In the latest episode of the Mission Matters Podcast, Akhil and I sat down with Dan Wright and Pradeep Nair, co-founders of Armada, which is building distributed AI infrastructure for the world’s most demanding environments.
Dan has a phrase that stuck with us: “AI drags infrastructure.” Everyone sees the value of AI. The hard part is deploying it where it’s actually needed, like oil rigs, mine sites, aircraft carriers, and national labs. These are places where the cloud simply can’t reach, and that’s where Armada comes in.
In this episode, we cover:
• Why the cloud breaks at the edge (latency, sovereignty, and cost)
• How Armada is enabling real-time AI in fully disconnected environments
• Armada’s role in the DOE’s Genesis Mission, the most ambitious AI project in U.S. history
• Why winning the AI race is just as much about infrastructure as it is models
• What it takes to build mission-critical technology for mission-critical customers
As always, please let us know what you think. And please reach out if you or anyone you know is building at the intersection of technology and national security.
You can listen to the podcast on Spotify, Apple, YouTube, the Shield Capital website, or right here on Substack.
One more quick reminder that registration is now open for the third annual National Security Hackathon, taking place May 1-3 in San Francisco! We are partnering with the U.S. Army to award $50,000+ in cash prizes to hackers building solutions to crucial U.S. military problem sets. 🇺🇸 Register now here.
Transcript
Maggie 00:36
In this episode of the Mission Matters podcast, were joined by Dan Wright and Pradeep Nair the founding CEO and CTO of Armada AI. Armada is building the future of distributed AI at the edge.
Akhil 00:50
Armada is a quintessential company addressing both a commercial and national security pain point, and that is how to get connected compute to the edge, to these potentially remote or disconnected environments for mission critical applications. That’s whether it’s an oil rig, a mining site, or even the US military and its allies. You can think of Armada as orchestrating both the compute via its containerized mobile data centers, as well as the connectivity associated with connecting those together to enable and accelerate AI applications, data analysis, where it is needed and in the timelines in which it’s needed.
Maggie 01:31
I mean, essentially you can think of Armada’s Galleon product as a bunch of CPUs and GPUs in a hardened Conex box with Starlink connectivity and all of the software infrastructure that you need to actually put AI applications on that edge. Armada’s technology is particularly useful in situations where you cannot assume reliable connectivity back to the cloud. Today, if you look at LLMs ChatGPT, the vast majority of these applications rely on having unbroken access to large cloud data centers to do all of the processing that you need to use those AI models. However, you know, you can imagine if you’re deployed out on an aircraft carrier in the middle of the Pacific during a conflict or on an oil rig, or even at a mine site or an oil and gas site, you might not have access to those cloud compute capabilities, but you still want to be able to use these powerful AI models to make your tasks your life easier.
Akhil 02:27
Yeah, Maggie. I think a lot of us here at Shield Capital, whether we served in the military or have been deep in the space, recognize that. A lot of these environments, you just can’t connect back to the cloud and you have to be able to analyze, assess, and compute near the point of friction or the point of influence, so where it needs to be done. Even if you have that connectivity back to the cloud, it’s critical that analysis, that timeliness associated is done nearest to those who are actioning.
Maggie 02:57
Yeah, it’s also valuable having this edge compute capability as it allows you to process the data potentially where it’s actually collected, and it allows you to have those data centers right near your energy sources. Dan and Pradeep have both had long careers as AI and data infrastructure, technology company executives. Dan served on the executive team at AppDynamics and DataRobot while Pradeep was a vice president at Microsoft Azure for 10 years, and a VP of engineering at VMware. Now onto the conversation. Dan and Pradeep, thank you both so much for coming on the Mission Matters podcast. I wanna start out with a question, Dan, where is the Department of War today in terms of AI adoption and infrastructure build out, you know, what’s actually being used? There’s lots of buzzwords, lots in the news right now about how the DOW is actually using ai, but what are you seeing are some of the specific. Use cases that you’re excited about that are driving real value for the war fighter?
Dan 04:01
Yeah, I mean, we have a saying at Armada, which is that AI drags infrastructure and what is meant by that is that typically how things work, not just with the DOW, but with any customer that we work with, is first they see the value of the ai and then they’re like, oh, how do I enable this everywhere? And so that’s kind of the stage we’re at now, where now the DOW has obviously understood the value of drones. I mean, I think everything that’s happened in Ukraine and other conflicts has really highlighted the necessity of not just drones, but drones at massive scale, you know, drone swarms. And then the obvious question becomes how do we process all of that data? Locally, including in these more remote areas. And then on top of that, you see a lot of the foundation model companies now working with the DOW in a very close way that wasn’t happening in the past. And those capabilities too, it becomes, okay, well how do we actually deploy this? So then you heard Secretary Hegseth recently at Star Base talking about, you know, deploying these types of technologies at the edge. So the Edge is now really a focus, but that infrastructure rollout is happening as we speak right now.
Maggie 05:25
Yeah, Pradeep, we’re starting to see some of these use cases expand drones, others, what are some of the existing technical challenges that still exist for actually adopting AI in these mission critical organizations?
Pradeep 05:39
I think a couple of things it’s what we are seeing with a lot of our customers, we are actually deploying into countries or places where there’s an existing cloud region, right? So that, that’s something interesting to notice like most of our deployments are starting to happen and hey, there’s a, you know, one of the existing hyperscaler cloud region, but they want an edge deployment and that’s driven by. Primarily three things. One latency. They want to have you know, the compute close to where the data is generated. The second portion is sovereignty, right? That’s where the AI stuff is very important because if you see what’s happening as Dan was mentioning, a lot of large language models being, you know. But many of the enterprise customers or even government customers want to take it, you know, fine tune it on their you know, confidential or sensitive data and then push it to the in front for the edge, right? So that’s what we are enabling. One of the struggle has been, Hey, I have all these models, but I want to have the compute on the edge. Now how do I bring all this together? Which is one of the pain points we are solving is, Hey, okay, you can have all these AI models, you know, generated in the cloud, but how do you put to action on the edge that requests like, you know, distributed compute, a single amazing customer experience that allows customers to focus, to switch between cloud and edge. So that’s one of the things. And the last one we see is cost as well. Like in terms of the problem we are solving overall. Hey. Yes, I have the models in the cloud. I don’t want to keep pushing all the data to the cloud if I want to like, you know, run inference and get output out the edge and then push the post-process data to the cloud. That’ll save obviously you know, significant on my, cost as well, because you know, people are spending lots still on ai, cost, all the training, et cetera. Now how do we start, you know, optimizing it so that you don’t keep spending just on the large models or the large data and push all of that to the cloud.
Akhil 07:43
Pradeep, you’ve experienced this built on the cloud side and now the Edge side. And Dan, you mentioned I love that sort of statement, AI applications are dragging the infrastructure. I think something we’ve observed, and this is not just in the national security spaces, when that does happen, I mean it takes a hot minute for the, in infrastructure to get developed, in fact, much longer than it does necessarily to take an AI application, right? And to build the potential use cases. And so Pradeep kind of curious on, as we think about the challenges, not just in the application development, but you know, as we think about that dragging infrastructure. What are the challenges of actually building and adopting that scaled infrastructure in a distributed manner for some of these mission critical organizations? And that’s not just national security, that’s oil and gas. These organizations that that, that have similar mandates when it comes to the deployability and challenges associated there.
Pradeep 08:36
Yeah so interesting. I think as a, as you said, I was you know, fortunate to be part of the, you know, building the cloud journey for over a decade. What I significantly saw was what Cloud enabled was, hey, you can think about developers around the world could actually focus on the application and start deploying anywhere in the world, right? They don’t really care where the infrastructure is sitting. It’s just a dropdown for them in the broader experience. And I just go focus on my application deploying that. That is a similar thing we are doing here. Hey, you can think about distributed edge compute sitting across everywhere around the world. You focus on the application now what challenges we see? It’s literally if your application is modernized running on Kubernetes, like you’re able to do that. Then, you know, moving to the edges relatively straightforward. Where we see challenges, if it’s set of legacy applications that have been done that has very legacy dependencies, then you actually have to go through sort of a transformation or the second thing we have seen is. If you are actually too much tied to one of the cloud provider, right? You have so much dependency on some of the cloud services or specific cloud services coming from a cloud provider, then you have challenges transforming. But what we are also seeing a lot of the like, and if you think about you know, journey of multi-cloud like enterprise customers are deploying on multiple clouds, which means that. The app developers are like transforming the application to say that, oh, I should be able to run my applications on an Azure or A GCP or on Oracle, et cetera, right? That’s how they are thinking. Which makes our job easier because, hey, if you have you know, transform your application or you built your application from the scratch, that it can run on primarily any infrastructure, any cloud provider, then running it on engine infrastructure becomes relatively very easy. For example, we are at, we are one of the applications we actually, they were running on Azure. We were able to actually, you know, you know, enable them to run on the edge within less than 24 hours, right? Because it’s all you know, native Kubernetes stuff. So it is very easy for us to, you know, migrate them over.
Akhil 10:51
Thanks for that, that’s great. And I wanna pull on something you mentioned with a lot of these legacy industries. You know, legacy is not necessarily the right word, but with a lot of industries, like in national security. Basic dev tools, infrastructural tool just don’t exist in those environments. And what we’ve been finding, Maggie and I, with our portfolio companies here at Shield Capital that are on the AI application side is they’re allocating 20, 30% of their critical engineering talent. Not to the application, but to the deployability infrastructure and how it relates to the edge. And so I’m curious a little bit about how do you think about working with startups in these. Really advanced AI application companies, figuring out what that edge deployability looks like and maybe taking the burden off of them so that they can focus on, I think as I’ve heard before, focused on just making really good beer or making that really good application.
Pradeep 11:43
Yeah. So super interesting. You know, we, you know, we spend a lot of time between, you know, Nvidia of the world and then, you know, and the cloud providers like Microsoft. If you think about the stack, what we are seeing is there’s obviously the infrastructure layer which is you know, hardware and all of those things. Then you have the past layer with Kubernetes. Then you have a lot of ML ops platforms, and then you have the AI applications, right? That’s. The stack is starting to work, and this is where some of the customers struggle because if they have to do it all by themselves, they have to go all you know, together by themselves. And then there are startups, to your point, is working on couple of these portions and oh, they have a really good ML off platform, or they are, you know, developing really apps. So what we have been is like we think about partnership very key. Like how do we make the experience simpler and easier for our customers? Is the partnership with these, you know, ML ops and AI apps. And we work with the, for example, have a specific team what we call as the marketplace team. What they are actually, their full-time focus is to like, you know, work with partners, onboard them onto our platform, make their deployment very easy. They don’t have to worry about Hey, what infrastructure, et cetera. Very similar as I explained in the cloud. So they can actually scale faster, right? And what that gives us the abilities, Hey, we deploy the infrastructure. The number of ML ops platform integration or the number of AI apps we support goes really well. So that’s and my team is. Deep infrastructure specialists and you bring in the application teams from the, you know, the startups or the app companies is what we are focused on. So that’s really what we are focused on and how we do through is partnership. We see that struggle like very much. They e even for example, even if I’m an app developer company and I can deploy it in cloud, then I don’t know how to move to edge, right? So we actually take care of that easily for them. So that’s really a problem. And through partnership is what we are solving. That
Dan 13:49
Just to piggyback on that, I mean, we say pretty, but I like to say our value prop is the three S’s. Speed, scale, and sovereignty. And the speed part is really important for the end customers that we’re serving, where it’s all about. Hey, I’ve got, you know, mission critical problems in these edge locations. I need them solved yesterday. And you know, for example, with the Navy, and this is public, we did a UNITAS exercise where we were able to rapidly deploy one of our Galleons on a ship. And we were able to show, Hey, you can take these really powerful models and for the first time, it’s never been done before. Run them at scale, you know, totally disconnected from the public cloud in these very remote locations. The speed piece of that is really important. And then once you do that, you create a ton of value, both for the end customer and then these model companies, which also look at us as a fast way to get deployed in these sites. And then you start to scale. And then the sovereignty piece is really important. I mean, I think what we’re seeing is that you need to take a broader lens when it comes to national security. And we work not just with defense, but we work in, you know, energy. Also thinking about like other critical infrastructure, like airports, right? All of it is really national security where you need to be able to do 24 7 monitoring of these sites, secure them, and then run all of these really powerful AI models securely locally at the edge.
Akhil 15:20
That’s awesome. Thanks Dan. And yeah, I think very much aligns with how we think we, I mean, it’s mission critical, right? It’s about how do we. Safeguard and enable the varied use of AI applications with the right associated security measures if and when uncertainties happen. And that could be just a commercial cloud outage, which happens more frequently than we care to admit, or something more nefarious that affects not just the national security domain but our businesses and commercial enterprises. Dan, can you gimme another example? Maybe not from the military. Going to Pip’s comment about how you think about partnering with the best in breed of AI applications.
Dan 15:58
Yeah. One example I really love ‘cause of the impact that it has is the example with Alaska Department of Transportation. And this started with us managing all of the connectivity as well as other connected assets for the state, like drones. They use Skydio drones, which are really great for emergency response, specifically using it for things like avalanches and floods. And obviously Alaska’s a huge, very remote state, and so a lot of times when these natural disaster strike people are affected, stranded in very remote locations and you can’t wait. For intelligence to respond, you need that in split seconds, just like you do in a defense scenario. And so what we did is for the first time, we deployed one of our Galleons, these modular data centers in Alaska, and we were able to cut their latency from over a day to near real time to respond to these things using all the data from the drones. Right, and that’s using AI models that are very, people are very, you know, familiar with. But the difference between deploying those at the edge locally versus deploying those in the cloud is literally, you know, life and death. You know, 2020 over 24 hours. I mean, people cannot wait in that scenario. You have to have the intelligence in the moment for it to be useful if you don’t, responding the moment the data’s almost like useless. And so that is our focus. And I like that example because I think that is something that is true everywhere in the world. And you’d be surprised how many different locations have this same problem around the world. I’ll just give you another example. I’m about to fly out to Australia. We’re making a big announcement with a customer in Australia and. I was talking to one of the former Prime Minister there yesterday in preparation for that trip, and they said, Hey, we have that exact same problem, you know, very remote part of the world. There are no large data centers outside of, you know, maybe Sydney and Melbourne. And you know, when fires break out or you know, even just to provide the opportunity to you know, educate the population, provide infrastructure for education, obviously huge area for mining and oil and gas as well. So going back to energy, your stuff is mission critical and we don’t have anything like it. So we just see a huge amount of opportunity with those types of high impact use cases in these more remote locations that have been underserved. From infrastructure historically.
Maggie 18:50
Yeah. Dan, an another pretty interesting use case. I saw you all announced it was late last year. You announced that Armada is going to be an official collaborator on the Department of Energy’s new Genesis mission. So I wanted to see if you could talk a little bit about what that mission is and the role that Armada will play.
Dan 19:07
We are really excited about Genesis mission. I’ve really advocated for a project like this for my whole career, so I’m just. Pumped that we’re doing this as a country. The things that we’re doing there are twofold. So one is we’ve been working with secretary Wright and Dario Gill and the whole team Anthony Pugliese over there. And they’ve been great about just connecting us to all of the 17 national labs so that we can go there. We can meet with them and we can understand what are the gaps in infrastructure across the labs. Because again, AI drags infrastructure. Genesis’s mission is the most ambitious AI project in the history of the country. We’re talking about doubling productivity when it comes to you know. Science and engineering within a decade. That’s the goal. So if you’re going to do that, you’re going to need to deploy AI at a scale that has never been done before. But you can’t do that unless you have the right infrastructure to enable it. So that’s the first thing we’re doing, is we have gone to all of the national labs and we found out. Where are those gaps in infrastructure? And again, our value prop is speed, scale, and sovereignty. We can fill those gaps in infrastructure right away to enable us to. Accomplish the objectives of Genesis mission, which include getting, you know, big wins this year, right? We can’t wait, you know, years for traditional brick and mortar data centers to be built. So that’s number one. The second thing is there is a goal to create a common AI platform across all of the national labs. They don’t wanna operate in a siloed fashion where each Napa National Lab is doing its own thing, right? The power of Genesis mission is taking all of the data from all of the national labs and bringing it all together and having them work in unison. And so what’s really unique about our platform is the full stack nature of it. The fact that not only can we deploy. These Galleons you know, faster, more flexibly where you can scale up versus overbuilding and do it totally sovereign and disconnected from the public cloud in a way that nobody else can. But the full stack piece of it where each one of these things acts like a node to a distributed private infrastructure that can be leveraged locally, but then also. Coordinated across sites. And a very tangible example, getting into the weeds of this a little bit is take you know, take one of these model companies, right? If you look at who are the collaborators, we’re one of them for Genesis Mission, but also all of the Frontier model companies are part of it, all of the major ones anyway, you’ve got pretty much all of the GPU. Know, companies participating in it as well, but then the question becomes, okay, if I deploy those models on one of my sites for cybersecurity and other reasons, I’m gonna wanna use that locally. It also helps with cost, which is a big issue. Running the models is much more cost effective at the edge versus running it in the cloud. And we can get into that. But then I want to take. All of the improvements that I get, fine tuning the models on all of that data, and there’s a huge amount of data at each one of these labs, and I wanna leverage that across all of my different sites. And so what you can do is you can take one of those models, whether it’s from OpenAI or Xai, or you name it, you can run it on that data locally. You can fine tune it to the data, and then you don’t have to ship all the data back to leverage that across. All of the other labs, you just ship the fine tuned model back and then you leverage that across all of the other labs.
Akhil 22:57
Hey Dan, you hit on something that I think some folks first learning about Armada may not entirely know or have a perception around. It’s not just about the data center in a box, right? Armada is not, Hey, I’ve got this galleon and I’m just gonna plop it somewhere and it’s gonna have some compute, some GP, you know, a couple of server racks. And have some power. The whole thesis is how do you interconnect those in a way that allows for this nearly hybrid compute environment, depending on the type of application and depending on whether you’re optimizing for cost or resiliency.
Dan 23:32
Yeah, exactly. And I’ll let pretty comment on this ‘cause he, he saw this firsthand at Microsoft where, what, like when PIP and I were first talking at the beginning of Armada, the opportunity was. Oh, wow. There’s 30% of the world that has these big hyperscale data centers. But what about the other 70%? There’s all these gaps globally in infrastructure, and that’s why we call ourselves the hyperscaler for the edge. And you know, the Galleons are super cool and they’re very impressive, which in some ways is a big advantage for us. But in some ways, you know, we try to really make it clear it’s not just about one galleon. It’s about. The full stack nature of it in the fact that each one of them is a node to a distributed system that can plug all of these gaps in infrastructure globally, and then allow you to leverage your intelligent across all of your different sites.
Pradeep 24:27
So I think Al interestingly, we talk about the world as, I mean, hey, there is you know, cloud for training. And then there is like what we see the world evolving to fine tuning and inference, right? That’s really it where we are with this thing. So what Armada is really playing a key role as being the platform for fine tuning and inference right now. All the edge boxes we are talking about is mostly for inference deployment. So that’s where we, basically, what we do is. We give the platform, but we also put the entire full stack solution with hardware and with the modeler data center. That’s what we’re doing. And there are cases in fine tuning, we are actually partnering with, you know, data center providers or colo providers where hey, they will actually bring in the hardware and it sits in a colo. And we give the platform, right? So what we are focused on is that customer wants a single platform for both fine tuning and inference, and then I bring down the models from the cloud then fine tune it like on my sensitive data. And then I click a button and I push it to inference for the edge. So that’s really like what. What Dan is talking about, that’s where an edge is actually evolving with that per premises. An edge can be like a small, you know, box sitting in a, you know, department of transportation vehicle, a police car. Then we are thinking about an edge box sitting at the bottom of a cell tower. And then you are starting to expand into a co, right? So that’s really like what we are looking at as, hey. That can be small to you know, a colo. Definitely not a cloud, but how do we have a single platform that spans across multiple? And where we are doing the hardware are both those like ruggedized boxes where we also have partnered with, you know, some of the folks on smaller boxes because they are already deploying those things. Then we want to run our software stack on top of it. So that’s how we are you know, helping customers.
Akhil 26:27
Thanks, Pradeep. I wanna nerd out for a second here on fine tuning and inference, if that’s all right. I think we have a group, and I’m doing this because I think our listeners tend to be somewhat technical, but also a lot of functional experts, right? That recognize that the current models and what the applications look like cannot get to the level of fidelity for their specific use case and deal with the tail of outcomes without some of that. Inference or fine tuning towards the edge, but then comes in all the factors, cost, resiliency, et cetera. So maybe my first question here, Pradeep, is what is wrong with the current state of infrastructure to be able to support this fine tuning and inference and five years from now? What do you want that to look like with, as you said, Dan, with Armada as the edge hyperscaler?
Pradeep 27:13
Pradeep: Yeah so I’ll give you a real example. We work with one of the you know, logistics customer. They had some physical security challenges. They had some like specific worker safety rules. And they had like literally somebody sitting in front of a camera or like a monitor looking to see if they were detecting some sort of violations, right? And they were like, you know, looking at maybe trying to look at their own models, et cetera. And their long-term plan was, Hey, it’s not just developing the models. They have you know, hundreds of sites and all of these models needs to run on those edge sites, right? So this is where hey, they tried a bunch of models. They said, Hey, there’s a bunch of consulting companies who came in, but everybody tries to solve the R piece. They’re like, oh, I will create the models for you, right? And then like then the next thing they have to deal with this. What’s the accuracy of the model, right? How do you keep you know, retraining the models or fine tune the models and then are you going to enable me to deliver it end to end, right? Many companies says, oh no, I don’t do this. You ought to go talk to somebody else and then it’s your job, right? So this is a case where you know, our AI team came in, we, you know, sat with the customer, they gave us like two weeks of data. Like we had we had our base model developed on the cloud. We, you know, we fine tune it or it drained the model on that. We actually pushed it onto the edge and then you won’t believe, like the first week we were able to detect seven out of 10 scenarios. Like before it was like a one or a one out of 10 or maximum two, because when the person looks away from the screen, they don’t see it. And from there we went to seven out of 10. But what the customer loved it and said, Hey, it’s not this one time. I wanna make sure that because new scenarios will pop up. How do you do that continuous training and then make sure that hey you can push it to the edge because we know what is the capacity available on the like edge. So it is not going to be hundreds of GPUs. So that is something which we are seeing, like where people are. Like in some ways when people are developing models, they assume that there’s infinite amount of capacity, right? And now how do you take that and run at hundreds of sites with a finite capacity is like the transition and that’s where. You know, we are helping customers or even partners are starting to look at saying that, yes. Okay, if I say I have got 4G Ps on the site, right? If I take like a small suitcase a deployment, you are barely one GPU, right? How can you like run the same models you know, what you need to optimize is what we are looking at.
Maggie 29:48
Dan. I’m curious, turning back to, you know, what it’s like actually engaging with some of these customers. You know, when I think of oil and gas, national security, energy, I don’t think of the most tech forward-leaning customers. So I’m curious, you know, how much customer education have you all had to do with these customers? How forward-leaning have you found them to be and how do you get them to trust you to actually help them deploy, you know, really valuable, meaningful technology.
Dan 30:17
Yeah. The good thing is when we first start talking to customers in general, they already have a clear set of goals at the edge, and they already have a backlog of use cases that are low hanging fruit. Once we deploy the infrastructure, there is an initial bridge that you have to cross, which is you have to demonstrate the value of each one of those use cases at these sites. And then they also almost every single one of them has a very stringent cybersecurity review that you have to get across. But then once you do that, and once you deploy at one site, going from one to many is much, much easier. And the reason is there is a huge amount of value that directly ties to their objectives. And if you look at what is a common pattern to go to, back to Pip’s point on pattern each one of these. Customers, whether you’re talking about defense or you’re talking about you know, oil and gas, mining, any one of them, they have a common set of things that are happening. So one is they have very distributed remote sites where they have a lot of data being generated and they understand that the value of that data is dependent upon it being used on the site. For primarily latency reasons, but also a lot of these are critical national infrastructure. And for that and cybersecurity reasons, you actually can’t take the data off the site. So you have to do all the data processing locally. And then once you deploy, there’s a huge amount of value to be generated. And I gave the, you know, the defense example and I gave the, you know, emergency response example with Alaska. But if you think about. An industrial example, like a remote oil rig or mine refinery manufacturing facility. I could go on. Each one of those things has a lot of value to be delivered in. Number one, not sending all of the data to the cloud because one, you’re reducing risk and secondly, you’re typically, if you’re, especially if you’re running AI on that data, saving a significant amount in terms of cost. But number two. Just making sure that those sites work 24 7 as they’re intended, has a huge amount of value. Every single one of them, whether you’re talking about an oil rig or a mine, or a refinery or a factory, they have the concept of unplanned downtime and unplanned downtime costs a lot of money. Then at the same time, each one of them is moving towards automation. And so you can help them on that journey as well. And just to give you a sense, like a large energy company, we work with some of the largest in the world. They’re all moving towards fully autonomous operations over the next three years. And so immediately by deploying this, we’re able to avoid unplanned downtime. And a lot of them have. Days or weeks of unplanned downtime every year, which can cost hundreds of millions of dollars or billions of dollars at scale. So that’s an easy business case right there. And then there’s a bunch of other use cases. But then the second thing is all of them are planning to go towards full autonomy at these sites, or nearly full autonomy. Maybe there’ll be one person on these sites. Within the next few years between now and 2028. And the only way to do that is to have local compute. So we have this same pattern across all these different areas. And Akhil, to go back to your point on inference, if you take a step back, that’s why you’re seeing this super cycle shift from just the training of the models now to the fine tuning and the inference of models is that everybody sees that they have now reached A level where if you do that fine tuning on all of that data at the edge. Then you can actually move towards autonomy and the use cases there are through the roof and the values through the roof when you do that.
Akhil 34:18
Super helpful. Dan, lemme just recap for the audience here. What I’m hearing, I think coming into some of these discussions, I think a lot of folks might see this as, hey, this is, you’re really paying for resiliency. You’re paying for whatever it is. What you’re saying is it’s not resiliency, it’s efficacy. There’s a matrix associated with, some people do want that resiliency, right? In a disaggregated, disconnected environment. Others are actually like, wait a second, this is the most effective way to do fine tuning and inference. And to be able to get that last, you know, not to use a football analogy, the last, you know, third and five, the last five yards to the application efficacy. The only way to do that with all the data sovereignty, all of the. Local nuances is to do it at that edge, and you need to be able to do it in a way that is connected across multiple nodes.
Dan 35:05
Yeah, I would almost say it’s first about enabling all of these models, enabling local data processing at the edge in the way that was never possible before, which creates a huge amount of value. And then the resiliency is like a nice bonus. It’s oh, and by the way, these are a part of a really resilient. Distributed infrastructure that if something ever happens with your cloud or something ever happens, even with one of the nodes, you have really resiliency between nodes. But that’s like a bonus. The thing that gets people to say, Hey, I’m all in, is number one, there’s like a huge amount of immediate value to be generated. And then secondly, it aligns with these you know, multi-year initiatives to move towards autonomy in each one of these edge locations.
Pradeep 35:56
Yeah, and I, just to add to what Dan said, we have a mining customer, like they have around 30 mining sites, a large mining customer. They are going through this whole process right now. They have all of these IOT assets sitting in. They all push that to a cloud. It goes to some central region. They were like, Hey. This is not the way I want to do modern edge or modern, like a distributed architecture. So they’re like working with us to re architect the whole thing from a distributed you know, architecture where hey, you want all of these it devices to actually send, you know, data to be processed in the gall. And then post-process data that pushes to the cloud for, you know, you know, for their compliance reasons, backup, et cetera. And that would also be used for larger training as well, right? You get you know, you know, the what the events which are really useful rather than pushing everything to go sit in some storage account and keep paying for the storage, right? So that’s why say, hey, it helps me in also that when you do this. At the edge, you can actually take actions because you’re immediately processed within seconds. You have the output so the people on site can take action. At the same time like I’m not trying to pay, I also, in many of these cases, egress cost is high as well, right? If you think about the network, if you’re pushing all of this data, raw data, it is you know, somewhere you’re paying the network bill too, right? So that’s, those are the way they are reaping at the architecture from a distributed computer is what they’re looking at.
Dan 37:28
Yeah, this was actually a big eye-opener for us and our co-founder John, when we started Armada. You go actually talk to these mining companies just to use that example, and they have what’s called a hard segregation of it and ot it from operations. And you might ask what, why does that matter? What is the practical implication of that? Well, in mining, like in most businesses, especially industrials, time is money, right? And we work with a lot of the largest mining companies in the world. They’re constantly doing exploration for new projects, right? And a lot of times those things can take years. Well, why do they take years? They’re almost all in very remote locations. Then because of this hard segregation between IT and ot, they have to do the exploration work at the site. And then they ship this stuff in batch, like the same way they did decades ago to, you know, hq. And they analyze it and then they send it back. And we talk to a lot of companies where we ask, how long does that take you? Well, every time they do that, it takes them a month or maybe more. And so by actually bringing. Something like a galleon to the edge and deploying it behind the firewall on the site. You go from more than a month to real time, right? Which speeds up these projects, which creates a ton of value and also allows them to out-compete the competition. And so it’s really eye-opening where a lot of people just assume Hey, everybody’s doing things. You know, the kind of real time way now. Not with a lot of these types of businesses because of this hard segregation between IT and ot, and the only way to actually enable real time is to deploy behind the firewall on the site.
Akhil 39:18
Very helpful. And Dan, as you were saying, that can’t help but think about the national security parallels of, you know, your analogy between the IT and OT world, right? We’re listening and observing what has happened over the past couple weeks in the Middle East. We’re recognizing that the ability to quickly sense what is in the environment and make use of it before it gets into the large IT systems or in the process of is absolutely just as critical. And a lot of those systems too are far. I’m reminded by some of the stories Raj here at Shield Capital shared about coming back from, you know, a, a fight or flight, a fighter pilot flight, and you know, manually downloading the sensor data. Just think about the latency associated with that. If you could do that in a way that that did not require that for the immediate analysis that needed to be done.
Dan 40:12
Yeah. And Shield’s, obviously, one of our founding investors, you guys co-led our seed rounds, so thanks for all your support over the years. But I remember talking to RA about that when we started the company. Literally, it’s and that is happening globally. You cannot have a situation where you have. Very important data just sitting there and you have, you know, days or weeks or months of latency to try to do anything useful with it. You have to be able to use it in real time at the edge. And that’s the common thing we see playing out across all of these different areas, whether you’re talking about, you know, defense or you’re talking about energy, or you’re talking about you know, any other area.
Maggie 40:53
Dan, I know that you all recently released a white paper on what it’s gonna take to build US AI dominance, and we know that a lot of our, you know, potentially adversary nations like China are investing very heavily across the stack in talent, in compute, build outs, in mining, in energy. What is it gonna take for the United States to maintain our AI dominance? What are some of the sectors that we’re going to need to invest in over the next five, 10 years?
Dan 41:24
Yeah. It was funny how that played out. We didn’t plan it this way, but when we released that white paper, I was in Washington DC with some of our team and we were at the. Winning the AI race summit that the president spoke at last July, and that was the same day that the White House released its AI action plan. And we had been talking to them a lot and so I don’t think it’s coincidence, but they were very closely aligned our white paper to that. And the idea is number one, one of the biggest blockers to AI is energy, right? And if you look at how that is playing out globally, China has made very fast progress there. And that is a national security risk, right? We have to solve the energy problem. How do you do that? Number one, you use all of the energy that is available to you and I mean all of it. And one of the things that we highlighted is that in the US there’s about six gigawatts of stranded energy. And it’s not being used for ai. Why is it not being used? Well, it’s because we’ve been requiring that energy to somehow travel to these far away data centers, and that is very difficult to do. And so what we are instead saying is, let’s flip the script a little bit and let’s actually bring the data centers to the energy and enable that energy to be used right there. And a nice thing is that’s also where a lot of the data lives. And so that’s kind of our philosophy with Armada. The data centers are an enabler and they should go to where the energy and the data lives, rather than this dynamic where you’re having to have the energy and the data sail to some far away data center. That’s one of the reasons why we call the company Armada, is like each one of these things is a ship sailing to the data. And that’s why we call the data centers galleons. And then the other thing that we need to do is we need to use all different types of energy. And that includes, you know, stranded natural gas that includes things like some of these, you know, micro nuclear reactors. We think that you gotta invest in nuclear, you gotta invest in solar. You gotta, you know, be willing to look at things like data centers in space because the sun is like the ultimate source of energy. We need to use all of these different. Energy sources and then have the data centers go to those energy sources and enable ai. And we need to do it really fast ‘cause China’s moving fast. And so that’s what we’re focused on. The other thing, and this is the third pillar of the AI action plan and also in our white paper is you need to work with allies. The allies also have huge amounts of energy that is available for ai, and we can utilize that. And that also has other good impacts, like making sure that we don’t end up in a 5G situation like we did, you know, with China last time, where you’ve got other nations adopting that technology. We need to make sure that the world does run on the US AI stack, and so you kill two birds with one stone. When you work with allies, you are able to utilize all of that energy. Constantly improve these models, improve the entire stack, and then, you know, five years from now, instead of running on Chinese infrastructure, the world be running on the US AI stack.
Akhil 44:46
Yeah. And the AI race, as we think about in this geopolitical competition, is not just about the large language models, right? It is also about the infrastructure and the deployability across the globe. And I mean, this is what you’re seeing, right? Pradeep, both your prior experiences and what you’re seeing as you deploy Armada, not just. The United States, but across the globe.
Pradeep 45:08Yeah. I think, actually great question. I’ll tell you an example. In 2006 I was working on some of the telecom infrastructure stuff in Africa. What, interestingly I saw was like, you know, Huawei and some of the Chinese providers were actually setting up almost giving away free infrastructure. So they would set up all the telco infrastructure. You know, I didn’t realize that point of time, but they were actually taking sort of the backbone of the country, right? When you have all the infrastructure in your control, you have you know, the data and everything flowing through under sort of your control. I’m actually seeing similar thing. If you look at you know, the minerals industry, the mining, et cetera be like in a, if you have to build a lot of infrastructure to run AI models, you need all the, like silicons you for the silicons, you need all the minerals, right? So it’s a super interesting thing while I look at it as a LLM race, I also think about it’s a, you know, a race. For grabbing control of all the minerals and the mining sites and the oil and gas, right? So it’s very interesting to see that’s what I’m seeing. Hey you want those minerals in hand to produce more compute infrastructure to run better models, but, and also you want those to be deployed. So who has controlled all of this becomes super critical. And that’s where we have been talking about, hey, having a full American AI stack is super important from that perspective.
Akhil 46:39
Yeah, interesting. Pradeep, if you actually tie that line together, you can start with AI dominance and the geopolitical aspect there. Then you go to the infrastructure dominance and then you tie to the use cases. I mean critical minerals, right? I mean we’re talking about that right now. We’ve been talking about it. Your experiences from 2008 honestly predates some of the news around Huawei. And we’re only just now seeing or have for the past couple years the. The real challenges and limitations and the desire to have a critical mineral access point, and all of that is predicated upon knowing where they are and being faster about acquiring those resources.
Pradeep 47:18
Yeah. And I think that’s what to like really underline Dan point. It’s not just about within America, it’s about the allies too, right? You have to help and be partnering there to win the stack. Not because there are a lot of natural resources outside of us two, so we gotta be helping the allies to that.
Maggie 47:39
Last question to wrap up, what advice do you have for founders looking to build mission critical technologies for mission critical customers?
Dan 47:49
I’ll start and then Pradeep you chime in my first piece of advice is very simple. Be clear on what your mission is and write it down before you ever hire anybody. Like the first thing that Pretty, and John and I did when we started our motto before we hired a single employee was we sat in a room at Founders Fund and we said, okay, let’s write down our mission, our reason for existence. Let’s write down our vision and let’s write down our values and let’s make sure that every single thing we do from this moment forward aligns to that. And let’s make sure that everybody who joins our company understands what we’re about. And it’s a simple thing to say, but 99% of companies do not do that. So if you do that, you will stand out and you’ll get the best talent in the world because people want to work for companies that have missions. They wanna work for, especially the best people they wanna work for companies that are solving the biggest, hardest, most complex problems. And so we did that. We, we had our manifesto that we published, you know, while we were still in stealth, people were like, who’s this company? But apparently they stand for something. And then we even did a founder’s film and we did a launch film when we launched outta Stealth. And it was all about our mission. And it’s funny how. You know, those things are now a few years old, but you go back and look at ‘em. It’s all the things that we’re doing now. It’s all the things that we’re talking about in this interview. It’s just real. But being very clear about what you’re gonna do, what you stand for, and then, you know, walking the walk is really critical.
Pradeep 49:27
Yeah. Just to add, actually, as, when we were writing that thing, I remember in in one of our investors office, that time we didn’t even have an office. It’s funny that we actually took some of the hardest industries to go after, right? Because we believed that was the right focus area. And in some of the, you know, some of those industries are very hard to move because as earlier mentioned, they are legacy. So you’ve gotta believe in it and then you gotta be at it every day. So don’t do it if you’re not passionate about it. Because solving the hardest problem, you have to have the belief, passion, and you have to grind through it in many ways.
Maggie 50:07
Great. Well, Dan Pradeep, thank you both so much for your time and for coming on the podcast. We’re super excited to be investors and can’t wait to see what you all do next.
Dan 50:17
All right, thanks a lot. Great talking to y’all. See good. Thanks. Bye.
