A (3:24)

All right, well, first, let's go into the categories. Okay, so A is architecture. And so architects use Soffler to do conceptual design and urban planning. And the architectural design, which is called bim, which is the latest kind of industry standard for the way you do architectural design and construction documentation. And then the engineering part, it's not mechanical engineering like in factories, although buildings are starting to get built in factories. It's called prefabrication. And those processes are more like traditional manufacturing processes. But the E traditionally stands for civil engineering. And what's called mep, which is mechanical in terms of H vac, like the heating and cooling systems, electrical, the electrical, which supports all those things, and plumbing, which is an integral part of any building. And also data centers depend on water for cooling these days. So. And then there's construction, which is as it sounds. And now there's specialized software for each one of the A, the E, and the C. And there's also software, specialized software for the. Oh, for operations. The software vendors are the ones everybody's probably heard of, like Bentley, Trimble, Otter Desk, Hexagon, Nemechek, and then there's special rendering software companies like chaos Group and D5 and Unreal Engine, Epic Games, et cetera.

B (4:57)

Okay, makes sense. So, like, in any industry, there's, you know, tons of different softwares. I mean, obviously they're all accelerated by Nvidia. I think a good question to kind of stop and ask is, you know, in terms of. And we don't have to get into specific models of Dell Pro Max. But more on the Nvidia side are these kind of lower end on the the GPU stack. Are we talking 1,500 Blackwell RTX Pros? Are we talking 6,000 Blackwell RTX Pros? Like, where does that kind of fit in the GPU stack for acceleration for those ISVs?

A (5:30)

Well, there's a lot of. It depends. At the most basic level, 3D software of any type, CAD or BIM, uses the GPU to power the viewport with DirectX or OpenGL, and that depending on the complexity of the scene and the size of the display in terms of resolution, et cetera, you need a powerful GPU for that. But that's really kind of like table stakes. Every software application needs that. But that's not the best way to measure the capacity of the GPU that you need, because many of these software applications do compute on the gpu. People think about the CPU as the computer as the, as the place where computing gets done. But more and more, a lot of the intensive, computationally intensive processes are being offloaded to the gpu. Why? Because the GPU can accelerate those processes much better than the cpu. And the reason is because the GPU has thousands of cores where the CPU only has a small number of cores. And those operations that can be parallelized, broken down into many small parts perform much better on the gpu. And those operations can vary depending on the software application. Some software applications offload compute to the GPU for file I o when you open or save a file file. Some do some type of simulation that's mathematically intensive and that can include like computational fluid dynamics, structural, et cetera. Some applications convert point cloud to mesh using the GPU or align photogrammetry using the gpu. And of course there's ray tracing. All the ray tracing applications benefit from the RTX stack that's included in the hardware of our gpu, but also in the software that we provide to software developers. Speaking of software developers, any of the tasks that I mentioned previously that benefit from the GPU benefit from a software layer that Nvidia has called cuda, which is available to every software developer to optimize and accelerate their software for our GPUs, and I haven't even got into AI, and of course AI is going to run much better on the GPU compared to other methods like the CPU or the MPU for the type of work that folks in the AC industry are doing.

B (7:54)

So, I mean, you kind of teed up the next question perfectly, Sean. I mean, we all know that AI is kind of taking center stage. I mean, AI has been around a lot longer than people actually think and goes back to many years to machine learning, neural networks. But in terms of commonplace, probably in the last three to four years it's really taken off. So with AI kind of taking center stage now, not only from an app development, software development, AI dev data science standpoint, but specifically with an AEC and geospatial, how is AI going to disrupt AEC and geospatial customers workflows like how are they changing and evolving with AI?

A (8:29)

AI has the ability to learn so it could watch the way we work and learn how we do it and why we're doing it and, and automate those things that can and should be automated, eliminating a lot of the mundane tediousness of the day to day workflows. So there's a ton of examples in AEC in geospatial and now we're starting to see, which is something I've never seen before and is this incredible propagation of AEC AI startup companies, because they recognize that across the workload, the A, the E, the C and the O, there are thousands of different operations that can be optimized with AI. And you know what's interesting? We have all this software and we've gone in aec, we've gone from drawing blueprints on stone tablets to blueprints on paper to CAD to bim, et cetera. But in a nutshell, the productivity in this industry has not improved despite this focus on technology. In other words, we haven't really leveraged the technology. And I believe from a productivity perspective, and I believe the reason is because it's still a very manual process. We're still drawing the geometry. Even if we can generate the geometry computationally, there's still a ton of manual steps involved and manual steps involved in each conversion process, going from conceptual to BIM to construction documents, et cetera, et cetera, down the line. And IT create all, all these different steps and processes create tons of confusion by the way, when it's time to build things on the construction site. So AI has the ability to optimize all of those things and automate many things that are just like hair pulling, repetitive, silly wastes of time that really eliminate the ability for folks, for practitioners to focus on iterating, applying their ingenuity to creativity, thinking about solutions and problems early in the design process because they're all too busy just doing this manual work. I'm really excited about the future. And all of this by the way, takes computational Capability. There's two sides of AI. There's the AI work that all of these startups are doing, which is AI development and software development, and the AI that all the end users depend on to do their work, which is AI inference. And we could discuss like from a computational workstation requirements perspective how they're very different.

B (11:26)

Yeah, let's, let's dive into that. Well, let's talk about like the traditional kind of end user kind of inference workflow. Right. You kind of described you know, a process where, and it probably is on both sides between the startup and the traditional ISV where hey, we can get mind numbing tasks done that take forever, very human intensive. I mean it's kind of an agentic process, right? We're with certain guardrails it can go kind of through. And I mean, and an example, this is a very old example, but it'll kind of illustrate the point is I have a coworker who, you know, drafting designer, etc. Who back in the day, you know, had to go in and manually tag kind of the doors in the, and the drawing and the design, right. And then came a feature that said tag all doors. So it made much easier to go kind of go through and do that. But diving into kind of the end user ISV related, you know, from a computational GPU standpoint like Sean, dive into some of those details on what those end users are actually kind of maybe doing and then specifically like the computational requirements.

A (12:26)

This is a great question and I like the example. Let's break it down into a couple things. So the first thing is computational work versus AI inference, which is prediction. So in, and by the way, my son is an intern in civil engineering firm and I see like they have them doing very menial and quite tedious jobs like the one you're describing. Very similar actually that can be completely automated with AI. So I tell my kid like, you better learn AI, you know, because these things are going to be eliminated. Like it's fine for an internship, but you don't want to be in that position when you're in your career. So traditionally like software applications have to plan for every function that they deliver, including like if you wanted to have a tagging feature, you would have to write an algorithm and figure out how you're going to identify these objects computationally, which requires a lot of computer thinking and processing and how you're going to tag them computationally, which requires a lot of thinking and processing. And by the way, if there's, if that can be parallelized, it can be done very well on the gpu, but it's still following an algorithm which, you know, might have to be updated and rewritten by a software developer as things evolve. And that couldn't possibly predict every variation or permutation that's to come, which is very different than AI because AI has the ability to learn and also to keep relearning as it works. It can receive feedback from the end user and be updated. So that act of the AI working is called AI inference. So when you type a prompt into ChatGPT as an example, you're getting AI inference. And basically the ChatGPT is trying to use tokens, which is the, the methodology of doing inference and compute terms to predict your next word just the same way. In this example that you gave me, the AI would number one, identify all the doors and number two, and using a prediction mechanism like the same way it could identify like fire hydrants. And why, how does it know to identify fire hydrants? Because you teach it every time you do a captcha. And that's the kind of like, you know, reinforcement learning that I'm talking about. The AI can improve and then it has to use some sort of prediction for the tagging. What's the right way to tag these things? And that might be based on a rag looking up their tagging methodology and the last tag in some sort of a vector database. And I'm sure you've talked about all these things with SOMA and other people on my team. So all of that inference work can be done on, can be done very, very quickly in real time, which is the advantage over the traditional computational methodology. And the other advantage is it can evolve and iterate and teach itself. And you could teach it without having to go back and have software developers rewrite the code. You know, we say that AI write software and here's a good example, but the interesting phenomenon here is the level of compute may in certain cases be lower than the computational case, particularly when things are massively computationally intensive and required to go out to like an HPC high performance compute queue where you require like roomfuls of servers to do the compute. A lot of people don't have that capability. AI inference offers the same, you know, functionality by using prediction instead of computation can reduce the compute envelope to your workstation, which is really amazing and is going to completely transform the way Dell Promax users get work done and what they can do in a workstation.

B (16:11)

Yeah, no, I mean, absolutely. And I'll give this example, my wife will probably hate it, but you know, in the kind of the prediction, the reinforcement learn. And we just, I don't know how we made it, but we just redid our, the first floor of our house. My wife's been on me for 13 years about it and we finally ripped the band aid off. You know, we redid the bathroom and all that. But in that example, right, from a reinforcement learning is we had a very standard door, like normal Home Depot door, right hand swing, like very easy. So from an AI kind of. And we're not talking the inference piece, we are talking kind of the training piece. But it's very easy to train that, hey, that's a door. And you know, there's lights around for the outside of the door. It's very easy to train that. But my wife, lover, but she wanted this big custom fancy door, right, like where we had to tear into the walls and all this kind of stuff. But that is a door. But that is not traditionally like a size of a door. Like we had to have it custom made and all this kind of nonsense. But in that example of reinforcement learning, you know, Sean's son, who is, you know, doing his internship this summer, sees this architectural drawing and he's like, oh, I recognize that's kind of a weird shaped door, but that's a door. Let me tag it. So then the AI learns and then from that you've built up this repository of what does a door look like? So AI then will inference locally on the GPU to say, hey, that's a little bit weird of a door. I've never really seen it, but hey, I have an example of that. That's a door. That's a door. That's door. So that's what kind of the inference piece means. And you don't need a ton of GPU compute, but there are things that you will. And we'll get into that in a second. But let's talk about the next thing. Sean, so we talked about, you said there's a whole host of kind of AEC startups that are focused around AI. Maybe let's start with what kind of problems are they trying to tackle from the ones that you've had a chance.

A (17:53)

To talk to every possible nook and cranny. I like to think of it like the example that I like to use is if you come home from vacation and you come home to a very dry potted plant and you pour water on it, the water is going to go everywhere. It's the same thing with AI. AI is going everywhere and it's filling in the gaps and in like traditional ISV stacks, automating things that are today manual processes. There's startup companies that are working with the traditional ISV applications to help automate certain things. A simple example of that is every software application has an API that allows you to do like custom programming to create features that aren't included with the software. But it's very complicated. It takes skill and knowledge to understand how to use that API. Well, ChatGPT understands languages. One such language could be the APIs of your software application. So startups are creating like chat. With my software application, you could say, hey, draw me a really cool custom door that kind of looks like this picture upload picture. And it'll use the application APIs to do that. But I want to get to one really specific example or area that I love that really gets into the power of Magentic is on the construction side, agentic AI really means taking the human out of the loop. And the ultimate manifestation of that is when you have multiple AIs working together without multiple humans. Each AI is focused on the thing that it's really good at. And a great example of that, where like humans would cause problems and slow things down, is in like a safety or security environment on the construction site. And of course, for general contractors, safety is priority number one. You're the humans that are working there are above all else the important thing. And you want to keep them safe. And you could use agentic AI for that. And the way that's happening, and the startups that are doing this basically have cameras all over the construction site that are monitoring everything that's going on. Now if you could imagine, you have like 30, 40 cameras. Are you going to have 30, 40 humans watching each camera like aircraft controllers? No, not at all. You have AIs monitoring the cameras. The AIs have been trained for specific things that happen on a construction site, like accidents and people walking backwards and things swinging into people. One example that I like to use is a crane. So the crane is carrying a load and swinging this way and the construction worker is like walking backwards this way while they're taking photographs or something like that. And if the trajectory of both continues, there's going to be a collision and the construction worker is going to get hurt. The AI, which understands physics and motion and trajectories, can pick these things up. And then agentically, without a human having to sit there or fall asleep or take a lunch break or miss the thing entirely, the AI picks up the that, hey, in a minute or 30 seconds, there's going to be a problem. Now, what are the steps I identically need to take to solve the problem? One would be call the construction worker's mobile phone, call the crane operator's mobile phone, flash something on the crane operator's screen, sound the alarm, call 911. Emergency shutdown of the crane. There's multiple steps that the AI can take completely autonomously, each one of which is like, if you think about it, a sub AI, an AI that calls 91 the right phone number. An AI that calls 911, an AI that shuts down the crane. So all these agents are having a conversation, you know, and escalating appropriately. Another great example of the construction site is using the same kind of cameras, plus other photographs from workers, plus reality capture data from photogrammetry, lidar, et cetera, to compare the progress of the AS built back to the construction documents and the bim and check for clashes and check for. To make sure quality is correct and check to make sure everything is the way it is supposed to be and check for progress against the construction schedule. You know, this was supposed to be completed. Let's. What's the level of completion? AI. Independent agentic. AIs, AI agents that are. One AI agent understands the schedule. One AI agent can understand what it's looking at and translate that to the CAT or the bim. Another AI agent could make projections on the scheduling and another agent can adjust the cost. Another agent can send alerts to all the subcontractors that there's delays and they should manage accordingly. Think about all the waste, mismanagement and human error that that kind of agentic process eliminates from the system.

B (22:42)

Yeah, I mean that, that's a fantastic example. And you know, one. And for those that fancy themselves kind of learners and someone who wants to educate themselves, go to Nvidia.com and I believe it's backslash, it's blueprint or build.Nvidia.com and look for the blueprint called LM router. And basically LM router is very. It's a. It's a very basic version of what Sean's describing. But it'll give you the ideas that I think people have the, I would say misconception that AI can do everything. And it's not AI can't do everything. AI has to be trained how to do certain tasks. It's kind of like, for example with my daughter is she's not very good at cleaning her room. We had to show her how to clean her room. Now she's Good at it, but she had to be trained on that. But just because she can clean her room doesn't necessarily mean that she knows how to go cook with a sharp knife and make spaghetti. Right? That's an independent task. So kind of what everything Sean's describing is, it has to be trained on each kind of those individual steps. And LM router is a boiled down version that can run on a Dell Pro Max with Nvidia RTX gpu. But we can show you that exact task where it's, where it's image, text, logical reasoning. It'll route that task to the right kind of, or route the inference to the right model, whether it's vision or whatever, and give you the right result, which is basically an agentic or multimodal system. And you know, it's funny that you mentioned kind of the extraction safety example. My dad, I won't name the name of the company, but he's retired now, but he still knows a lot of guys. And recently he was telling me about that it's a, it's, it's not construction. Let's, let's just say anyways, it's it. They have to wear hard hats and they get in big trouble when they don't wear their hard hats because it's, you got to wear a hard hat. And the company, you know, created kind of exactly described it. Not quite multimodal, agentic, but we'll look and we'll find examples with computer vision of people not wearing their hard hats. And then they will get a warning kind of in their union file. Right. If they're not wearing their hard hat. But that is just a very interesting way of how instead of having someone, a safety officer walk around, you can use AI to ensure kind of the safety of your workforce. So I, I love that example. So you brought up a good question, Sean. I want to ask because I've seen it kind of diverge in two ways in the market. And I'm curious how AEC kind of which angle it's taking. Right. Is that, you know, with kind of the onslaught of AI, I see a lot of companies across many different verticals kind of create their own software. Right. And some are going to be very successful. But at the end of the day, customers don't necessarily want to add another software, another application. What I've seen, especially in M and E, is the integration of some of these processes and tools and IP into, for example, you know, one of my partners, Blue Moon AI, like the, the adoption or the integration of their application within side of creative cloud like within aec, have you seen more of the standalone creation of new, you know, app AI applications and workflows or are you starting to see more of a convergence of those into via maybe an API into the traditional ISVs?

A (25:47)

I see three things. I see one, as I previously described, tons of startups doing every possible thing you can imagine and it's great. I mean you can't ask for something better in terms of using technology to improve productivity because you've got this entrepreneurial spirit from people who are in the industry who have experienced frustrations on a day to day basis. Now see AI as a tool that can be harnessed to solve these problems. They move very, very quickly because they have one specific problem to solve and they don't, unlike ISVs have legacy code bases that they need to manage. The traditional ISV applications have lots of users and millions of lines of code and a schedule of things they need to fix and improve and new features they need to add that. They've been promising customers for years and it's very difficult for them just to like, you know, throw something. They can't just throw things in. Very difficult for them to experiment. So what's likely going to happen is you're going to see some of these startups get acquired by the traditional ISVs. You're going to see innovation from the ISVs and inclusion of new AI based functionality in their traditional software applications and you're going to see ISVs creating and launching new software applications and services that leverage AI. But the other variable here that we're seeing is the companies themselves, the customers now developing their own AI, develop AI software development and AI development capabilities. And this is really interesting and never really seen, we've seen a little bit with computational design like Grasshopper scripting in Rhino and that kind of thing, but that was, you know, very localized and limited to a number of firms. But here we're seeing a lot of firms for tons of different use cases including things that have nothing to do with the design process. Just like reading and responding to RFPs or managing HR, doing costing and scheduling, like Pure, like LLM based things that have nothing to do with design. We're seeing AI use cases across the board, across all the firms in acno, aecno. And so this, the other really important element about the reasons why companies want to build their own AI is because AI training needs, AI needs data to train on and all of that data resides on their hard drives and most of that data is what they would consider proprietary. They don't want to train on somebody else's data. They don't want to train on generic data because that data doesn't reflect the way they do things, the way they do their scheduling, the way they do their costing, the way they do their design, the way they manage their construction, whatever it is that they want to. They have a hundred years of experience doing things a special way that's got them to this point. What if they can train an AI based on that data to automate certain elements of their workload? They need to understand AI and AI software development in order to do that. And then the question is, well, how do you do that and integrate that capability with the software coming from these startups and from the ISVs. For example, you get some models from the ISVs or startups and you want to do what's called fine tuning with your own data. You have to either have a process that's defined by the software vendor or you develop your own process. I see a lot of both of those things happening. For companies that want to get into AI software development or fine tuning, note that the computational requirements are quite different than inference, just like the people are quite different inference. You have architects, engineers, VDC managers, et cetera, using traditional software that's powered by AI and that's called inference. Using traditional workstations with GPUs that are going to be doing more things with AI, you need better GPUs, you need better everything, frankly. Better GPU, better CPU, more memory, more storage because more is going to be happening on the workstation. But it's more of the same type of architecture that you have today with rtx, et cetera. Over on the AI development side, it's completely different and you've probably talked about this with Sama, that you now need specialized hardware and lots of it to do training. Like for example, the Dell Pro Max with GB10, which is Grace Blackwell supercomputer that is available in a very small form factor, but scales up all the way to data center with lots of GPUs for doing complex AI training. And this is the path that our companies are headed into, our AAC companies. But there's these two paths, you know, and they don't share the same workstations architecture. And our customers need to understand this.

B (30:42)

Yeah, no, I mean that's a great point. I mean think of so inference training, fine tuning, right? Like I kind of think we have inference, you have fine tuning, you have full on skill training like inference. Think about consuming, right? Whether it's a chat GPT or automated tagging all the doors. You're not writing the model, you're just consuming the, the intelligence out of the model. Right. Then you have fine tuning where you're taking kind of a model and you're adding kind of your own data. That can absolutely happen on a workstation. A bit, a little bit slow. Right. But I mean you could run that on a 96 gig Blackwell RTX Pro.

A (31:15)

On or on four of them.

B (31:17)

Or on four of them if you want to run it faster. Exactly. But if you're doing. Exactly. And then what you said, when you're full on training, you're talking data center, very, very heavy compute. So we've talked about aec, we've talked about geospatial. Let's kind of take just a bit of a different twist. So if you haven't heard, I wish, I think most have heard the term kind of digital twin, you know, something that's through Omniverse, which was recreated kind of a replicative world in 3D of an actual physical space. How does digital twin fit into AECO or Omniverse?

A (31:49)

Currently, yeah. Digital twins are required to have an understanding of what's going to happen in our universe in the physical world. But to have that understanding in a digital form before any concrete is poured. And the types of digital twins that Nvidia is getting involved with account for every variable. All those variables are grounded in the laws of physics. So what you see, like going back to sketching, orthographic and perspective sketching, we're trying to show what this thing is going to look like, look like being the key term in the physical world. But that doesn't do justice to the way the thing is actually going to work in the physical world because it doesn't factor physics into the equation, it doesn't factor in the way the environment is going to interact with the built thing, whether that's water or wind or vibrations from the earth, like earthquakes. And the way people are going to use the building or the, the, the asset, the way it's going to degrade over time and how to manage maintenance, et cetera, to manage that asset, the different ways that it can be used, like if you're in a hospital, there's optimal design scenarios that you need to factor in the full physics of the deployment, the way that the building or the asset is going to be used with everything, people, processes, electricity, lighting, everything factored in. And then you want to design to an optimal outcome. Normally what happens is the design process is you design something on paper, the Architect or whatever designs it, sends it over to the engineer, the engineer runs some standard calculations on it, like just what they need to do to get it signed off and then they send it over to and oh, probably the engineer says, you need to change this thing. It can. That thing, you know, won't work. It's those, those trusses are going to support the load. So go back and redesign. Then you resimulate and then over to the construction team, have to figure out how to build. Well, what if you kind of, you know, had all that information in one place where like here's all the design, the engineering, the construction data. Let's start simulating different scenarios, let's factor in absolutely everything and get the math results. In terms of how efficient our design is, does it return the KPIs we're looking for in terms of energy consumption, in terms of lighting in certain places, in terms of the way people use the space, in terms of if it's processes like oil and gas, in terms of how efficient it is in terms of production of that product. All of this can be done in the digital twin and it's not today. So there's tons of opportunity ahead of us.

B (34:51)

Yeah, I mean if you watch. Jensen's keynote didn't speak necessarily directly, I mean definitely about digital twins, but really physical AI, right? It's like how does the physical work? Like you said, how does it introduce itself into like the AI and the 3D aspect of it, right, whether mapping a street for 3D car to know how people walk out of the street, to know where the car should drive, all of that type of stuff. So it's definitely on the front wave. It's very, very exciting. The stuff I've seen is super cool. So I know we're getting up close to the end. Sean, couple of rapid fire questions for you that I think the audience will definitely want to know is, you know, what is a future looking use case in AEC that you think is on the horizon in the next two years?

A (35:31)

Well, I think what's happening now is first of all there's a recognition of this value of the digital twin, especially for large scale assets like airports and stadiums, entire cities and how the things that I talked about in terms of the value of having all this information digitally before any concrete is poured, but also because AI is now going to be part of the building operations solutions, you need, as you say to have a digital twin to train the AI because an AI needs training data. If you're building a building that doesn't exist how are you going to train an AI? Nothing exists to train it, so you have to create this synthetic version, this digital version of the building. It has to behave as it will in the real world with physics, and then you could train an AI based on that data. So we're starting to see this convergence of AI and digital twins. And a great manifestation of this that I love is the idea of robots on the construction site, because robots are basically AI robots. I mean, to have an efficient robot, you want the robot to be autonomous, but it has to know what it's doing. And to know what it's doing, it needs to be trained. And to be trained, it needs training data. And that training data is going to come from some sort of construction digital twin, ideally the digital twin of the actual thing that the robot is building, the specific bridge, the specific building, so it can anticipate every and go through and plan in a digital world. Everything that needs to be done when it goes to the site, it doesn't make a mistake.

B (37:03)

I love that. So, you know, Sean, this has been a great conversation, like learning about kind of AI from, you know, the traditional kind of AAC workflows to the introduction of AI and how they're kind of morphing the industry and also talking about digital twins. So, you know, as we wrap up, give the audience, you know, one or two minutes just to your kind of final thoughts and takeaways and then we'll go ahead and wrap up the episode.

A (37:25)

Yeah, well, one final thought that we didn't talk about is coming up of next week, Dell and Nvidia are going to be together at the S3 user conference, which is Esri's, the largest geospatial software company. And we're going to be talking about how AI is powering future workloads today inside of S3 ArcGIS based on Dell Promax with multiple GPUs. And this is doing exactly what we've been talking about this whole session in the context of Geospatial. You can imagine how many satellites or flow applying today thanks to like SpaceX ability to inexpensively launch satellites. These satellites are monitoring our earth for all sorts of things. There's tons of data coming down. What human is scanning and analyzing all of this data? It's far too much. So ArcGIS Pro from S3 makes it possible to use AI to get insights from this data, which is ultimately what you want. Show me the erosion, show find me this lost hiker, whatever it is. And all of that can be done using AI on Dell Pro. Max with Nvidia GPUs and we'll be out there next week talking about how to do it.

B (38:35)

Amazing. Well, thank you Sean. Really appreciate the time. So if you haven't, definitely check out the Reshaping workflows page on Dell.com and definitely check out LM Router on Nvidia or build.Nvidia.com so with that being said, this is Logan with Reshaping workflows powered by Dell Pro Max and Nvidia RTX GPUs. Until next time, keep your AI running locally on Dell Pro Max and Nvidia RTX GPUs. With that, have a great night.

A (39:20)

Do what you want.

B (39:22)

This podcast was produced in partnership with Amaze Media Labs.