ECEN 681 Seminar Spring 2026 Mark Papermaster

Thank you, Dr. Ready. Thank you for coming. So glad to be here. Well, it's such a pleasure to be with you here today, and it's been many, many years since I've been on the campus. So I could hardly recognize it, frankly, when I pulled up because it's clearly more than doubled in size. And the innovation here is astounding. So we always like to joke, my undergrad was at University of Texas. But as soon as I got into industry, my goal was to hire the best engineers, and so I started hiring right away, you know, diving into the pool of engineering graduates coming out of Texas A&M because it's such an outstanding program here today, and every year it's getting stronger. So thank you, Dr. Ready, for your leadership, those of your faculty, and of the administration here at College Station.

Well, look, I just can't imagine actually a better opportunity for me to come back onto the campus, as I said, after a very, very long absence, because what a time that we're facing right now. I had the fortune of entering the industry in a time of incredible disruption. It was the early 80s. The PC had just come out. The internet was just starting, and so we just started getting interconnectivity. And that spawned, obviously, the amazing revolution of computing that we all have with us everywhere we go. And the interconnectivity that we just depend on. I mean, think about it. If you didn't have a connected device, you almost feel nervous, like, what's going on? I'm not fully connected. So, you know, it was certainly a time of disruption and upheaval, but what we face now, I'd say, certainly matches that time of disruption. I'd also put it akin to the electrification that we did, the whole industrial movement we had. That's really what we're facing now.

And I get questions all the time. I was just Monday speaking to a number of financial analysts. I mean, they're analyzing in incredible detail the trends, and many of them see this, but a number of them are still skeptical. They say, 'Mark, isn't this a fad?' like we saw the internet of things, and we always hear about these disruptions, not all of them are truly disruptions. But the reason I don't have actually any doubt that this is a major disruption that we're seeing is because of the impact that we're seeing, the impact that AI is now really just in the last couple years has risen to a set of capability that is going to fundamentally change how we learn, how we discover, how we communicate, how we practice our daily lives at work, at school, and our home lives. I mean, I know it sounds like hyperbole. It's just not. We are right now on the cusp of inflection. And I'll say even just the last three months we've seen at AMD that cusp, and I'll describe a little bit of those trends going through here. I'll describe what does it mean for computing. I've spent my four decades in computing across IBM and Apple, Cisco, and as Dr. Ready said, the last 14 years as CTO and most all of that running the engineering team at AMD. It's been an incredible opportunity, but now is an incredible cusp of inflection.

You just have to look at a few stats that stand behind it. Look at that. Over 350 million users of AI tools. I mean, it's the fastest adoption of anything. I'll actually show you some stats later on some of the meteoric adoption that we have. And when you look at the size of the market, obviously, AMD, this is something we focus on. We are a company that develops that fundamental computation. It is going to be a $1 trillion market by 2030. And we were one of the first to really call that out. Lisa Su called out that projection a couple years ago. And again, there's a lot of skepticism. Is this a fad? Is this not? And people said that seems quite optimistic. Nobody's saying that now. Our peers in the industry are calling out the same. This is going to be a trillion-dollar industry imminently here. And that is the semiconductor portion, the compute portion of the industry. So it's going incredibly rapidly. And you know what, I'll just point to the last stat there that shows, you know, projection, obviously it's quite a bit of a guess, but it's an educated guess of a number of 78 million new jobs that will be created by AI. But I'll caution you that as well as creating the number of jobs, it will also displace existing roles. So it really will be a time of disruption, and it's a time to think about, thoughtfully, how AI is transforming and how it changes our ability to be more productive and to move with that technology trend. And I'll comment about that at the end of my presentation.

Well, first, my basic premise as to why this is not a fad and a trend is that AI is not a technology impact on a server, on a section of the industry. I mean, you heard me say it affects every aspect of our lives. It affects every aspect of computing. So AI is going into every computational device, from the biggest supercomputers to the smallest devices, are all becoming, or most of them are already, AI acceleration enabled because AI is running everywhere. But in addition, it's affecting every industry. I talk to customers every single week. I'm talking to CIOs and CTOs across pharma, defense, industry, manufacturing, transportation. I'll go on and on. You pick the area of focus of business, and top of mind is the discussion around this inflection point we're at and how we can partner to move more quickly and to be more productive. And I'll talk later on the last graphical icon I have on here in science. AI will fundamentally change how we do science and will fundamentally speed our time to discovery.

So I talked about this incredibly rapid adoption. It's hard to believe that the first accurate, you know, natural language capability of AI to be able to generate natural language from an interface was really 2011, 2012. Dr. Hinton did some similar work that got the accuracy to where it caught everyone's attention because AI has been around for decades. But it wasn't productive, and it was the start of really neural nets of the forward path of creating a set of statistical weights based on a training. But the recursive loop, the recursive loop to adjust those weights and to iterate billions of times to create more accuracy. So that was just in the early part of the last decade. And think about it, it was just over three years ago that ChatGPT was released onto the world. Just over three years ago. And now you don't even think about it. You're using, whether it's ChatGPT, Gemini, or your favorite generative AI model, you're using it like you're using your phone and your PC together. It's used in your daily devices, and you are using it, if you're like me, you don't even think about it, it's just part of your daily routine. So we've gone from very few to over a million active users that just shot up right within months of that introduction in November 2022. It just was the steepest adoption at the time of any new application to over a billion active users today, on its way in just a few years to being over five billion. It's hard to even use that term, but it is indeed going to be over five billion users every single day leveraging this technology.

And I comment that we're focused on the massive computational demands behind it. We think the biggest clusters will be really needing 10,000 times the level of computation we have today. So that's what drives the state-of-the-art. It's the big foundational models that are collecting the moats of information that can be pulled together. And then there'll be a continual of models which are smaller, more focused on the areas of expertise that they need to be able to respond to. So you're going to see an entire continuum of both models and the associated amount of compute that you need to service those models. So it's literally from the largest supercomputers to the smallest of devices and model sizes that match with that.

So I showed you that ChatGPT, you know, went from like zero to a million active users today faster than any application. But now that we're at a billion active users daily of AI, I found it astounding what happened with what's going on really right now with the agentic AI. So you see on this an adoption rate of PyTorch, the adoption rate of Microsoft Visual Studio, you know, a very, very steep adoption curve, that's incredibly fast adoption rate. But when Open Claw came out and allowed all of us, all those billion users that are running AI daily, to figure out how to chain together their kind of task that they do, not one task, but a workflow. You know, if it's at home, it's like I want to check my calendar, I want to do this, I want to fire off a check of the weather, I need to send this message. I mean, you can chain together whatever you want as long as you can access those apps that you were manually doing spread out over time, and it can be done literally all practically instantaneously. And look what happened on that Open Claw adoption rate. It's just through the roof. I was shocked. I used this in a presentation about two and a half weeks ago, and I saw Jensen use it in his GTC presentation last week. So he's equally impressed with this adoption. We all are. I mean, it's just incredible. And I'll talk, I won't have as much time to talk about it, but it is agentic AI. It is the ability to take what we all do over a course of a day, weeks, and months, and being able to collapse it into an agentic flow that can run, if it's containable, run within minutes or hours, or it might take an overnight, but it's literally collapsing what we've done in days, weeks, and months to hours. This is the mark that when I say the last three months have convinced me we are at that cusp of inflection, it is agentic AI that was the missing piece to really provide all of us the opportunity of the promise that we all had of AI productivity gains in what we do every day.

Well, for those of us in the chip industry, this comes with extreme challenge because when we run the computation, it's got to be fed an amazing amount of data from memory and I/O. It runs an amazing amount of computation, and it burns an amazing amount of energy. I mean, the power of our CPUs and our GPUs is going up dramatically. I'll show you some stats on that in just a minute. But you know what it comes with it. You hear about, I included this graph on memory because you hear about the memory shortage. Memory prices have gone up multiple fold just in the last months. Why? Because that insatiable demand. There just wasn't forecast, and the demand far exceeds the supply. You're going to see it affecting price of laptops, phones, even the availability to get that. The data centers will be commanding and paying top dollar for a lot of these componentry, not just the memory, but the CPU and GPU. So we're actually going to have a temporary shortfall of computing devices as this takes off. But from those of us that are designing and producing that computation, the systems underneath, what we see is the graph that you see in the lower right. We see that if we don't solve and mitigate, we won't ever really solve power efficiency. It's an ever-ending quest for more and more power efficiency. But if we don't make dramatic improvements in energy efficiency of computation, then even when we can build all of the memory and I/O and the CPU and GPU and acceleration devices underneath, we're still going to be gated because we just won't have the energy to power these devices. It's an inexorable trend that we're on. This is a nonlinear, exponential graph, of course, that I'm showing.

And you look at this zetta-scale type of computing. I mean, this is just massive computing. We are currently at an exascale. So AMD actually powers the number one and number two supercomputer in the world with our CPU and GPUs. Number one, Livermore Lab, number two at Oak Ridge, and they're above an exascale of high-performance computing. So that's full and double-precision math being applied to solve the world's biggest problems. But look at the power consumptions, you know, many, many megawatts for that single compute cluster to provide those exaflops of high-precision compute. And what are we working on now? We're working on zetta-scale and even the concept for yotta-scale beyond that. So the innovation that we feel is needed and what we are doing across our research as well as our development teams is equally how do we get the computation more efficient, how do we get CPUs, GPUs, networking, all the way through the software stack working more closely together, but how do we do it with much more energy efficiency because we are in the gigawatt scale. It's just astounding. Like you saw AMD signed a deal with, first with OpenAI to build out 6 gigawatts of compute in just the next four years, just recently with Meta another 6 gigawatts of AI compute capability. So this insatiable demand for computing needs insatiable curiosity and innovation as to how to solve the computing needs in different ways than we've done before. You again, you'll hear me say probably 10 times in this presentation. You could not be, for those of you that are students, you could not be entering industry or academia at a better time if you like a challenge. You can't have a better time if you're, as faculty, to be looking for challenges with your grad students and your undergrads to take it on. I mean, it's just phenomenal need for innovation, and we need that industry. We're driving that industry internally. We want to collaborate across academia and government to speed the innovations. I called it from exaflops to laptops, the biggest supercomputers to the smallest devices. And that's what I'll jump into here.

I want to walk through and just share with you our view of that opportunity space, how we think about, I'll show you some examples of the products that we're developing and why we made some decisions we did to be able to both drive up the computing and drive down the relative power consumption. Every generation demands more power. But we're providing many, many more flops of computing versus the additional power ratio that we provide each generation across everything from the cloud to PCs and the edge, to what will be certainly a large market today with embedded, but will be an exploding market with physical AI, which again I'll talk to a little bit. Let's start with supercomputing and data center. We found ourselves as a chip company, we had to reinvent ourselves to be able to design at a rack level. I'm showing you the rack solution that AMD will be shipping at the end of the year. It's already back in our lab. It's our newest 2-nanometer EPYC CPU. So, it's an incredibly advanced CPU. That version of our CPU in one chip gets up to 256 high-performance and very, very energy-efficient x86 CPUs, and it's designed with our new generation of Instinct MI450, and we build up actually 72 nodes of this CPU and GPU complex in a single rack.

And why did we have to, as a chip company, why did we acquire a company called ZT Systems that designs racks? We sold the manufacturing. We're not actually in the rack business. But we determined that we had to reinvent ourselves. You can't design at a chip level. You can't be in that bunker designing and optimizing chips if you're not thinking about how you're going to optimize the performance and the energy literally at a rack level across hardware and software. So it just redefined the entire problem for us. But we went about it the way that we always do at AMD. We're committed to open standards. So we partnered actually in this case with Meta. So we co-designed this Helios Rack with Meta. We donated to Open Compute Project. We're all about open standards at AMD. We want to win on the merit of the technology we provide, not on how we give people flexibility to create different variants of the computing that we need. So we're very much open in enabling our partners to really run with what we do. And I'll talk later, it's equally across the software stack. So Helios will be the most powerful AI computer at the end of this year. It's targeted to take on Nvidia's Vera Rubin. And it won't beat it at every workload. It's going to be, you're going to see lots of racing as to who's got the best, and there's many, many different workloads across AI, of course, but the ones that we targeted we expect to have absolute leadership performance.

And I just want to show you, as I said, a little bit about how we had to design differently. This is the sled. So we pulled the sled right out from the rack of the compute tray. Equally, there's a network tray that has all of our switches and interconnectivity, but this is the compute tray. Again, this builds up to 72 nodes of CPUs and GPUs. And what you see is a very, very tight design. And why does, again, why did it change the way we thought about design is it comes back to energy. So you had to think about the proximity of all of your computes. Where are the bits flowing? How are you minimizing the path from memory to the computation and access to the I/O and storage devices that you need. And so we optimized this very, very dense sled. What you see in the back is the GPUs. Each of those GPUs, it's not like if you pull the cover off, you see one chip. If you pull the cover off, you actually see over a dozen chiplets. You see, what you see is that we broke out of the GPU computational devices. The I/O and memory controllers, the actual high-bandwidth memory is vertically stacked up to eight, even up to 12 high. So it's a combination of 3D stacking and lateral connectivity all on a silicon carrier. So it's all silicon to silicon on each one of those chips. Taking all those chiplets and putting them together, driving down the energy consumption as they talk to one another dramatically, driving it down. This was an innovation we drove, the first to drive this type of 2.5D and then 3D where you have vertical stacking and later lateral silicon-to-silicon connectivity. Our first product of this nature was in 2015. We were the first one working with TSMC, what's now called CoWoS, the technology that allows us connectivity. AMD was the first because we knew we had to take this problem on. We knew computing was going to be heterogeneous. It needed different devices. We architected for modularity. So we're the first to chiplets. And now the pinnacle is in this Helios design. You see this very, very tight integration of those GPUs in the back. There's two CPU trays up front. Those CPUs are also chiplet-based, mini CPU Zen core chiplets around the center I/O and memory controller, and then a highly programmable AI network interface chip. And so it's an amazing amount of technology. And again, this is just one sled. You can go to the back side, very much collocated is the network tray that you can pull out, and then the back side is very, very high-efficient connectivity. Today it's copper, lots of space because it's going to photonics in the next two to three years. You're seeing photonics is finally finding its day. It's been in the labs for a long time, or it's been used for network scale-out, but now it's coming to rack scale-up to provide yet again more energy efficiency and more performance of how we build these computers.

So why are we so motivated? Obviously, we're a business, and so our shareholders demand that we'd be out there, we'd be pressing our innovation to have competitive and leadership products out there. But we're a different kind of company. Again, we don't do proprietary. We're all about being open. We're all about collaboration. And the other thing that's different about AMD is we're motivated to just really make a difference. You go talk to any of the engineers at AMD. Talk to some of you, I've talked to, actually have some internships at AMD. If you're interning with AMD, talk to others. Ask them what inspires them. And you're going to hear something different than you hear in some of the other companies. You're inspired to make that difference in society and to make lives better, make people more productive, and to speed science and speed direct discovery.

And you know, we look at, I talked to you about the supercomputers that we're powering today, that type of collaboration that we have with the US government and academia, gain their trust. They saw running on open systems, running that collaboration, they could move science forward. I've talked to, you know, so many with examples of Pratt & Whitney used the supercomputer at Oak Ridge National Lab to run high-performance analysis of their jet engines and to find efficiencies. Think about the fuel that they're saving on the new airplane engines that they designed using that supercomputer. It was double-digit percentage. And when you just think about what that means in terms of fuel on an annual basis, it's huge. And that was access to the current level supercomputing. So what I show here is with the US Department of Science, Dr. Dario Gil, Under Secretary for Science under Chris Wright, the Secretary of Energy, have launched the Genesis mission. We're standing up right now as part of Genesis mission a cluster at Oak Ridge National Lab of MI355s. It'll be stood up by the end of Q2, so about another three months. And it's a massive amount of our current generation MI355, again supporting both AI approximations but also that analysis like Pratt & Whitney did where you just have to have, I mean, they're doing fluid flow analysis that has to have high precision, and we have both in our computers, and it's being used to speed those kind of discoveries. And then our next generation, I showed you that Helios rack, we have a variant of that that again will support both HPC and AI approximations called the MI430. And that has been awarded the Discovery supercomputer. So that will be the follow-on to the Frontier system and is targeted again to really drive that next generation of innovations.

What that means is things like fusion that have just been forever just a dream. And now you're going to start to see timetables of when it will really start to bear fruits because we're going to be able to model, rather than, you know, if you go to Lawrence Livermore National Lab and you'll see every few days they can fire one shot in, and the NIF is the National Ignition Facility. I've been there a couple times, it's incredibly laborious and timetabled to step their way to advances in fusion. But with these type of supercomputing capabilities, it's going to be a digital twin. I'll talk about that more in just a minute. But I want to talk about another accelerator I'm really excited about. And that's quantum. Our efforts at AMD are how do you build quantum, which is an amazing accelerator because think about binary and all of our classical computing to the massive states between a zero and a one that you have with quantum computing. So it is problems that are multi-state, statistical in nature can be solved that could not be simulated and solved before. And we are now within a few years of getting sufficient control and error correction to have real-world problems solved by quantum computing. And again, what we're doing was led by our research team. We have Angela here from our research team with us here today at A&M. We have our FPGAs, our field programmable devices, are at the heart of the control and also the error correction. So we're working with really every, almost every quantum startup and major company out there delivering these new capabilities. And equally, we're architecting with them how does that quantum accelerator work with classical computing. Quantum is going to be very expensive. And so what you need to do is solve your problems on whatever you can on classical CPU, GPU, classical accelerators. But those aspects of the problem that need quantum, you need to be able to send those to the quantum accelerator and then amalgamate the results back. So very, very exciting time.

I'll shift gears a little bit and talk about PCs and the edge. Also exciting time, and particularly for those of you in students because this is what can bring AI right into your hands. It's the imminently affordable AI. And what we've done at AMD is we've used the same architecture of our CPU and GPU. So our software stack, our tools, our enablement come right over. We're actually, again, in our research team have innovated some tools that make it very, very easy to optimize. And again, we open-sourced it so you can just go out to a web page and find it, or it's promoted at GitHub. And so everything from laptops to what we show is Strix Halo, where, you know, Nvidia just announced a PC that they're coming out with that's going to compare, as we have our Strix Halo has a leadership position over what they'll be coming out with this year. So we're already out there combining incredibly powerful CPUs, GPUs, and 128 gigabytes of coherent shared RAM around that to make it a very, very productive AI machine. It's not as inexpensive as a $500 laptop, but it's imminently more affordable than trying to go get on the cloud and run your tokens at AWS or another cloud provider. It's very, very economic. But again, this is what's critical is it's such a disruptive force. It's a motivation at AMD to democratize, to get this out to more people's hands. It's not good for society if the access to the computing for AI is limited to the very few. And so, you know, it's a major, major focus that we'll have.

But those of you who are gamers should feel happy to know that we're R&D and applying AI everywhere, and that includes gaming. So you think about how we used to do the upscaling. It was all physics, all math. We had math models doing the ray tracing. So every time you saw light bouncing as you were playing games, it was running those physics engines. But now, you know, just like AI for science is mixing AI with high precision, we're doing the same in gaming. So if you look at the latest release of our FidelityFX Super Resolution, it's doing the same. It's got a very, very close mix of AI with the math modeling, and it's just amazing realism that you see in the gaming. You'll see the same kind of technology in upcoming game consoles. So our partners are Sony and Microsoft. I can't speak for them on all the features of new game consoles when we come out. They've already hinted at it, so I can hint at it as well, but I can't share those details. But stay tuned. Very, very exciting advancements in gaming, and I imagine there might be one or two gamers in the audience here.

And then lastly, I talked to you about embedded AI, and when I say AI is going everywhere, this is where AI goes into literally every device around us. It's already starting in a lot of very, very small devices. We have lightweight inference engines. I'm always an adopter. So if you see me walking around my home, it's all voice command to those inference engines that are embedded in everything from my thermostat to my microphone, my speakers if I want to change the song. But what's coming is a much more rich AI than natural language recognition that we mostly have today. It truly is going to be the AI that you're seeing today with billion-parameter-plus models that's going to come locally, and it will be able to be fine-tuned for tasks. And so now you start thinking about reimagining how these embedded solutions can be not just smart, we think of smart today as, oh, I can talk to it and it can talk back. That's not smart. Smart is when it can really have the kind of capability that you're getting today when you go to that favorite generative AI model that you're using, and that's coming. And imagine when you start tailing those models to specific tasks. So it's in the factory floor, that's a very much optimized, very small model when you don't have to think about answering any question in the world, but how to optimize those factory operations. So the power goes way down. Again, it's part of that energy wall that I told you about. Part of that's going to be not having AI be the broad application in every case, but really tailoring it for tasks at hand. And we're just going to see it go across every industry. Again, it's already happening.

And it's leading up to the era of physical AI. When you think about physical AI, it means that it's leading to what will be the ultimate use case, and that's a humanoid robot that really mimics what we can do with our human bodies. And it used to just be dreams. We've all been watching the movies showing humanoid robotics for decades here. It's now imminent. I was in the Bay Area on Monday of this week with a leading company in the area of humanoid robotics, and it's astounding the progress being made. What are they doing? They're investing massively in modeling, the accuracy of the modeling, and it's AI for science. It's a mix of high precision, low precision, and they've been perfecting the models. They're using, you know, combinations of devices, but they're using all of that modeling, and they're creating a world model. That's why we show this vertical line there. You see in the slide says we're entering a new frontier. So, we've been optimizing specific parts of robotics and these embedded devices, but the world model puts it together. It allows you to have a digital twin that mimics being in the real world. So you can simulate a humanoid robotic in this incredibly complex world that we just take for granted. We take for granted the massive sensing capabilities that we have of touch, of vision, of hearing, how that sensor fusion puts it all together, the reasoning of the brain, the joint activity that we do. Think about the computation our bodies are doing of how to balance as we walk. I mean, it's just massive computation our bodies are tuning in at every moment of every day. And that's what has to be tackled for human robotics to take off. Again, it's got to be incredibly power efficient. So, it's going to have to be very, very finely tuned models. And it's going to have to be designed in a world of digital twins with world models.

But it's coming, and you can hear projections of anywhere from, you know, years to within, let's call it 15 years, but it's in that range. And I could have said that until just very, very recently. I couldn't have told you that it's in the definitive range that we're seeing this coming. And again, this is part of the same cusp that I talked about because what are those humanoid robots? What's going to enable their intelligence, the brain? It's going to be basically agentic processes that are starting now, will be much more evolved by the time you see humanoid robotics. So these trends are all connected. And you know, as you think about the brain activity, the joint activity, the spine activity that we have, it has to be all modeled and working closely together. The other aspect I'll tell you is it has to be real-time. So not only incredibly energy efficient, it has to be virtually real-time, and it has to be deterministic. You can't design this such that there's arbitrary outcomes that may occur. Do you want that robot in your home that has arbitrary outcomes based on the task that it's facing? It's not feasible. So it's really, I think, the most demanding use case of AI, and some of the best brains in the world are on it. And we're very proud at AMD to be partnering with them and providing devices across CPU, GPU, these FPGAs, and helping them speed their discovery.

So I'm remiss if I don't talk about the equal and actually the majority partner in the hardware device development we do. We're not a hardware company alone. We haven't been for years. It's all hardware and software and how it works together. So we are very, very actively involved in our software stack. But we do it in an open fashion. So this is our software stack. ROCm, it's the open compute model that we have out there. And you shouldn't be surprised we support all of the frameworks that are out there. We are a GPU, just like Nvidia, we've been competing with Nvidia for well over two decades on GPUs, and so the constructs are very much the same. So if you're on Nvidia and you want to port over to AMD, we port over very, very quickly. It typically takes about a day to use our heterogeneous portability tool and to map it over. And then it used to take about a week or two just to tune it up and get it to full performance. But AI's changed all that because now we have an agentic process that does that tuning. And so we can port over and tune almost instantaneously. So again, AI affects everything and is drastically affecting software. Now you know all the stories about coding, but it's also making it vastly easier to move across different platforms.

So look, I want to wrap up. Let me see. I think we're okay on time. Dr. Ready, one minute. Perfect. That's what I need. One minute to wrap up. Because I just want to really highlight what my suggestion and ask would be to you. It's just simply to embrace AI. Please embrace AI. Please think about it. Am I getting up to like 10x more productive? Am I embracing it every day in what I do in my studies and my research across every facet? Because that's what you will need to be competitive. I just need to be candid with you. This is going to be essential going forward. Those of you going to industry, you're going to interview and you're going to be asked, tell me about how you apply AI in your studies, how you apply it in your daily life. So please don't miss out. Please embrace, supercharge your activities. Because you are in the most disruptive time of technology. If you embrace these tools and make yourself more productive, I'm just jealous. I wish I was starting my career right now. For those of you students, I wish I was in your shoes. And look, I just want to end with a huge thank you. We are so appreciative of the collaborations we have at Texas A&M. It's across many of your design centers. It's with your faculty. It's the interns and the extensive hiring that we do here. Thank you very much, and we look forward to deepening those collaborations.

Really appreciate your time.

It's a great question, Nicholas. And your example is exactly right. Look at those purpose-built computing that we have in homes. Look at what you see. Purpose-built computers that are, you know, in the factory floor. I visited Diligent Robotics in Austin a few months back, and Diligent has robotics that you now start, they're already deployed in a number of hospitals. They're purpose-built. They're not humanoid. And so the question is, isn't that the most efficient? And it is, and will continue to be for purpose-built activities. The reason of humanoid robotics is to ultimately get this kind of scale and impact that we as a society will be demanding before we'll be truly acceptive of a physical AI embodiment, is to accept it into our human world. So the reason for humanoid robotics is simply human acceptance. We won't accept the, you know, proliferation of these for broad applications, and people won't want a dozen devices purpose-built in their home. And frankly, we built a world over centuries that's tailored to our human form, and that's what's driving humanoid robotics is that it drops the barrier of the use cases. So it moves it from purpose-built to general use in a human world. That's just, you just can't change how we've tailored this world over generations. So that's what's driving it. But I'm sure there's going to be a continuum of solutions from the very, very defined purpose-built to an ultimate humanoid robotic that we're going to see. Engineers, like we most of us all in the room, are incredibly creative and adaptable. Well, let's see. But that's the driving force behind humanoid robotics. I'll let you, Dr. Ready, choose the questions.