopes that makes it possible. And so as we think about working together to build the intelligent infrastructure for the intelligence era responsibly to make all of these new advances possible, I want to once again thank all of you for your community, for the collaboration, for the engagement, for all the innovation and creativity. I hope you have a fantastic next 2 days ahead and thank you for all your time. Thanks.