NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hi, everyone. Thanks for joining us today for Demystifying Edge AI From Proof of Concept to Production with NVIDIA and Lenovo. Before we begin, we wanted to cover a few quick housekeeping items. At the bottom of your screen, you can find various widgets. Once you open them on the screen, they are resizable and movable. If you have questions during the webcast today, including any technical issues, you can put them through the Q&A window. A copy of today's slide deck and additional materials are available in the resource list. Now without further ado, we'll turn it over to Amanda to begin the presentation.

Thank you so much, Lana. So here we are today. Thank you for joining us for Demystifying Edge AI from Proof of Concept to Production. We have hundreds of people joining us, and thank you for taking the time. For this webinar, I'm joined by Blake Kerrigan from Lenovo, where he's the General Manager of their ThinkEdge Business Group. So glad to have you here, Blake.

Yes. Thanks, Amanda, for having me. Super excited to be here.

I think we've got a lot of great stuff to talk about. So let's just jump right in. I wanted to jump off by looking at this overall trend of digital transformation that's the top focus of almost every industry these days. Businesses are looking at how they can use new technology and data to drive business agility and flexibility so that they can improve efficiency, improve that customer experience and reduce their overall cost. Sometimes we call this the Fourth Industrial Revolution. A new wave of technologies like AI, 5G, big data, digital twin, simulation, IoT and robots, all of these coming together enabling companies to move their business to this sort of software-defined practice. Now of course, many of these new technologies are taking place outside of the data center. They're happening actually where the data is being generated, which is why we see this huge upsurge of folks who are interested in edge computing and AI. Let's talk about the path. Today, most organizations are there collecting data. They've got sensors. Cameras are probably most -- the most common sensor that we see out there. But there are trillions of IoT devices measuring traffic, temperature, signals and so much more. And companies start by collection, data lakes, making sure that they have this data. But the move that we're seeing today is a lot of companies investing in how they can use this data to produce some sort of intelligence through AI inference. While a lot of the training is taking place in the cloud using these massive data lakes, the actual inference itself is moving to the edge. That's where AI can help customers detect anomalies or segment an action. And the final step, once you get that intelligence, is about integrating the detection into some sort of action. This could be as simple as alerting a human. But it can get more complicated as we look at intelligent machines and robotics. But that's the ultimate goal, is to go from collecting data to making intelligent signals from this data to actually automating our processes based on this data. So Blake, I know you're working with a lot of customers today. How do you see them going through these phases? What stages are they kind of at?

I think it's a great question, Amanda. One of the things that I think we've observed through kind of the last 10 years of this massive kind of what we call traditional IoT practices that I think a lot of digital transformation, 4.0 leaders and large enterprise have really just been focused on how much data can we collect and how much analysis can we do. And that story always kind of begins with the data capture with small devices and moving data into the cloud, doing some analysis and then coming back in and doing some sort of changing of some sort of a process. Maybe it's a people process. In other cases, it might be a machine process. And what we're starting to realize now is that a lot of customers, Lenovo being a large infrastructure company, of course, we love to sell those customers storage for all that data or hyperconverged instances to do that analysis. But the reality is that with that proliferation of data, there's less and less business cases that justify massive amounts of data lakes, right? So to your point, right, it's all about not only capturing the data at the edge but applying analytics at the edge. And I think the center of this slide really kind of illustrates performing AI or inference at the edge is the real opportunity to build kind of an ROI for these types of legacy IoT applications that's 3, 5, 10x what the original investment was. So a lot of customers are super interested. But I think the big question is how do they get started, right? Kind of a daunting task.

Absolutely. Yes. And let's dive into that why this is happening at the edge. I mentioned a lot of the training, and we talked about, a lot of the training is happening in the data center cloud. But for AI applications, they need to process the data. So in the cases where these AI applications run in the cloud, we have to actually ship the data from wherever it's being collected to the data center over a network. The AI then processes this, makes that intelligent inference and sends it back, where it's either going to learn or take some sort of action. Now simply physics-wise, this data movement adds time to the process, which makes the AI less efficient. And in some cases, depending on the amount of time that it can take, it can make some use cases impossible to run. It also means the application is subject to the bandwidth of the network. Any outages can disrupt a service. And of course, the cost of data movement grows exponentially. So that's why the value is to move that compute power to the edge. And that's what we see with edge computing, right? The compute servers are moved closer to the edge of the network, sometimes directly next to the devices that are producing the data, so it's the shortest distance possible for the AI application. So by investing companies in this smaller distributed infrastructure at the edge, they reduce a lot of the challenges of running that AI application in the store, in the factory or even on the city streets, for example. So while many customers have been investing in edge computing, this isn't exactly new, many of the devices that they've run there historically, they're not powerful enough to run AI. So it does require some net new investment. And like most transformational changes, you have to see immense business benefit to make it worth their while. So let's talk about some of the benefits that we see with edge AI and just how incredible they are. Because of course, that's what's driving this demand. So the first one, real-time intelligence. I talked about that briefly. But it means that we can invest in use cases that we simply could never run from the cloud. Industrial inspection is a great example of this, where you have hundreds or thousands of objects being inspected on an assembly line. This requires you to process and identify defects in potentially milliseconds. We can do that now. Robotics or anything that involves the safety of humans interacting with machines, this often requires split-second decisions that rely on that real-time inferencing. So that's one of the first benefits that we hear about a lot. Reduced bandwidth. I hear this a lot more and more from customers today, especially as they start rolling out these AI applications at scale. I think if you think of a lot of stores and factories, they may not be built in areas that have a high network bandwidth or even that reliability. Edge computing means you only need to rely on the local network. It not only reduces the cost but also improves the resiliency of wherever you're running this AI application. So even if you don't have that connectivity, the applications can still run. And I think that's incredibly important, especially as you start talking about mission-critical AI applications. Data privacy. This is particularly important when we come to regulated industries, where moving data can present regulatory or safety concerns. Processing data directly at the edge, this ensures that the data remains in the country or location that it's being collected. And I think we also now have the opportunity to run AIs that can obscure any potentially sensitive data. So whether that's faces or license plates, we're able to remove that data from the original capture before that we're sending anywhere off-site. And the final one is improved efficiency, which is, of course, a combination of all of these benefits. So many of the use cases we see with edge AI are about increasing automation and therefore reducing the burden of low-level repeatable tasks from humans so that they can focus on the higher-value tasks. I think a great example here is restocking inventory. Today, humans look around the store for what inventory needs to be for stock. They go get it from the warehouse. And it's a highly manual task. And AI can very easily monitor for this and alert their associates to where they might need restocking. It simplifies the task. It makes it really critical. And it reduces the burden on the humans in the store. So Blake, as you look at these benefits, are there any that you hear more often from the organizations as they're particularly starting that journey down the edge AI path?

Yes, it's funny. Probably the one that I hear the most from -- most C-level executives who are -- there's usually a top of mind, a couple of topics that keep them up at night. But I think data privacy and security is number one. I think when we were thinking about -- again, back to the proliferation of IoT devices and gateways capturing information all over your enterprise network, that might sound very scary, right? But then you start thinking about, well, based on data sovereignty and privacy laws that exist from each -- every jurisdiction, sometimes it's state and local governments, municipalities may even have different types of regulation around data. One of the things that we find is a lot of our customers, our enterprise customers, even though they may be capturing sensitive data today, they're not necessarily that interested in it, right? A good example, you used a retail example. But when you're in a retail environment, the only thing that I want to understand about somebody walking through a brick-and-mortar store is maybe some level of demographic, right? I may want to understand this person is shopping with two young children, right, versus this person maybe shopping on their own. Maybe I can understand, based on where they go in a store, what they might be focused on in terms of what they might select. But using edge computing, I think you really hit the nail on the head, right? We can anonymize that, right? So no longer are we storing raw video in the cloud or locally, right, we're completely anonymizing it and turning it into the ones and zeroes -- proverbial ones and zeroes that you can actually build into a database and act upon. That's certainly number one. I think the number two, the biggest one is real-time intelligence to me is that's just the ability to learn more and understand more, gain new insights. But that really should translate into improved efficiency. There's probably two advantages when it comes to edge computing, one of which is we can take siloed IoT and enterprise workloads locally at the edge and we can virtualize all those, right? Once we get those into a virtualized system, where we can run a maybe a CPU-intensive workload with a fair amount of storage, we can then start to look at, "Well, how do I correlate different enterprise applications, maybe IoT applications? How can I correlate those and then do analysis across both types of data that might be being collected at any given time?" So you think about improved efficiency in terms of I'm creating efficiency within the store and the infrastructure. But then I'm also creating efficiency and that now I can create -- not only do I create those new insights by being able to compare vast amounts of data locally, but then I can augment maybe whether it's the experience in the store or it's the inventory planning and the distribution center. Now I have a larger data lake. But instead of it being in the cloud, it's now local, right? So I can discard the stuff that I need, act upon the stuff that I don't need and then act upon the things that I see that provide insights.

Absolutely. It allows us to have that really powerful data management strategy, which is critical with the amount of data we're generating.

That's right.

Let's talk about the benefits of edge. What it is? What are the benefits? Let's get into some of the examples. I think a lot of the edge AI work that we're seeing today is computer vision. It's one of the most mature areas for that AI development. And of course, because we have such a ready available network of camera sensors that can detect, inspect and segment out some sort of information, it makes it a really powerful use case for a lot of these industries to get started. So as you can see from here, there's many examples across industries that drive these powerful use cases, whether it's predictive maintenance and understanding when a machine might break down and alerting before it does. I really always want to see this in my car before I have an issue, right, all of that predictive information, right? Store analytics. I think you were talking about this. Where are people going? What are they shopping? Where is that heavy traffic area? And how can I use that information to maybe provide more value to the brands that I'm bringing in my store? Patient monitoring. We all know the doctors and nurses of this country and every country are overwhelmed and overworked. So how can we make them more intelligent about how they do their work and ensure that patients are taken care of? And then, of course, factory floor optimization. Factories shipping, online retail is becoming so incredibly powerful. And they're trying to get these shipments out as fast as they possibly can. So how can we optimize all of these processes and make those faster and more efficient? So I think really anywhere that you have this data -- and again, the visual data -- video data is usually one of the most popular we see, you can build these applications or even work with application providers who've already built these applications to bring that automation into these environments. But I don't want to just talk about the high-level examples. Blake, can you talk us through some of these examples that we've seen from our joint customers?

Yes. Of course, Amanda. I mean, one of my favorite use case is actually one of my favorite verticals. And there's definitely a lot of sub-verticals underneath it. But retail is a -- and when I talk about retail, what I mean more specifically is brick-and-mortar retail. One of the great use cases and actual deployments, it's nice to have solutions that you've actually deployed that you can talk about. And this is one of them, right? So most people are familiar with the company, Kroger. Kroger is one of the largest grocery chains in the world. And Kroger, obviously through the pandemic, has seen a lot of different things change throughout their store. And one of the things that we tried to focus in on with Kroger, I think they knew that there were changes that were happening in the store. They needed to figure out a way how to flex up and down in terms of the physical footprint of the store, but then also how to augment their patrons or their customers' experience within their stores. And I don't know about you, but one of the things that drives me crazy anytime I'm shopping, whether it's going through a retail grocery store or a quick-serve restaurant, is not only do I want to find what I need to buy fast, but I also want to make sure that whenever I'm ready to check out that I'm not waiting in a long line. So one of the things that we did with Kroger is we worked with them on leveraging a partner with obviously NVIDIA but also an ISV partner called Everseen. And what we did was we worked together on how to deploy the right amount of edge computing technology with the right amount of GPU performance from NVIDIA to create a completely frictionless experience for a customer. So if you think about a self-service checkout and scanning and clicking buttons and what it means to actually self-serve, we took that a step further. And what we wanted to do is kind of create the frictionless experience for the customer. And so obviously, we're improving customer satisfaction when we can improve some of those key indicators for a customer. But the really interesting thing here is we could also help Kroger on their bottom line. One of the things that they run into is that through these self-service checkouts, they run into a lot of loss, if you will, right, or what would also be referred to as shrinkage. And one of the things that I've noticed and a big takeaway for me is that shrinkage doesn't just define purposeful theft, right? It doesn't have to be a bad actor to equate to shrinkage. And in this scenario, what we've been able to do with Everseen and NVIDIA with Kroger is we've been able to reduce, or basically whenever somebody is going through a self-checkout lane and if there's an item that they don't stand, I would bet more than half the time, probably most of the time, it's accidental, right, when somebody doesn't stand something and puts it into the basket. What we've been able to do with computer vision, leveraging AI at the edge, is we've been able to immediately cue to the customer, "Hey, you might have forgotten something," right? "Hey, did you need -- did you mean to put that in the bag without scanning it?" So it's just these little cues that we can use to drastically reduce shrinkage for a customer like Kroger. And when you're managing thousands of grocery stores across the world, these little bits of mis-scanned items add up really quickly. So this is a great example of being able to create using edge AI, being able to create a significant amount of business insight and outcome for the customer. So look, just switching gears here a little bit, another -- obviously, another area or another vertical that's ripe for disruption with edge AI is industrial manufacturing. I think that there's been hundreds, if not thousands, of companies that have been trying to disrupt and improve efficiencies throughout manufacturing and supply chain. But one of the areas that we start to see seeing real value in manufacturing is this idea of visual inspection. Whenever we go to a lot of customers in manufacturing, most of the time, we -- our customers seem to be focused on visual inspection right down to, say, a quality of one part, right? But what we end up finding is that a lot of our customers who were doing visual inspection today may be doing visual detection or visual inspection in very siloed manner, where they may have one smart camera, one smart system managing one part of the assembly line. And in this scenario, we worked with Bosch, who I think a lot of our listeners might be familiar with but might not understand all of the different businesses that Bosch is involved with. So Bosch is a large Tier 1 OEM in the automotive industry. And in this scenario, we worked on a project in a factory just outside of Milan, Italy in Bosch's -- their vacuum oil pump manufacturing facility. So basically, what their problem was not so much that they weren't doing visual inspection, but that this idea of that being able to virtualize and do multiple points of visual inspection across an entire assembly line and then orchestrating all of that inferred information. So basically, at the beginning of the manufacturing process, we may be using video to monitor the inflow of inventory of some components or maybe their gaskets or they could be nuts, bolts, screws, right, and tying together that very early stage of manufacturing all the way down the line, right, where they might be doing casting in one place, it might be assembly in another one. But basically, the idea was that if we can do inspection of every single step of the process and then we can compare those from start all the way to finish, not only can we detect areas where there might be an issue with a defect in a specific part, hopefully avoiding that part making its way into any kind of an automotive vehicle, but then we can also validate that not only was each individual station correct, right, and that they passed some visual inspection quality check, but that the process took place in a known and understood amount of time, right? So one machine wasn't operating faster than normal. So this kind of goes back to my earlier statement around being able to once you have -- once you can gain new insights across an entire process, and it doesn't just apply to manufacturing but basically any process in any enterprise, you can gain that 3, 5, 10x amount of insight. So I'm going to slither to one last scenario here, and it's one of my favorites because it's probably not that familiar to a lot of people. But in this case, we're addressing a real problem that exists in the airline industry. And this problem is very specifically, simply put, the issue of birds and drones within a runway path of an airport. Most people that fly in and out of airplane -- I spend a lot of time on aircraft, commercial aircraft, I'm sure you do, too, as well, Amanda. But one of the things that you don't realize is that there are a lot of people, both on the ground and in the cockpit, looking, scanning all the time for foreign objects that might be in the path of a runway but might also be on the runway itself. Most of the general and commercial aviation incidents happen below 1,000 feet above the ground level. And a large majority of that is essentially impact to foreign debris, drones, bird strikes. While we don't hear about these very often, other than the infamous story of the airplane that went into the Hudson River due to losing both engines to this rare and unusual case of a bird strike, but what it does -- there's kind of this other issue of when birds are detected by a human on the ground, it also creates this massive logjam. So if you think about planes that might be on approach, if there's birds or drones detected in the path of that runway, essentially the tower controller has to tell the plane to go around, right, causing -- burning thousands of more gallons of jet fuel every time that happens. And so what we did, working with the edge company, which is an ISV that's part of our NVIDIA-Lenovo ecosystem, is we worked to deploy the ThinkSystem SR650 with Quadro RTX 4000 GPUs at the local airport. So basically, if you think about an airport and you think about where the tower sits, it's centrally located, right? So when we deployed these servers, we connected them to a network of different cameras throughout the different runways and taxiways. And essentially, what we've done, with the help of the edge company, is we've been able to create a neural network model that basically looks and detects the birds, drones and any other foreign debris or objects in the area. So not only can you get better accuracy, you can also track what direction these birds are in or whether this drone flying in this area maybe have some level of impact to an incoming flight. But the key here is that the business outcome that this provides for airports is the fact that you were not having to divert aircraft, causing more and more delays, which I know you and I both hate, but it also creates a more streamlined flow of traffic both around the airport and on the ground.

Yes. And what I love about all these examples is they're all computer vision, they're all edge AI and that they're all so different that it makes you realize just the broad impact this use case can have across so many of these different industries. So I think that's awesome. All right. So now that we took out some of the use cases that people have in their heads, what is edge AI, what is it doing, how could I implement it from my industry, I wanted to talk about some of the components. Because when we talk about edge AI, it's not about simply hardware or software. Or it's not simple, I guess, is the best way to say that. And I think one of the really tricky things about these edge deployments is, most often, you're talking with a brownfield environment. Your factory, your store location, it already exists. In many situations, you're already going to have sensors in place. Your network, your storage, that's probably already defined. So it's a question of fitting the right pieces into that environment that match the space and power that you have so that you can add this additional capability of AI infrastructure, of AI applications so that they can do the work that we're talking about. So that's sort of the first challenge, is getting this edge infrastructure, which is many different components, to work together in the environment that you have. The next thing I think that's really interesting is building on top of these, so management and orchestration of these environments. Typically, we're talking about highly-distributed environments. You're not going to have skilled IT staff on site. So a lot of the tools that we're used to using in the data center and the security, even pieces that we're used to relying on in the data center, they do not apply in these environments. So management and orchestration needs to be rethought. And one of the areas that we see a lot of development is in this cloud-native architecture, very popular in the cloud, also very popular at the edge. And while these are more and more what we're seeing is sort of the default management, many teams don't have the expertise to go build and design these themselves. So looking at what are the tools available to help control, update, manage, all of these edge environments is really critical. And then of course, the final piece, or the first piece, depending on how you look at this, is those AI applications. We're going to talk about this a little bit on the next slide. But these AI applications, they're designed differently from what we would think of a traditional edge application. So they're updated regularly. You're constantly retuning them with new data, with new use cases, things like that. And so you need access to these environments so that you can make those changes. And so all of this is what makes it so complicated to go from what feels like a simple use case that we're trying to implement into getting it successfully rolled out and scaled across our environment. So let's jump into some of the challenges. And here, we took the top ones that we see. And so Blake, I want to sort of ask you about each one of these and what you're seeing. And the first one is around that POC to production scale. How are -- where is the pitfall? And how do companies avoid it?

Well, the first one for me, especially when it comes to POCs, is a couple of things. The first of which is it's for the right developer, right? And there's no right or wrong. I mean, I say that a little loosely. But there are several different ways, is the way that I should put it to go develop an edge AI application. But I think what -- most of the time, what we find and where people struggle and fail is that they may be working with a specific ISV in the market or they may be building some homegrown AI application. And almost always, it seems like the last consideration is the hardware, the infrastructure layer, the edge infrastructure. We run into customers all the time basically using either desktop compute or even general-purpose server infrastructure to POC and edge use case, right? But when it comes to deployment, when you get into the field and you want to go run a large-scale deployment of, say, let's call it, 100 retail stores, what happens is they'll deploy maybe a few hundred workstations or desktop products. They get out there. The solution runs really great right off the bat. And then over time, what you see happen is that you'll run into either maybe there's a security issue, maybe there's not enough resiliency built into the system, maybe they haven't built in enough management capabilities to do some of the security and update stuff that you talked about. So what ultimately happens is they spend a lot of time running trucks, right, just to reset systems or just to find out that power supplies have failed or maybe somebody poured a Coca-Cola down the backside of a desktop computer sitting in the back of the store. These are all real scenarios. And I would say they often inhibit moving from a POC to production. And I always say one of the things about the edge is you're dealing with constrained resources. So if you're going to deploy a resource and you're going to do testing, it's best to test on the equipment that you're going to deploy at scale. And here at Lenovo and working with NVIDIA, we've developed purpose-built edge computing platforms, scaling to devices that are sub-1 meter, all the way up to what would be considered a server-class product in half width, short depth, for a lack of a better term, an edge server, right? So these are the scenarios where I think we find that people have a lot of starts and stops, right? It's because they're not really developing with the intent of scale in mind, right? They're just always searching for how do I get to that business value immediately, right? But as you know, lifecycle management, that's just as -- that's -- deployment and lifecycle management is, in a lot of cases, is just as hard as the development work itself.

Yes. I think that's so true. And it hits on our next one, which is about AI lifecycle support. Because I think the other thing about the edge is a lot of the environments that we see today, there's the set it and forget it. I hear from a lot of customers, "Oh, it hasn't broken yet. So we just like to keep running." And unfortunately, that's not the way AI works. You are constantly tuning it over time. And so you need access to these environments somewhat regularly so that you can make those updates. And you're right, that requires management and tooling and things that people aren't thinking about. Because in the POC phase, yes, you can do that relatively simply with one server or a handful of servers that are all sitting there right next to your development team. So I think that one is definitely connected and something people really need to think about with AI. You touched a little bit on security. Do you want to talk more about what that challenge is and how -- yes, again, how people can think about it and avoid it?

Of course. Yes, so I spoke a little earlier on about maybe more data privacy and sovereignty. But the other reality of this, and I think I mentioned about this is something that keeps the C-level folks awake at night, right, which is now I just went -- if I have 100 employees that have 100 PCs, I just went from having 100 employees to now having 100 different pieces of equipment sitting in every store, right? And I think that not only is having all of the kind of the modern security features that you would expect from a Tier 1 OEM like Lenovo in place on your personal computing platforms, it's just as important. An edge computing box is no different than a PC operating on some of these desks at home, other than the fact that nobody is sitting in front of that machine. Nobody has it in their possession. So these systems, in a lot of cases, they could be in a store, maybe in a locked telco cabinet. But in other cases, they might be in a janitor's closet in, say, like an airport. So you can imagine an airport being a pretty sensitive area, right? And to allow somebody to just physically come into one of these closets and start plugging into different I/O ports, you need a lot of security baked in. Not only do you want physical security, like things that we've implemented would be like movement and tamper detection. So if somebody were to take one of these servers off of a rack, we can actually wipe the drives in them. That's an automation that we can build into it. If somebody were to rip one of the covers open, right, and try to recover a drive, same thing, we can lock it down. The other thing is that I would say is probably -- more often than not, we find that there could be, and this is somewhat security-related, but it's more about resiliency, we may have somebody in the back of a retail store may be plugging their -- they may take the Motorola phone and plug it into one of the USB ports, right? And depending on what they're trying to do, could have catastrophic consequences for something like a desktop PC. In the BIOS levels of these systems, we have the ability, we've thought -- we've had the foresight to think about, "Well, what are all those scenarios, right, where while it may not be malicious by nature, but what are the scenarios that could create havoc for somebody who's probably entrusting the entire efficiency and operation of their store on one of these edge servers?" So those are things that I think people -- again, we talk about POC to production. These are just things that you kind of learn as you go. And hopefully, you don't get bit for your own ignorance, right?

Exactly. I think IT has relied for so long on the data center. That's their safe space. Nobody can go in there. Nobody can touch these systems. The edge is not that way. So rely on the folks like NVIDIA and Lenovo and the other partners out there, who've deployed these things before, so we can help you identify here's the things you need to think about, right? Yes. And the final one we'll touch on briefly is not all edge environments are created equal. Even if you're a retail store and you're looking to roll a solution out across all the stores that are in your region, you might have some smaller stores, you might have some medium-sized and you might have large. And you need solutions that can scale across those. You talked about some of the Lenovo ThinkEdge line and being able to scale from the smallest possible footprints for power and performance, all the way up to -- I've seen half racks deployed at the edge, right? Like it depends on what the use case is. So having solutions both in the hardware and in the software that can be scaled so that they can bridge these different types of environments, I think, is really important.

Couldn't agree more.

All right. Let's get into the platform. And I'm going to let you kick us off. We've talked a little bit about the hardware. But can you talk us through the sort of the portfolio that Lenovo has, of course, with NVIDIA GPUs? And then we'll work our way up the stack.

Yes, for sure. So number one, I have to start with the portfolio, which I -- of course, managing the ThinkEdge business here at Lenovo, I'm very proud of the portfolio that we've put together. We definitely think it's the broadest. We think it's the most purposefully designed and, of course, the most innovative. There's a couple here that we've -- kind of some notable mentions across the portfolio. Starting from left to right. The SE70 is an NVIDIA Jetson-based product. So this would be your one-to-few cameras, right, doing some light AI. Although I say light, it's actually very surprising, the amount of compute that's in here to -- comparing it to just a few generations ago, it's probably 10 to 20x the performance. And moving kind of that -- moving from left to right here to the SE350. The SE350 is -- again, being purpose-built, you can tell it's half width. It actually fits on a backpack, believe it or not. But this product can run an NVIDIA A2 or T4, run an entirely virtualized stack in that retail environment, maybe doing inference on, I don't know, 20 to 30 different cameras at one time. And then once you get into kind of the flagship of our portfolio that you see the ThinkEdge SE450. This product, again super compact. It looks kind of large in comparison to the other two products. But it's basically is a telco cabinet depth. It's 2U tall, can fit into a telco rack. So if you've ever been into a store, remote office or branch office, there's not a -- there is no place to rack a server, right? There's pretty short -- if there's a rack at all, it's usually pretty short and it's very limited space. And the other thing about all of these products that you see here is if you think about physical location, not every place has a rack that I just talked about, right? So we have all sorts of wall-mounting options, VESA, DIN, things that are important when it comes to deploying at the edge. That kind of covers basically the server side of the equation or the system side of the equation. One of the interesting things that I -- and this is the reason I love talking to customers because we're known -- Lenovo is known for being a Tier 1 infrastructure provider, among a lot of other things. But one of the things that we do well, right, we've always done infrastructure very well in the data center. And one of the things that we've worked on and focused with is not only doing that certification with our partners, like VMware and Red Hat and Azure, working to make sure that you can orchestrate Kubernetes, K3s, K8s on these systems, but we're also focused on taking tools like XClarity that exists in the data center. XClarity is our server management platform that is used in basically -- probably any data center that leverages Lenovo ISG products is leveraging a XClarity today. One of the things that we've done is we've augmented that software offering to run either on, on-prem, completely disconnected from Lenovo, or in a hybrid cloud environment or even in the cloud but basically allowing our customers to deploy and create a zero-touch provisioning environment. So when you ship an SE450 to your small ice cream shop, your chain of ice cream shops, you're sending one technician that's basically just racking the system or mounting it on the wall and plugging in a couple of cables. And then they're calling a network administrator, who might be thousands of miles away, and doing all of the bring-up, bringing down the BIOS configuration, all of the firmware and OS updates, actually provisioning and loading the OS, setting up the virtualization layer. And then ultimately, from there, once you brought that into that environment, that's when you get into the next part of it, which is orchestrating all of the different applications. So in a sense, it doesn't really matter if you're running any of these different container orchestration or management platforms, whether it's VM's or Kubernetes'. Lenovo makes it easy to bring a device like any of these that you see here out of the box and provision it with the least amount of time spent on site. So for me, that's a pretty powerful statement when we're talking to our customers about this daunting task, right, of deploying server-grade, GPU-rich products in these environments, right? This is one thing that if we can help save them time on the deployment, it's a win-win, right?

Absolutely. Yes, I think that's so important, and giving access to the tools they're familiar with but making it that much easier to control that sort of fleet of devices that are out there. Yes. And then of course, in addition to the GPUs we make on the NVIDIA side, we're very focused on the AI applications and ensuring that we're providing that sort of seamless layer of AI application, software and support across all of these different platforms. And so NVIDIA AI Enterprise, this is our accelerated data science pipeline. It allows you to streamline that development of your AI applications as you're developing them, all the way through to production deployments out there at the edge. We have frameworks, we have pretrained models, we have development tools, all of these pieces that not only allow you to build something from scratch. We also work with many, many application providers, some of them that were mentioned in our use cases earlier, to ensure that not only do these run in an optimized fashion on top of the accelerated infrastructure, but also that they can be supported by both NVIDIA and Lenovo so that when you have a problem, you know that we're there to solve that for you, which is particularly important in these edge production environments. And as you talked about earlier, these obviously run across all of the different industries we've talked about.

Yes. I think one of the beauties of having a partner like NVIDIA is obviously the frameworks and the software, the things, the tools that NVIDIA develops to enable this marketplace are really one thing. But when it comes to the partnership, NVIDIA developed some of the most powerful GPUs in the world, right, some of which are made for the data center, others of which we've collaborated on to work at the edge. But the important thing is that when NVIDIA makes a statement about this many tops, or flops in certain scenarios, that a company -- when we partner, we are actually taking these scenarios, right, these -- probably between the two companies, thousands of different ISVs that we're validating on our systems, that we're actually validating that, "Hey, we see the same level of performance that NVIDIA states," right? So in a way, we're working together to make sure that when an ISV or an end user comes to us looking for a certain amount of performance at the edge, that the applications that they deploy will essentially run as advertised, right? And so I think that's kind of the key to the partnership to me.

Absolutely. All right. So the next to last thing, of course, we're going to allow time for questions, we want to allow those steps to getting started. So we've talked through these -- throughout the presentation. But just really quickly, the first and foremost thing is identify what that problem is and ensuring that, that problem not only has -- is technically feasible but also has a strong business opportunity. You heard from all the examples we gave, that return on investment is what's going to get the fuel into your POC process. That's what's going to give you the visibility to help you push these things through. Once you've identified the problem, you then determine what is the data? What are the application requirements? Do I want to build this? Do I want to buy this? And reaching out to Lenovo, NVIDIA and all of our trusted partners, we can help you with this part of the process. You don't have to come to us right at the end just when you're doing hardware sizing. We actually can help you figure out what those unique data and application requirements are or connect you with one of the thousands of partners that we work with. You then want to analyze your edge capabilities. What is the infrastructure you already have out there? What is it capable of? What is the environment that you can build into? And that's going to help as you make decisions on the investments that you need to make, the hardware that you need to purchase. Of course, then you want to roll out the solution. POCs, we talked a little bit about the challenge of rolling them out. You want to make sure that you're looking at that POC with the scale that you intend to get to in mind and testing it on the production-level hardware and software that you plan to actually deploy. So typically, we see these take anywhere from 3 to 12 months, where you're testing these out and then ideally scaling and rolling those across your entire environment. And then the final piece that I talk about a lot is sort of what we call celebrate success. The really interesting thing about AI is once you deploy one, you get the bug. I say this a lot, AI is like a tattoo, you don't just get one, you get many, right? So you're going to get the full sleeve, eventually. So celebrate that success, bring the learnings back to your organizations and then see where else AI can help you solve additional problems down the line that can either leverage the same infrastructure or you can scale out and expand as needed. So anything you want to add there?

Yes. Look, I couldn't have put it better myself. I think if there was one thing I would add to it, it would be probably making sure that you don't get overwhelmed with the amount of business improvement opportunities that might exist at the edge. You are -- you as the customer, you understand your business better than anybody. We at Lenovo, we like to think that we have the technology to solve some of the humanity's and business' largest problems, right? But the reality is that if you're going to start somewhere, you need to start with the problem statement, right? And you need to be able to quantify what the objective outcome is that you want to drive, right? And once you've identified that, then you can kind of work backwards. And I always -- again, Lenovo, we're infrastructure experts, right? We have taken the lessons learned through being one of the greatest PC manufacturers of all time, the greatest infrastructure providers all of the time. We've already applied that knowledge and know-how. But what we don't -- what we can't do all the time for customers is pick that million-dollar use case, right? We can make the connections to the right ISVs, create -- help select the right virtualization and storage options. So that would be one thing. And if I had to add anything else, there's kind of two ways to look at edge computing. Edge computing is basically just the consolidation of enterprise workloads that either are happening today or may happen in the future outside of the data center, right? So consolidation can essentially be the catalyst of change in an edge location, right? That's kind of the -- to your -- I think, to your example, the first hat you get is you build an ROI for workload consolidation. Then when you get into starting to analyze all of the data from those workloads you've consolidated, the AI portion of this, whether it's natural language processing, predictive analytics or computer vision, that's where you start to see this ability to unlock this again 3, 5, 10x return on investment, right? So consolidation can be the baseline business case. But the acceleration through AI is really what's going to -- that's going to be the wind in your sail for the future. So that would be my only add, Amanda.

Awesome. Well, I think we've shared a lot. And hopefully, everyone on this line has gotten a lot out of the conversation. This is not the end. You can learn a lot more. NVIDIA and Lenovo have been working hand-in-hand together to create a ton of resources for you. So wherever you are on your journey, if you're just getting started, if you want to learn what other companies have done on their journey for edge AI, what are the benefits they've seen, how do you position this with your leadership, there's a great campaign page that's available to you. You can actually see this in the links alongside the webinar, so you don't have to type it in yourself. But I do want to take a couple of minutes to answer some questions. So why don't we turn it over to the Q&A?

All right. So I know questions have been coming in fast and furious here towards the end, Blake. I'm just going to start with a couple, and I know we've got a few minutes to answer them. So the first one I wanted to ask you is what is the best AI workload for someone getting started?

That's a tough one. I think it goes back to you can start simple and small or you can go big, right? You can try to solve multiple different AI workloads at the same time to drive some level of value. But I think it's critically important, as I kind of just covered, that pick a simple problem that you're trying to solve. And it could be one of many problems that you aim at solving in the future. So if you're looking at a manufacturing process and you're thinking about, "Well, hey, I have a problem with defects that cause -- the initial problem is the defect itself, but then there's this chain of other problems that it creates in the process of manufacturing." If you could solve the simple project -- or the simple problem first using inference through defect detection, I think you'll gain kind of the sharp skills that you'll need to go solve and tackle the next one. And of course, I'd be remiss if I didn't mention that if you're not -- or maybe you don't have the technical chops or the team or the resources to go solve or build your own AI application, between NVIDIA and Lenovo, we have a broad range of independent software vendors that we work with and validate with.

Yes, I think that's a great answer. And making sure that it has that business value would be the piece I would add is I've seen organizations say, "Well, we're just going to experiment over here." And they pick something that's small but also something that doesn't really matter. And then they struggle to get the attention and the resources and the funding that they need because it just doesn't meet those requirements for the org. So start small, but start with something that's important to your business. Because that's going to make the difference down the line.

That's right.

All right. Another great question. So at what point should a customer consider the security required for their edge infrastructure?

Well, look, I mean, I'm going to be a good steward of our business and everyone. And I think that is -- the correct answer is immediately, right, from the get-go. You want to come into a design, an AI application design, whether it's you're focusing on the hardware infrastructure and maybe you know what type of acceleration you need. But a design principle and pillar that has to be implemented upfront is security. I always say this to customers as they're getting started is the big question I ask is how many people throughout your organization, from different technical discipline, maybe it's infrastructure discipline versus OT, right, versus either product security office or your enterprise security office. Engage with them early. Let them know what you're doing. There's -- oftentimes, there's a lot more resources in pockets of areas of your business that you may not know exist. And a good example is here at Lenovo, we have -- our legal team sits very closely to our product security office, right, so -- and there's a lot of nuances. We talked earlier about data privacy and sovereignty and security from a data perspective. But then from an infrastructure perspective, there's all these other considerations that you have need to take into consideration. So to summarize, unless your POCs are going to exist inside a firewall, inside a physical area of -- some area within your company, security has to be paramount, right? Because the first thing you're going to want to do when you nail that killer app is move it outside of the lab and into a store, a manufacturing floor or maybe a city. And this is typically where companies get bit.

Exactly. Frame in them early, it feels like you're putting too many people in, but it will save you time in the long run getting this out if you're thinking about those considerations ahead of time, 100%. And that's definitely all the security people who are listening to this right now, they're going to appreciate that. All right. I think we have time for one more question. So I've seen here, and I'm going to expand it a little, it says what's the average length of time it takes to set up an SE450? But I'm going to expand that to be a little bit like how quickly can folks get going with the infrastructure and things required for an edge POC? Let's start there.

Well, look, so there's a couple of different elements of the installation itself, right? So when I talked earlier about having a specific installation technician right on site, if you're leveraging Lenovo Open Cloud Automation with any of our ThinkEdge products, we have the ability to ship the system bare metal, right? So there's the physical installation, which is what we might refer to in the data center as rack and stack, where this might be mount and plug, right? But essentially, I would say the average physical installation is probably anywhere from, I don't know, 10 to 20 minutes, 30 minutes, 0.5 hour at most. And then going from a bare metal state, doing the -- making sure that you have the right BIOS settings, downloading, updating all of the firmware and security patches, all the way up to bringing in an operating system and getting it to a fully functional state, your total time with physical installation is definitely under an hour. Depending on what type of clustering you're doing or how many different virtual machines you're running, that can vary. But it's typically dependent on whether you're orchestrating Kubernetes on Linux or maybe you're going to run VMware and you've got maybe 20 different VMs to spin up. Those are kind of the considerations. But I would say, on average, we've seen that through leveraging it to like a Lenovo Open Cloud Automation, we've seen a reduction of over 3x, so basically taking something that might take 3 hours and moving it under 1, right, but -- and subject again to every specific scenario, such as the case with all things edge computing.

Exactly. Yes. Well, I know we're right at the top of the hour, so I want to be cognizant of people's time. So Blake, thank you so much for joining me on this. I think it was a great conversation. And for folks who are still waiting for their questions answered, we did get a ton in here, so we will collect those and share out information. But I want to thank everyone for joining us. Thank you, Blake, for being here, and hope you have a good rest of your day, rest of your evening.

Thank you, Amanda.