Warning: This is a rush transcript generated automatically and may contain errors.
Asbed Bedrossian: Hello everyone and welcome to this Conversations on Groong episode. Artificial intelligence is often described as the next great industrial revolution. Governments and technology companies are spending vast sums of money on chips, data centers, electricity generation, and the infrastructure needed to train increasingly powerful models. Major technology companies are expected to spend roughly $650 billion on AI infrastructure in 2026.
Some analysts see this as an urgent geopolitical race. They argue that the development of frontier AI could become a national security challenge comparable to the Manhattan Project. Our guest today, Richard Campbell, has taken a more skeptical view. In his talk, After the AI Hype, What’s Real and What’s Next? he questions whether current expectations and valuations are running ahead of the physical and economic realities of computing infrastructure.
These questions have direct relevance for Armenia. NVIDIA says the Firebird project plans to establish a major AI computing facility in Armenia using tens of thousands of Blackwell and Rubin GPUs. Today we want to ask two broad questions. Does the global AI infrastructure boom make sense?
And if so, does Firebird make sense for Armenia? We’re honored to have Richard Campbell to have this discussion. He is something of a celebrity in the software development world, and many of you may know him from .NET Rocks, RunAs Radio, and from his talks and keynotes at IT conferences around the world.
Hovik Manucharyan: Folks, before we go to Richard, just want to ask for a quick show of support. You can do that by subscribing to this channel if you haven’t yet. For our new listeners and viewers, we cover a lot of geopolitics. This is going to be a very interesting discussion with Richard.
He knows a lot about everything, but also a lot about data centers. And specifically, we’re going to talk about various aspects of data centers and also then cover Armenia’s investment in AI data centers. So stay tuned and on with the show.
Asbed Bedrossian: Richard, welcome to the Groong Podcast.
Richard Campbell: Thanks so much for inviting me. I’m excited to be here.
Hovik Manucharyan: Honor is all ours, Richard.
Asbed Bedrossian: So we are really excited to have you on this podcast and this is your first time. So if you would actually tell our listeners a little bit about your background, what makes you tick and what keeps you awake at night.
Richard Campbell: Oh, goodness. Well, I’ve been in software since before. It was cool. My father was an electrical engineer, so I burned my fingertips on soldering irons in the single digits.
I remember him getting me to solder and desolder components when I was six or seven years old. I ran into my first microcomputer when I was 10, and that was a TRS-80 Model I with 4K RAM. And a Tiny BASIC, not a Microsoft BASIC, a Tiny BASIC that only had three error messages. What, how, and sorry, which are,
Hovik Manucharyan: you know, good error messages.
Richard Campbell: What’s object not found, but sorry. And they grabbed me. With that strong hardware background, always being in the electronics side, I was big on building my own machines. My first job after school was repairing TRS-80s at a little company called H&S Microsystems, putting in lower case kits and fixing drives and doing disk drive alignments, that kind of thing.
And so even through high school, I was writing software on the side. That was my income. And I’ve just never really done anything else for better or worse.
Hovik Manucharyan: Actually, I want to interrupt you a little bit because that’s how I also got my start in software very early. I had an Apple II Plus and I remember my drive to learn how to code. Was to build my own games. I don’t know if that was yours, but whoever I’ve talked to does that and always worries me because nowadays kids have like these, you know, very immersive games.
And it seems like there is no, you know, motivation for you to actually tinker and code. I don’t know, what was your…
Richard Campbell: Depends on the kid too. You put them in front of the Unity starter kit, like they can make a side scroller in an afternoon. Like you can do a lot now with the modern tools. I remember, when you think about that era, it was magazines like BYTE and SoftSide and the games were printed in the magazine.
And so you had to type them in. Which was a heck of an education in debugging, you know, to actually get all those lines correct and get the game to run. It was many days sometimes to type them all in and get them right.
Hovik Manucharyan: Yeah. So then you went on to podcasting, right?
Richard Campbell: Oh, yeah. There’s a couple other steps along the way, but sure. You know, I was a happy IT guy, you know, running a team of developers and writing code myself for a few different companies and And I fell into writing magazine articles because it was an opportunity to get tax breaks. At the time, the Canadian government, I think they still do this, if you could show you were doing original work, they’d give you a tax credit for a certain amount of wages.
And one of the proof points was, are you published in quote unquote trade journals? And so I sort of took up the mantle of I’m going to do this writing because it’ll expand my budget, but I can get those tax credits. Maybe I can hire another guy. We can get more of the stuff done, that kind of thing.
And I fell in love with the writing part, like telling the story. Of technology to me just became incredibly fascinating. And that leads to conference speaking and other lectures. And, and then I bumped into Carl Franklin at a conference once back in 2004.
It’s a bit, you know, historical now, but yeah, we were both in a show in, in, in Montreal. And I knew who he was. He had no idea who I was, but I knew he was because he had one of the first Visual Basic websites on the planet, which I had referred to as I’d moved into VB. And so we started talking and became friends.
Sometimes you meet a brother. You know what I mean? Like it didn’t take long. It was very obvious we should hang out more.
And he had he’d had other co-hosts before me, but he asked me if I would co-host with him and show 100. And previous to that, Mark had done 50 shows and Rory had done 50 shows. So I thought, I’ll do 50 shows with them. I mean, how hard could it be?
And we’re at 2017 now, so missed it by a little bit. But that’s, you know, the very early days of podcasting too, right? Where really it was only the geeks that could make it work because you largely had to write your own catchers and things like, and you were burning to CD-Rs. This is before the iPod or smartphones as we know them today.
In fact, the first time we did a conference as the .NET Rocks guys, people brought their favorite CD-Rs with them, shows they listened to over and over again, and we were signing them, which is a little weird, but you do the thing. And so, yeah. And meantime, I also decided I needed an IT podcast as well. So I started RunAs Radio.
And in just the past few years, I signed on with TWiT for Windows Weekly as well. Because three shows a week just scratches that itch, right?
Hovik Manucharyan: Yeah. I mean, .NET Rocks is a cool podcast, but I really love you. You took an interesting turn there with your Geek-outs, and you established essentially, I think, a format of storytelling. I love the way you tell these stories about nuclear power, about AI and data centers.
So that’s how I got reacquainted with you when you gave a very good talk at NDC, both on AI and whether it’s a hype and also building data centers in space. So I think we will link to those talks for our viewers. Highly recommend that you guys watch that. But was it easy for you to change topics?
And did you keep your audience as well while doing that?
Richard Campbell: Well, the Geek Outs were Carl’s idea, if you can believe it. But that’s not surprising. Carl’s got a vision. You know, he’s really great at what people like and that sort of thing.
When we were making a lot of .NET Rocks episodes, there was a period where there was so much sponsorship. We were making three shows a week. And they were paying well enough that that became like most of the job. Like I’ve always been involved in conferences and consulting and things like that as well.
But in that sort of halcyon days in the 2010s, we were just making a ton of shows. And it was in 2011 when the last space shuttle, when Atlantis was landing for the last time. And I went on a bit of rant about the problems with the shuttle and how frustrating it was, the potential it hadn’t in some ways never achieved while still doing great things. And Carl said, you know, we should record that.
And I’m like, for what? He says, no, I think people would really like to hear it. It’s like, no, it’s not software development. Anyway, I was wrong.
Because the reaction when we published that first one, even before we called it a geek out, was people just asking for more. What about this? And what about that? And that’s usually the kind of feedback you want is to go further into that.
So the geek outs became a monthly routine. And these were all subjects I was always interested in. As my career expanded, And I worked as much on the IT side as the software development side and started working in other countries, got a taste for geopolitics. It’s like, why does this place work the way it does?
And how can you be successful in those different locations? All of those mechanisms become important. And I’m prone to keeping notes. I’m always thinking about how we’re going to do the next thing.
And so the geek outs are actually from piles and piles of notes. And in some ways, making those shows It was about forcing me to finish thinking about something, at least for a point in time. So you’re actually going to tell the story of how nuclear power works. You’ve got to organize that hour properly and try and lay it out for folks who want to understand.
And I avoid shying away from the complexity of things. I believe in the intelligence of others and the fact that, you know, the reason these problems are unsolved is not because people are stupid, but because they’re hard. It’s not, they’re not simple problems. And I think most people value understanding the difficulties.
You can see where, why things struggle where they do. And then, and they’re for a variety of reasons and you have to respect those reasons. Uh, the, the term I fell in love with was ontological humility. Just a recognition that there’s been smart people in this world for centuries, for millennia, that made the best decisions they could with the information they had at the time.
And if you understand those, then you can maybe move the ball forward a bit. You know, we’re all trying to make some progress. And if you can teach someone to understand something a bit more deeply, maybe they can make a difference with it.
Asbed Bedrossian: That resonates with me as well, because time and again, we’ve seen how intelligent the listeners are, actually. There’s so much information in each person that when we’re talking with them, they’re at least as knowledgeable as we are, it seems.
Richard Campbell: Sure. Yeah, and probably coming from a different point of view, too, right?
Asbed Bedrossian: Oh yeah, definitely. Definitely.
Hovik Manucharyan: All right, well, let’s jump into our topics, which are not few today. So we hope to keep you all. There’s going to be a lot. We’re going to go start with just talking about AI in general and the issues around sovereignty, security around using AI, as well as then we’ll jump into Armenia and Armenia’s potential use and plans of using AI.
So we’re going to be talking about the AI data center later on in the show.
Hovik Manucharyan: Of course, I’ve been keeping track of the debate around AI, but one interview I think I caught two years ago really jolted me in how honest, but also Scary it was. It was an interview by Leopold Aschenbrenner, who essentially said that, you know, yeah, we’re going to achieve super intelligence. We’re going to and it’s better that the US does it than the Chinese CCP. And that term was kept repeated on and on.
But essentially, it doesn’t matter what kind of government you have. I don’t want to put words in his mouth, but it seems like no loss was too big in order just to sort of beat China in this race. Sure. Now, Aschenbrenner…
Richard Campbell: The ends justify the means kinds of arguments. It doesn’t matter who dies along the way as long as we get to this goal.
Hovik Manucharyan: Yeah, yeah. Aschenbrenner and the people and the viewpoint that he represents argue that scaling AI would lead towards super intelligence. And there is a huge race right now between US-China over chips, models, power and industrial capacity. And his argument essentially rests on the In part on the construction of computing clusters that could eventually consume extraordinary amounts of electricity, as if they’re not today.
You, Richard, emphasize the physical limits of such projections, and you say that such projections could overlook Serious issues like electricity generation, transmission, cooling, and practical problems of operating very large computing facilities. So tell us your take on this. When you look at the AI race today, which picture is closer to reality? A technological competition comparable to the Manhattan Project that we must win at all costs, or an investment bubble that is built around expectations that may never materialize?
Richard Campbell: And to be clear, even the Manhattan Project was founded on questionable grounds, right? The Germans never did take building an atomic bomb seriously. Nobody did. Uh, and the, now the, the expenditure paid off, you know, to some degree and, and created a new world and arguably, you know, we can, we can go back to the relative merits of the mutually assured destruction model, uh, and, and the consequences of all of that, but that’s a little off topic.
Uh, The folks that are not financially incented to push AGI largely say, I don’t see a path to AGI, that these are much more complicated ideas. And the Chinese certainly aren’t speaking that way either. Like this is a perceived race of the West. China does what China does for the most part.
And while they’re certainly interested in the space and they see value in the technology and are utilizing it in ways that perhaps in the West, we wouldn’t find anathema, their whole citizenship scoring system and continuous surveillance. Not that we’re surveillance free in the West or much less respecting privacy, You know, those aren’t that unique. You can debate their innovations, but you can’t debate their cost effectiveness, that they are very efficient. And part of that is the constraints that are being put on them.
And part of that is that they, you know, that’s the angle they can take advantage of. But I always see this AGI argument as this escape route. It’s the excuse for all things. We don’t have to worry about the harm we’re doing because as soon as the AGI reveals, it’ll refix everything.
Like that’s very magical thinking. And we’ve seen forms of this for many years in different tech. It’s not a good strategy. Better to deal with the reality you have in front of you.
Like this is a profound technology, you know, at the simplest level. We are already as a society contending with far more data than we can actually manage. And often the information we need exists in that data. We just can’t find it in a useful form.
You know, we struggle to organize information all of the time. For better or worse, these neural nets have a remarkable ability to synthesize that data together. Now it’s a bit of a lossy compression strategy. It’s like taking a whole lot of raw images that are massive and unwieldy and JPEG-ing them.
You can still see the images, but you’ve lost some data along the way. And in many ways, these large language models do the same thing with language, where they make the information far more accessible, but also a bit inaccurate. And so there’s this, you know, you need to fact check, but first you’ve got to press against the idea that there are facts that are important out there you want to get to. And so the two things go together.
And I think in software development, especially because of our practices around How we build software, the strategies of testing, building specifications, and so forth. We’ve had more success with LLMs than most other fields, just because I think it works well with our procedures. And you’re seeing really, really skilled developers today starting to utilize those tools more and more effectively over time. So we can see value in the tech, but none of this leads to a super intelligence.
So this idea of racing for a sort of mythical goal, it’s unwise. And meantime, there’s so much value around otherwise to take advantage of. And a part of that reason, like why do we have this problem? Well, part of it is that So much money has been invested and so much more money is available to invest, although it appears to be running down now, that you have to keep making bigger promises for your investors.
That’s the only way you get the next round and the next round and the next round. And so in some ways, the pattern of investment forces this behavior that you have to find ways to keep projecting the line upward. And that means making crazier promises. And after Three or four years of these promises, now the promises are so outrageous that most folks are looking and going, that’s not going to happen.
And you’re bumping up against physical limitations like you can’t make that many chips in this amount of time and you can’t generate that much electricity to keep that line going in that direction. And so we’re hitting very reasonable thresholds now of what’s possible in the next few years.
Hovik Manucharyan: Now, there’s this talk about, of course, frontier AI, which means essentially the latest generation of technology and LLM specifically. But, you know, we’re looking at projections where a single data center could require gigawatts of electricity. Where do you think the bottleneck will be in the future? Will it be in power generation, cooling, transmission, construction, or the chips themselves in your opinion?
Richard Campbell: Well, it’s all of the above, isn’t it? Ultimately, right? Every one of those resources is actually constrained. Much of this has to do with duplication of effort that you have many competing companies.
One of the points I made in the keynote you referenced was, That companies like Amazon and Microsoft and Google are looking at this hype cycle, this event, as an opportunity to cement their leads as the world’s hyperscalers. You know, I look at the same effect that happened with cellular telephones a few decades ago where the few big telcos that jumped on, they got as many cell towers out there as they could so that today, In most developed parts of the world, you just couldn’t put another shell tower in if you wanted to. It’s not feasible. And so the incumbents have locked the market.
And it sure looks to me, when you look at where data centers are being proposed, that a lot of them have to do with getting all the best locations locked up so that there will never be a new competitor in the future. And that’s good long-term thinking. Knowing the hype cycle will end, knowing that AI will eventually just be a normal product, just another workload in a data center, How do you make sure you have the data center space?
Hovik Manucharyan: Are you worried? I know you spoke against this competition at all costs, but China is rapidly expanding its electricity generation and industrial capacity. Are you worried that that may be a structural advantage that China has in the future? We’re not seeing the US or the West in general rapidly scale up their power generation capacities.
Richard Campbell: No, the West, because of the way that its bureaucracy works, does this at a much more gradual rate. The upside to being essentially an autocracy in China is that when Xi says do it, people do it. Now, China’s dealing with serious issues of its own. You know, its demographics are in a very serious state.
They don’t have any immigration. They mostly are losing population and their population is rapidly aging. So the original goal, if you go back to Deng and those folks that started this motion in the eighties for China was about trying to get to a consumer economy, that the primary consumer of Chinese goods would be China. That hasn’t happened.
And in the meantime, Policies like the one child policy and the general disruption in lifestyle has meant there are very few children, far fewer than the Chinese government realized. There was a set of reports out of Shanghai that finally admitted that there’s 200 million people missing, essentially, that didn’t exist and abruptly their demographics aged dramatically. And so they’re dealing with that crisis. China’s not building power plants because they want to compete with the West as much as they’re building power plants because they need them.
That’s actually what they require. I have been in Beijing when the AQI was over a thousand because the coal smoke was being held in the city. There was a strong push to move away from those highly polluting power plants into cleaner ones, whatever is possible. They have driven down the price of solar power for everyone in the world, effectively.
One of the things that happens when you have Government-led industry is that overproduction is a very likely behavior. And so they overproduced solar panels and then essentially effectively dumped them on the market. As beneficial as that may be, it drove down solar prices all over the world and made it hard for anyone else in the world to make solar panels at a reasonable price. They’re doing much the same thing with batteries.
And again, I wouldn’t even speak to this as a strategy so much as it is a need for China and their top-down directed industries. They tend to have this happen to them. It’s hard to, you know, make the market move precisely. And so when you talk about the West moving slowly, that capitalist-driven model naturally leads to more caution.
You don’t want to be, you know, if you overproduce in the US, you take your own market and end up bankrupt. In China, the government keeps you going, just sells it at a lower price and keeps your people employed. So, you know, There’s distinct differences in those two different systems and the advantages they have. We’re benefiting from much of China’s overproduction in many respects, while also having to deal with the much more pointed concerns around dependency on the Chinese industry.
Asbed Bedrossian: Richard, I’m kind of interested in knowing if the economics of all this AI build-out can work out. Most people probably think that you buy a computer and you plug it in the wall And it goes. Everything’s fine. But the financial challenge is not simply the initial cost of construction.
In my career, we had the capex when you make a multi-million dollar purchase. And then you, of course, have the opex. The GPUs are going to depreciate just like Like CPUs, just like your laptop, everything eventually gets old. New hardware can make previous generations less competitive.
These pizza boxes have a lifespan, you know, three, four, five, whatever. They’re engineered and there’s electricity, there’s cooling, there’s networking, maintenance. So there’s constant operational costs. And critics have generally questioned if these projects can actually produce enough revenue to justify all of this CapEx and OpEx put together.
Many of them call the current build out an insanity bubble. What’s your take on this? What does a genuinely profitable AI data center look like to you?
Richard Campbell: I mean, we don’t know right now. And to be clear, there are no AI data centers. There are just data centers. The workload may be AI at the moment.
And what would be the distinction? Well, it’s more power per square meter. But that has been going on long before we started tagging AI into the name too. Racks for many years were in the 10, 20, 30 kilowatt range.
Even before the big NVIDIA chip shows up, they were pushing 100. An MVL72 is 140. They’re talking about 300 per rack. At that day, I dealt with data centers back in the virtualization era in the early 2000s where we had rooms that were three quarters empty because we couldn’t get enough electricity into them now.
The racks themselves had densified so much and consumed so much power per rack. You didn’t need that much space. So, you know, in the end, a data center is actually a large concrete box with a whole lot of electricity pulled into it, some kind of cooling methodologies, and we can talk about the range of them, a lot of connectivity. And then you, today, largely just bring in racks on pallets.
You don’t even take the pallet away. You put it down on the pallet, you plug it in, you run it for five years, and then a forklift comes back and takes it back out again. And it gets refurbished and repurposed through other things, and new racks come back in. And so the asset is that power, cooling, connectivity, and some structure around it, and some infrastructure to protect it.
Everything else is transient.
Asbed Bedrossian: So you have in the past described AI infrastructure as the ship of Theseus because the GPUs, networking equipment, all this stuff that costs money, they’re continually replaced. Does a so-called AI factory, and I understand that’s just a data center full of computers, ever reach a steady state of profitability or is there continuous capital replacement simply part of the business?
Richard Campbell: It’s always going to be replaced, and there’s always going to be a consumable. There’s always going to be new hardware. But you’re hinting towards the lines that we’re trying to merge now. The same thing happened in the dot-com boom.
When is a website actually profitable? It costs a certain amount to operate, and it provides certain services that have an overhead, that have a cost of goods associated with them. What’s the right price? And when a new market emerges, we often, as we certainly have this time, lowball the prices, provide unlimited service because we’re trying to get customers engaged.
And we’re also trying to see how things get used. I think an awful lot of the creators of these tools did not Understand how they would actually be used. And so it took time just watching how the customers use these tools to get feedback. I was talking to a friend, I had lunch with him yesterday where he was talking about the fact that Anthropic asks to have access to his sessions with Claude because he’s pushing the limits of what that tech can do.
And they really want that data. They’re making him an offer essentially to use his data as part of their understanding of how their products are used. So we’re in this race right now where the price is going up, the capability is lagging behind on that, and we’re finding the intersection point. It was really easy to use LLMs for all kinds of work when it was flat rate $200 a month.
But as you start getting into per token prices, the return on investment has to shape up and that’s going to create a balance between the cost and the consumption. Consumption is likely to decrease as the costs come into play and people start to actually have to assess, is this worth it? The optimization strategies, what workloads make sense, what ones don’t. The localization strategies where now I can buy a Spark and run that particular workload locally and just keep my own room rather than go off to a frontier model for it.
I think we’re still feeling around for what that’s going to look like. And that is just not that different than what came before when we moved off of mainframes into PCs, but the mainframes never went away and we rebalanced the workload a couple of times and the PCs got more capable and we create clusters of them and networks of them and they took over other workloads, but still the mainframes persisted. I think we’ll never get fully off the cloud with AI. But we’re never going to be all on the cloud with AI either.
It doesn’t make sense to do a one size fits all here when the real prices come to play.
Hovik Manucharyan: What is the point at which you think the prices will reach an equilibrium?
Richard Campbell: Yeah, I think we’ll see it and it’ll be a mix and it’ll be complicated. It’ll depend on the project. You know, the other side of this when you get into the LLM theory is that as the scope of the problem becomes more defined, the size of the model can shrink. So certainly in the software realms, We’re talking about things like you use frontier models for your first two versions, but as the software architecture is well defined and understood and the feature set of the software is better defined, you can move into a smaller perhaps local model.
I even talked to a group where they were talking about we’re going to end up generating a model for each application for the maintenance of that application.
Asbed Bedrossian: Richard, I may be taking a little bit of a detour here from our main topic, but when we hear about Chinese models that have cost apparently a lot less in terms of much weaker GPUs compared to the frontier models that we have here in the United States, and they say that what we get from the Chinese are generally about six months, seven months, eight months behind a frontier model in the United States. Are those going to ever come in and become financially more viable even for U.S. companies to use because they are a lot cheaper to use somehow? Well, I think we’re seeing that happen already.
Richard Campbell: I mean, that’s happening right now. As the costs are coming in, folks are now looking at alternatives and saying, is this good enough? You know, do I have to keep, you know, you don’t always buy the i9, sometimes you buy the i5, right? You’ve got to decide on the workload.
And so as soon as we start to talk about real costs, now all the customers come to play. And as is usual, Privacy takes it back, seek to cost. So, you know, you can debate, am I willing to run this model off of Chinese servers? It’s like, well, for that price, sure.
But also, when you start talking about running local models, those models are very efficient and run extremely well on Western hardware. So, yeah, I think all of these things come to play. It’s not the first time in many different product lines where another part of the world builds a lower cost approach to it and folks have to consider it.
Hovik Manucharyan: Richard, let’s change the topic a little bit. There is this big debate about having AI sovereignty. I think it’s associated with the debate around AGI and superintelligence, but in general, I think that this is a capacity that already has implications for your critical infrastructure, for your national security. And then there’s a distinction between Building frontier AI and also spreading AI throughout the economy.
There are some economies, I think Singapore is one, that have decided, okay, we’re not going to join the race to build the next biggest chip. We’re going to try to have specialized applications that we excel at, like localized applications of AI. So the debate is like a media CEO, for instance, Jensen Huang, says that countries should not outsource their intelligence. It helps his company’s bottom line when he does that.
Richard Campbell: He’s very willing to sell you that too.
Hovik Manucharyan: But at the same time, these foreign policy think tanks associated with governments normally, like Western governments, say that you should hold on to this AI capability that includes talent, it includes the capital infrastructure, and you should control the ability to diffuse AI throughout the economy, and especially who can use these technologies. So for a small country, We can take Armenia as an example, but in general, what does technological sovereignty and AI sovereignty actually require?
Does it require owning advanced GPUs and data centers, or can sovereignty come from control over data, software, digital identity, and local language models and applications, which are going to be more needed anyway?
Richard Campbell: I think you have to think about the crisis moments. Like, let’s be clear. I don’t think there’s any debate about data sovereignty, first and foremost, right? Like, yes, a country should protect its people’s data.
And we should all be doing a better job of that universally. Let’s draw a corollary. And I’m going to use Canada as an example. When the COVID crisis hit, everybody needed vaccines.
There were new vaccines coming. And Canada had shut down. Their last vaccine facility in Quebec many, many years before, because they simply bought the vaccines from the US. They produce it on a much larger scale.
It was far more cost effective. And so they didn’t have any. And unexpectedly, because this had never happened before, at least not since we’ve had vaccines, the Americans simply had none to provide. And so Canada bought their first vaccines from the UK.
And so there’s now an initiative in Canada to have one expensive Vaccine facility. That’s not just a building and a factory where you can push a button and say, make a vaccine. It’s also a set of skilled people that operate that facility, that understand how to make those vaccines. Like that’s a national asset.
Every country should probably be able to make their own vaccines. Again, they won’t be as cheap as you probably make somewhere else. And you can look at compute the same way. It’s in Armenia’s best interest to have some amount of compute, not because the compute is the important part.
It’s the ecosystem that grows around the compute. It’s the educated people. It’s the access to resource. And when the crises come, and let’s face it, Armenia knows more about crises than a lot of other places in the world.
When the crises come, you have a resource that you may need at that particular moment. So take advantage of the ecosystem first and foremost, but also recognize strategic assets where they exist. You take care of your own water, you take care of your own food, you take care of your own medicines, take care of your computer.
Hovik Manucharyan: So we, months ago, we had an Australian professor, Warwick Powell, with us, and he’s a proponent of, I guess, the thermodynamic theory of AI competition, where the The real competition is more about how efficiently countries convert energy into useful computational output, not just having the latest chip. And he argues that China does a better job at it. What do you think about this? Could efficiency ultimately matter more than simply building the largest possible clusters?
Richard Campbell: Well, it always makes sense to be efficient, recognizing that efficiency takes time, right? So often we’re, and you certainly have seen this in the first few years of this AI wave, speed took priority over efficiencies. The models were more power hungry than they need to be, less effective than they needed to be, to be quick to market. And gradually the efficiency equation has come in.
As real costs come to play, efficiency matters far more. Because the Chinese were constrained in other ways, not the least of which coming from behind, they played the efficiency card first and worked on it. Starting from the very first public DeepSeek model, the focus was on efficiency for them because of their constraints. And it did paint a picture of the reality that these tools can be more efficient.
I don’t like extremes in any scenario, right? Like the world is more than compute, right? There’s a whole lot of other things we also need to do, but using resources efficiently is always a good idea because it provides more and more benefit over time. You know, that might be a minor improvement efficiency, a 1% improvement efficiency today.
As those resources grow, they’re only going to become more relevant. So you, but you can’t only do efficiency, you must also build, but you’ve always got to pay attention to both. I just think that the arguments are more complicated than that. And, you know, getting your feet wet, getting started is an important part, which, I mean, Armenia, that’s already happened for them, right?
The first Firebird facility is up and running and it’s small, but it’s important. You think about the value of an end-to-end moment. That you can actually have done all of the things to have compute actually up and running. That means you were able to build, you were able to import the resources that were needed, you had the skilled people to do the initial work, you’ve got things up and running.
It’s no different than the first transaction in a business. Or the first set of solar panels that you install. It’s not the whole rig, but it is an end-to-end test. You made power and you distributed successfully.
You’ve got computing country now in the form of these NVIDIA chipsets, Blackwells. You’ve proven it’s possible. That tells people a lot once that’s up and running. That’s a major milestone.
Asbed Bedrossian: Given what you just said, that runs into some of our next questions, actually, because indeed 70,000 Blackwells are being deployed by Firebird. Something on the order of 300 megawatt expansion plans exist all the way to 2 gigawatts. Based on everything you said, Does hosting a frontier scale AI facility make sense for Armenia in a strategic sense? Is it too big, too small?
Is it just right?
Richard Campbell: Well, if you can’t power it, it’s not going to get that big anyway. I think the, you know, like many things, you can get to build this incrementally. I mean, just this past week, it’s 15 megawatts, right? Is the current rig that’s up and running now, which is not trivial.
But it is well within current power consumption. Now, if I get the names right, please help me with this. It’s Hrazdan. So there’s a huge Soviet-era thermal power plant there, fuel, oil, and natural gas, plus you have one combined cycle gas generator, the Hrazdan 5.
Which is 400 plus megawatts by itself. So 15 megawatts, not a big deal at this point. But this is still, you know, it’s good to think in terms of a data center as simply industry. Think of it as an aluminum smelter.
Consumes a huge amount of electricity to make a product that can be sold. And so do you turn off the lights to run that thing? Probably not. Like that’s not a good idea.
People get very angry very quickly and it’s not, you know, you can’t make money off of that at the expense of everyone else. So you’re going to play a balancing act here. And in most parts of the world, and I’m, you know, the talk to follow the AI hype cycle and data centers in space is building data centers that don’t suck. So I’m deeply immersed in this research right now and talking to the folks that are building them.
The norm for most contemporary scale data centers, and 300 megawatts is contemporary, is that you come into this with a behind the meter mindset, AKA you are also responsible for building the power. So the timelines are a challenge. Yeah. Cooling has been made into a bigger deal than it really needs to be, right?
Like at the simplest level, and to be clear, like cooling is important. For every megawatt of energy you put into a data center, you get a megawatt of heat. That’s right.
Asbed Bedrossian: And you have to dissipate that.
Richard Campbell: It has to go somewhere. And, you know, the focus on water has been largely because the data centers have learned that air conditioning is expensive, that using evaporative cooling is far less expensive. The challenge there is to put regulatory rules around using water that make it appropriate, doing as much capture, limiting the amount of water consumed, getting away from potable consumption entirely, getting the price of using a nation’s water to the point where they have to debate the ROI reasonably. As long as you make it the cheap alternative, they’re going to use the cheap alternative, right?
These are profits seeking structures. And so whatever reduces cost is what they’re going to be interested in. It is a people’s responsibility to say you have to do this responsibly.
Asbed Bedrossian: Richard, since we’re talking about the cooling and the water, a lot of people are under the impression that AI consumes water. We don’t destroy water. Water runs through the systems. And in my view, we heat it up.
And if we’re going to start heating it up with hundreds of megawatts of power, then that’s going to accelerate. Basically, that’s going to accelerate global warming, isn’t it, on a global scale?
Richard Campbell: Yeah. The impact of the water temperature increase is not as large as, say, the emissions of carbon dioxide from those natural gas power plants. Uh, in, in terms of overall heating potential. And if you look at the consumption, you know, and you’re, you’re absolutely right.
Asbed, water is not destroyed in the process ever, right? Whether we consume it or we blow it, you know, in the end, this is, this is evaporative cooling techniques where we’re just pumping water through and carrying the heat out and radiating it off. And we tend to, when given an opportunity, we use an open loop approach because it costs more to cool it back down. When you look at a nuclear power plant, they have those big cooling towers because you need such pure, clean water to run the cooling loops, it’s more cost-effective to cool that clean water yourself than pump it back in and use it again.
We’ve not put those requirements onto data centers to recirculate water more thoroughly, to make the cost of using new water higher. If we do, then the equation changes and they’re more interested in managing that water more effectively or using other strategies. Water is not essential to the process. It’s just cost effective.
We could be doing cooling loops, closed cooling loops with ammonia. It’s just more expensive. So always this balancing act of what’s cost effective versus what’s appropriate for the environment, what’s safe and reasonable. We have choices there.
That being said, if you deplete the water table to the point where people are struggling for potable water, nobody’s going to like your data center very much. Like that creates issues. Uh, and that’s where you get into this bigger point of like, we tend to use all water the same and, and drinking water is different from simply recyclable water. It doesn’t have to be drinking water, cooling data centers.
Hovik Manucharyan: Right.
Richard Campbell: It can be, you know, other water sources.
Hovik Manucharyan: Yeah, I think that is my worry in that I have seen a tendency of large data center companies building data centers outside the US primarily because of regulatory concerns in the US where now States and local residents are fighting back for their water. My worry is that if a data center in Armenia is built out, will the companies that are involved have an easier time reaching palms, if I may say so, or compelling the government to look the other way or try to reduce the regulations? That is, I think, a serious worry.
Richard Campbell: I think you’re right on there, Hovik. These are the issues. There’s a lack of knowledge when they go into smaller municipalities that simply don’t know what the issues are, and there aren’t large-scale federal regulations, and that’s true everywhere in the world. And then also, you deal with a certain amount of corruption .
No place is immune. The challenge is the sufficient checks and balances, and largely revealing the truth. There are overreactions here. If you’re that concerned about water, why are there golf courses?
They consume a huge amount of water. And I point to a golf course not because I dislike golf, but because that is an optional use of water also. Any more than lawns are a foolish thing. Why do we pour water into green stuff in the front of our houses?
You can go more practical on that. Responsible use of water should be something that is regulated well everywhere because it takes time to make more potable water. The water will be back. The question is, what will you drink in the meantime?
Managing this is important and one of the upsides to making this much noise is to simply help the officials that exist to be more aware of the concerns and to make sure we have reasonable regulations in place. The overreaction is a problem too, where you have entire states now saying total freeze on building data centers. Like that creates another set of problems. So I think when you take a balanced mindset to this, you’re gonna be okay, right?
Because these companies are reasonable. They are there to make a profit, but they also know that destroying where they are doesn’t help them either. So picking that balance and encouraging that balance, nobody wants to hurt people. They just want to be successful.
Hovik Manucharyan: Now, earlier you mentioned the industry forming potential for having a data center and supporters of this data center. I still remain on the But supporters of building this huge data center in Armenia argue that researchers and companies will benefit from having advanced compute physically located in Armenia. But to my understanding, they will still need to access that capacity On viable commercial terms. If I start up a company in Armenia, I can just use AWS or any one of the alternatives as a data center and use compute resources.
So I just like, I don’t buy that argument that much, but what is the real economic advantage that physical proximity actually provides?
Richard Campbell: I would say the Silicon Valley didn’t become the Silicon Valley because you gave free semiconductors to the residents. I think the real power here is having a data center in Armenia, a major data center like this does attract a certain amount of attention. Now, hopefully the universities are working with them and that perhaps they have students that are part of the operations teams, that there is some trade mechanism going on to educate more folks to be a part of this data center like the construction teams. Are one thing and they’re larger and construction is going to go on for many years.
These data centers are built up incrementally, but the operations teams are smaller. You would just hope that we have a wise set of folks working with those teams to make sure we are educating more people and that there is an incentive of some form to utilize the data center in situ in country. But everyone’s looking for a catalytic moment, a moment that creates all of the other things around it. And my understanding of Armenian startup culture is like, you’ve already figured this out, that what makes startups work is not just smart people, but also a financing strategy, an education strategy, a regulatory strategy.
It takes players in finance, in education, in government and so forth to make all those things work. That same pattern just gets amplified when these data centers are there as well. And so you may or may not run the compute off that data center. That wasn’t the main reason.
But that anchor, that sense that we are part of this system, it’s not just happening elsewhere. That’s a powerful force. And that encourages folks to do more there. There are more Armenians, I don’t think I need to tell you this, but you know this, there are more Armenians living abroad than living in the country.
And by a lot, like double? I’m excited to see Armenia on the map, an in-the-world conversation, part of this story, because there’s so much more to be done.
Asbed Bedrossian: Richard,
Asbed Bedrossian: I’m still thinking like an IT nerd, I guess, since we’ve installed a lot of computers in our careers. And your high-performance computer is always as good as its weakest link, its next bottleneck. And it’s either going to be your backplane or your back interconnect or the connection to the rest of the world. And when it comes to that, Armenia is landlocked and depends on terrestrial fiber links through neighboring countries.
And given the importance of these physical fiber networks, I think you have had talks about these issues on the global undersea cable systems.
Richard Campbell: Yes.
Asbed Bedrossian: Could connectivity become a bottleneck? Even if… Connectivity is always the bottleneck, as I said. …World class. That’s right.
Richard Campbell: Yeah. Look, I’ve spent time reading, understanding Armenia better. You get your gas largely from Russia, but you also have a pipeline to Iran. Why?
That’s right. Because sole source is a mistake. It’s always a mistake. So you always have an option.
It’s smart. It’s what smart places do. Like this is a crossroads nation, right? You can call it landlocked.
I think crossroads is a fairer term. Everything can pass through here. You’ve got a bunch of neighbors in varying states of relationships. You should connect to them all.
The more fiber you lay, the better off you are, because then you provide choices for yourself, but you also provide choices for others. The power of being part of this data center world is that you create a sense of trade. Your neighbor benefits from you and you benefit from your neighbor. I think we can all agree that peace is better than war.
And trading partners tend to fight each other less. And one of the trades you can make is data. So you sit in a location where you can easily run a cable to Iran and to Turkey and to Georgia and to Azerbaijan. The fact that all of those connections can be made and you just create a point where the whole system becomes a bit more resilient.
Everyone can communicate with everyone else. It’s just, it’s another asset. It’s an important one.
Hovik Manucharyan: I want to go back to the power issue, because just for our listeners to understand, Armenia has currently a small surplus of energy that, for instance, in 2023, that was the latest number I could find, Armenia’s surplus was around 178 megawatts. But I think that the announcement for this Firebird AI was that they’re going to have 300 megawatt power by 2027, next year. That’s more than two thirds of the nuclear power plant, and that’s double the surplus that we have.
Richard Campbell: And let’s be clear, Metsamor is an old power plant, right? Unit 2 is from 1980. And it may have spent a few years turned off after the earthquake, but it’s got 40 plus years on it. I think the current licensing is extended to 2036.
Hovik Manucharyan: Well, there’s a debate on whether, and it’s related to geopolitics, so it has the option of being extended to 2036, but due to Armenian geopolitical course, I think that now is in question, and now like Russian, Rosatom is saying that 2031, And if you want Russian, if you want to go until 2036, I think the deal is that you have to employ majority of, I guess, there’s this technicality where most of your contractors have to be from Rosatom. So they have to actually be the majority stakeholder in the renovation. And I think that’s just a, I think, precast for the geopolitical issues related.
Asbed Bedrossian: It’s not a technical issue. It’s a political issue.
Richard Campbell: And believe me, the two go side by side all the time, right? Like, you know, that’s kind of unavoidable.
Hovik Manucharyan: And so, yeah, I mean, the bigger backdrop, geopolitical backdrop is Armenia wants to join the EU, the Armenian leadership. There are a lot of questions around this, but I think that’s the whole play. If Armenia leaves the Russian orbit completely, then the lifetime of the nuclear power plant will get cut down to 2031.
Richard Campbell: And don’t bet on that. Rosatom makes deals on nuclear power all over the world, even during these embargo times. They’re not just supplying fuel to Armenia, they also supply it to Turkey, who is a NATO member. They also are involved in the fuel cycle for France.
Believe me, I think those deals can be made. But you’re right, the reactor’s old, and even if it’s 2036, it takes 10 years to build a reactor. Like, that’s part of the problem.
Hovik Manucharyan: So what are your thoughts about that? Because there’s also now this, also, I think, geopolitical, where the United States is saying, hey, Armenia, we can sell you these small modular reactors. They cost a lot, and they may be great for US market. But in terms of the cost efficiency for Armenia’s market, you know, I personally have not, I’m not convinced that that is the best technology.
But so there’s a choice whether to build out a completely new large reactor, which Russia is now offering Armenia. They’re saying, you know, we can help you build a completely new reactor that’s gigawatt plus.
Richard Campbell: Yep.
Hovik Manucharyan: The catch is that you’re going to be married to us geopolitically for 30 plus years or 50 years or however long that- 50, 60. Yeah. Yes. Or now, like, I mean, Armenia is, Armenian leadership saying, oh, well, we’ll just go with small modular reactors.
What is the cost efficiency for that? But in general, what are your thoughts about small modular reactors? I mean, I know that they’re probably a little more safe, but also that hasn’t been proven.
Richard Campbell: The biggest problem you have with small modular reactors, you cannot tour one right now because there are none. So you would be part of that experiment. Now, all contemporary reactor designs, and I would even put the Russian reactors in this. These Gen 3 Plus reactors are very safe.
They are the passive cool-down style reactors. In the West, there’s really two that are fully mature and built. You can walk around them today, and there are teams that know how to build them. You have the Westinghouse AP1000.
This is the reactors that just finished in Georgia the past couple of years, Vogtle 3 and 4, which for many political reasons took far longer and cost far more than was necessary. All those were the first of those models, although the Chinese have also built a couple. So you could get them at a better price. You’re also signing a deal with the Americans for 50 or 60 years.
So you can contemplate how comfortable you are with that. The other alternative in the West would be the EPR, the European Pressurized Reactor. That’s originally a French design. The UK has adopted and are building eight of them.
And they’re of course also over time and over budget because the English insisted on making their own modifications to it to make it more difficult to build. So you could look at that design too. And both of these are gigawatt class reactors. And again, you’re going to build a long-term relationship there.
As much as I love everything about SMRs, I’m really fascinated the prospect of it. I think it’s not Armenia’s interest to be first on any of this stuff. You really want to go with proven designs. But realistically, when you talk about the time horizon, what’s happening right now and your demands for power, I don’t know that nuclear is even on the table.
It’ll take so long. By the time you have our new reactor up and running and people using it, it’s 10 to 15 years.
Hovik Manucharyan: Meanwhile, as I said, we have like potentially 300 megawatt consumption in 2027. Next couple of years. How do they plan? And I think that’s one of the answers that we haven’t gotten.
And we’ve actually tried to reach out to the Firebird people, but we hope if you’re listening, please accept our request to come on our show and talk to us on how you’re going to do power for it.
Richard Campbell: When I’m talking to other data center builders, what do they talk about building for power? They talk about solar, wind, and natural gas. And again, it’s all environmental. What can the region support and what’s cost-effective?
And wind and solar are very inexpensive, right? $20 to $30 per megawatt LCOE is incredibly inexpensive and operable in stages. You know, you have to build a whole gas power plant before you make any electricity where you can put up 10 wind turbines and have them running immediately or the first hectare of solar panels and have them running immediately and then add to them over time. The combination of solar and natural gas is very compelling. Daytime power, non-emitting and inexpensive, cheaper than gas by a lot.
And then when the solar can’t keep up, the power plant gets lit and run. A natural gas power plant running full load all of the time, 80 plus percent utilization, is far less expensive than what we call a peaker plant, a plant that you turn on only when you need it and turn off again. On the other hand, you have that balancing act. You also then, especially when the data center company gets involved in this, it’s normal for data center companies doing behind the meter build outs to build far more than they need.
That they make an agreement with the power company so that that is access power that can be sold into the grid. Building out 300 megawatts of solar is completely feasible. And to also have 300 megawatts of CCGT, well, that’s one turbine, right? Rather than five is 400.
So you add one turbine, but also the solar, and now you have this combination of sometimes the data center is only running on solar, sometimes it’s running on some of the gas. And when there’s more demand elsewhere in the nation, you have that residual available to you.
Asbed Bedrossian: Is it doable in the span of a year to 18 months?
Richard Campbell: None of them are doable in that amount of time. It’ll be three years. Solar is the closest, but you need 24-hour day power for a data center. But as fast as you can go on a gas turbine, because of the demand that’s going on, and now especially, it takes two or three years.
And that’s really fast compared to any other power plant.
Hovik Manucharyan: Let’s talk a little bit more about the ecosystem. Now, you covered this a little bit, but the data centers require major investment in construction while employing relatively few people. I was amazed that I think one of the US states where they had a data center idea and they actually gave a lot of tax breaks to this company, I think it was Ohio, to bring in employment. And the end result was net 10 employees added because it’s all automated and efficient that way.
Richard Campbell: Data centers were actually run dark.
Hovik Manucharyan: Yeah. Yeah. Lights out. So, so, so what kind of permanent jobs does a facility like the, um, AI factory in Armenia Firebird actually create and how many, uh, you know, what should be the expectations and projections and what is the benefit from, from a jobs perspective in general?
Richard Campbell: From the, again, from the construction side, you’re going to do stage construction, which is already happening. And so you’re going to have a set of construction teams that are going to operate over a long period of time. Because the data center is being built close to a town already, it’s a town of 40,000 or so. So there’s already residences, like there’s a lot of pressure to move data centers away from city centers and away from population areas.
But that means that you have to house the construction team somehow, and that adds a whole lot of other costs. So the fact of where you’re building right now, you already have people there and people could move in there that want to be part of the construction team, because you’re likely going to be building on that site for five, 10 years. Especially if you start building up industry around it, you know, the data center is industry is interesting, but it only becomes a catalyst to a bunch of other industry that can be part of that process.
And now the post the non construction employment is small data centers need relatively little supervision. You know, they need a set of security guards and they need a few folks to come in to do some maintenance. But for the most part, the servers are no longer serviced. They are run as is for three to five years and then they’re removed and repurposed and new ones are brought in.
So don’t look at the employment side of the data center directly. Look at the potential employment for the ecosystem that can grow up around it, that making agreements with universities and with startup elements and so forth so that new companies can emerge around it. That’s what the real potential here is, is to use the data center moment as a catalytic moment towards more technological industry. You have the educated people.
Armenia is a small nation, but you’ve always taken education so seriously. You’ve already got a startup culture in place. There’s so many powerful ingredients that are there. That you got to hope that the folks that are running this thing are recognizing that potential and saying, how do we create a business park adjacent to it that can be, you know, just being near the data center has its own gravity.
We know perfectly well the data is going to travel the same. The speed of light down a fiber optic cable makes no difference. They can use the AWS data centers in Turkey. But why would you when you could be using the one that’s down the street?
And so I think there was going to be a balancing act there. This is just an opportunity to ride the hype wave of building a major data center in Armenia to catalyze more technological industry in Armenia.
Hovik Manucharyan: I forgot to ask you a question from earlier, but I actually want to ask. Do these data centers equally, are they equally used for training and inference workloads or is it that some data centers are exclusively used for training, whereas maybe less expensive data centers are used for work for inference?
Richard Campbell: These days when you’re talking about the frontier models, they’re so massive that you need very large data centers, right? Those racks are ganged together with InfiniBand and better connectivity to be able to build out the models. And I don’t know that you even want to be in that business. Now that it’s up to you.
I mean, ultimately what they do with the data center is up to them, but we’re not really in a leading edge science moment anymore. The technology is largely known. We’re mostly tuning now. We’ve kind of gathered all the data now.
We’re just trying to build, refine those models better and better. It’s the engineering side of using this technology effectively that’s going to be the future. During the dotcom boom, yeah, we improved HTTP and HTML a bit, but mostly it was thinking through the business models and making a difference. The business models of AI are still wildly immature.
There’s lots of innovation to be done there and having local compute is going to be great for that. But the actual discrete workloads and whether we’re building models there or not, I don’t think it’s that important. It’s not big enough for the biggest models anyway. I think we’re rapidly discovering that the model itself is not the asset.
The harnesses and the tooling around the model are the assets. And that’s where the innovation can come. You know, you can probably name all of the major models that exist today. There’s like a dozen.
But can you name all the harnesses and tooling? Because it’s thousands. Because we’re still experimenting, because there’s still an exploration going on to say, how do we get real value from these models?
Asbed Bedrossian: Richard, you just mentioned about the benefit of having this data center in Armenia and Firebird may be physically located in Armenia, but the hardware is produced in the U.S. or by U.S. And access to advanced chips remains subject to American export policies. That creates a distinction between the physical hosting of this compute infrastructure and then having sovereign control over it. And this term, maybe even at the buzzword level, sovereignty, has been a big factor for the Armenian government.
Richard Campbell: Sure.
Asbed Bedrossian: For its own sake, has increasingly treated advanced computing hardware and model development as a national security issue at this point. And the 2025 Framework for Artificial Intelligence Diffusion sought to regulate access to advanced AI computing and model weights. This is going on even today and this week when, you know, Some new regulations are coming down from the government. I don’t really have a good assessment of whether they’re good or bad, but at least there’s something coming from the government.
So we’ll put many of these links in our show notes, folks. Go to podcasts.groong.org/episode-number. Richard, if Armenia hosts the hardware but depends on U.S. companies and U.S. approval for replacement of these chips, the software, the future upgrades, how much technological sovereignty Does Armenia really have?
Richard Campbell: Well, I don’t know if you want to get in the chip business. That’s its own can of worms, without a doubt. You know, maybe there’s a startup that’s going to want to try and take that on, but that’s not ultimately the answer there. This is always going to be a risk.
It’s unavoidable. But, you know, sovereignty starts with the land. And so the fact that it’s present in country gives you a say one way or the other, no matter what. That being said, what happens next is always an interesting question, right?
Once money has been put into an effort, there’s going to be an effort to protect that investment and it encourages additional investments. So you do have to pick a path of what you want to be on. And looking at the recent geopolitical realities around Armenia in the past couple of years, This swing, you know, in some ways you’re looking at what has happened to Georgia in the past few years and going, maybe some of that for us. You’re going to have to have a relationship with someone.
And maybe I’m biased because I’m in the West. I’d like you to be in the West too. You know, all of that is good, but none of it is free. You know, it comes with consequences without a doubt.
And that being said, who knows what other alternatives are going to come down the path over time. You know, anything Your country has shown its resilience, its ability to make a deal where a deal makes sense. So there might be some alternatives. And ultimately, a data center is just a large concrete box with power, water, and data pulled into it.
What you run at it is almost secondary to the point. Those things could change over time, but you’re going to have to make a commit somewhere. And yeah, there’s always going to be a conversation. You know, you own the padlock to the fence.
So you certainly have a part in the decision making of what happens with that data center.
Asbed Bedrossian: That’s right. But what happens inside that padlock? And could the US export controls or licensing conditions limit how Armenia could even use the facilities in the future?
Richard Campbell: Almost certainly the question is, is it beneficial? You know, who’s going to benefit from limiting those things? We’re in a funny time with the US right now where they have a tough time following their own laws, much less the laws of the international system. So it’s hard to know exactly how they’re going to behave.
And it’s made the data sovereignty issue a much more sensitive one. You’re seeing a lot more push everywhere to having more data centers in country and more data center control, more regional. So I wouldn’t be surprised if you head down that path as well. It’s just a question of, does that make sense at this particular moment?
Things will change again. And the nature, the alternative is not being part of the conversation. And I think that’s a worse alternative.
Hovik Manucharyan: Well, couldn’t the alternative be having Technology that is not export controlled, but still multiple decentralized data centers. Or do you think that that immediately takes you out of the conversation?
Richard Campbell: No, I think the same way you have multiple gas pipelines, it’s in your interest to have multiple fiber pipelines and it’s in your interest to have multiple suppliers of hardware, ultimately. And you’ve already done the hard thing. You’ve proven it’s possible. The fact that there is a major, modern, admittedly pilot, but a beginning of a real data center in Armenia just shows to the rest of the world, you could run workloads here.
Let’s make a deal.
Hovik Manucharyan: Right. So previously you mentioned, I think, which is very important that Armenia should capitalize on this and work with educational centers, research to help promote research. Armenia has so far committed to purchasing 25 million of Firebird’s computing resources over five years. I mean, honestly, I think that’s way, way too little because- It’s a drop in the bucket.
Yeah, it’s a drop in the bucket.
Richard Campbell: But so is 15 megawatt data center. These are all starter projects, Hovik. That’s right. The fact that they’ve done one says they can do more, right?
I think that’s what I would want. That’s exactly what I want to see when the pilot project is up.
Hovik Manucharyan: Yeah. I love your positivity. No, I’m super excited that I know that now, right?
Richard Campbell: The fact that the university has already done that means smart people are working on the problem.
Hovik Manucharyan: Yeah.
Richard Campbell: Now do more.
Hovik Manucharyan: Yeah, exactly. Armenia’s national debt has increased over the last eight years by $8 billion or $7 billion. And they have nothing to show for it. That’s what I’m worried about where, you know, Rational investment is not being done.
I’ll get off my soapbox, but what I also don’t like is people trying to hype this into something it’s not.
Hovik Manucharyan: So people who support this data center are saying nod nod, wink. All this usage that Armenia will get from this data center can also be used for military purposes and therefore it can be used to secure Armenia. But when I read the validated end user program by the Department of Commerce, And I see a lot of design conditions and regulations on how this can be used. So how much military value does access to a large GPU actually provide, given that it is using export control technology?
Richard Campbell: Well, I think we know, you know, considering your neighbors across the Black Sea and what they’ve taught the world about how militaries can function differently, we know compute is going to be important or is important now. And sure, you’re building out a data center that has fairly strict rules on that, but you’re also learning how to build out a data center so that nothing precludes you from building another one. And one that the military could have more flexibility with to learn the potentials of compute in modern conflict one way or the other. There’s no reason to violate the export controls.
There’s other options. The big thing here is you’ve got a data center and you know what it takes and you know what the sources are. You just have to go about it differently if it’s going to be an entirely sovereign data center used for its nation’s purposes.
Hovik Manucharyan: Yeah. And I guess it also depends on what military means, right? So could you use it for analyzing, you know, aerial visual, you know, surveillance and, you know, pictures and things like that? Yeah, I would pretty much exclude the use of them for controlling drones or missiles, but it’s a question of actually even like this adjacent dual use and how much Armenia would be allowed to do that.
Richard Campbell: Yeah, but I mean, ultimately, the sovereignty angle means you’re probably going to want a solely controlled data center of your own at some point. And again, going down this path means you just open the door to the skills and resources that are necessary to make it possible.
Asbed Bedrossian: But even if you had complete control over the, let’s say the GPUs or the data center, the software itself can limit. I mean, these LLMs are going to be smart enough to say that is not what I was designed to work on. Sure. If you are driving a workload.
Richard Campbell: Everything is buildable for a sufficient time and money.
Asbed Bedrossian: Richard, this whole topic basically brings up the fact that these data centers are becoming high value targets as well. In the war on Iran, we saw actually Iran target some of the data centers in the Gulf countries because they said that they are being used to support the war effort. By the West on Iran. So they targeted them.
Now there’s also Brookings analysis that argues that overseas frontier AI data centers can create national security risks because they are physically vulnerable. Difficult to conceal, exposed to espionage, sabotage, military attack. I’ve worked in an IT environment data center. We fortified it.
We started limiting access. We even put cement pylons so that a truck couldn’t ram it. But I’ve never seen a data center that’s protected by a Patriot missile system. Where
Richard Campbell: are we going with this? You know the other thing that’s hard to hide? Hydroelectric power. Is it infrastructure?
Yes. Is it vulnerable? Also, yes. This isn’t any different.
They are important assets. They have a high value. There’s reasonable levels of protection. I would strongly advocate for a peaceful approach.
As much as is feasible. But you’re not going to see them any differently than any other strategic asset in the country.
Asbed Bedrossian: Well, so Taiwan’s semiconductor industry is often described as a silicon shield. Everybody loves Taiwan Semi. They don’t want anything to go wrong with it, whether it’s in Western hands or Chinese hands or anything. So could Firebird Create even a small version of the shield for Armenia, or is Armenia too small to worry about, to be honest?
Richard Campbell: I think as long as your neighbors are benefiting from your capabilities, they all have an incentive to keep that capability intact. Creating those interconnect between your neighbors is important. Various opportunities for computer are important. This is the power of trade.
Globalization isn’t very hip right now, but coming out of World War II, Creating that interconnect between nations through trade largely ended major conflict in Western Europe. You know, there are other issues that went around that, brought more people to the table. There’s a really strong case here that a 21st century trade is in electrons or in light, right? It is this compute capability.
I just hope that Armenia takes advantage of that to make sure that the neighbors are also benefiting and see the value in it so that everyone’s on the same side of let’s keep that resource running.
Hovik Manucharyan: I love your positivity, Richard. Unfortunately, I think in terms of thinking about peace and mutual coexistence, our region is a little bit lagging from West and Europe, which is what keeps me up at night. But you’ve made
Richard Campbell: progress, Hovik.
Hovik Manucharyan: 2026 is a
Richard Campbell: pretty good year for Armenia. You know the path like let’s we if you go all the way back to the end of the Soviet Union like look at the path that Armenia has been on it’s been tough from there and I’m I’m doing Armenia discredit by starting there because there’s been thousands hundreds of thousands of years of humans in this part of the world right even before there was homo sapien there were hominids in the area like it’s the the history is unbelievable. But this current cycle, this past 40 or so years since the collapse of the Soviet Union, has been a challenging time. And it’s been far darker than it is right now.
You do seem to have moved forward on a more peaceful path. I think you can look your grandchildren in the eye in a different way today than you could even 10 years ago and say, like, we’re trying to make a better world for them. This is an ingredient into it. It’s only possible because of the path that Armenia is on right now, that suddenly these sons of Armenia, this diaspora, the guys from Firebird, are bringing back to their home country an incredible resource that brings them into the 21st century, that adds more to the path that you’re on.
You know, it’s not the reason you’re going that way. You’re already going that way, but it’s one of the benefits and it strengthens that path to have more options and to have more peace.
Asbed Bedrossian: Richard, I agree with you a hundred percent. There is a lot of positive in this and the fact that it’s a collaboration with the diaspora, that is a huge positive as well. I think that living our history We also are not sure exactly how some of these decisions are made in Armenia. I think we are trying to understand some of those things rather than just I think we came out a few episodes ago and said that we both love to have a data center in Armenia.
This is a positive thing. We just want to know how they’re going to be taking care of the kind of issues that nobody’s talked about, for example.
Hovik Manucharyan: Yeah, and my worry is, I always worry, so sorry for that and don’t feel pushed too much, but are there any strings attached? And we’ve seen a lot of geopolitical adventurism in the region that has not resulted in good things. I think that we covered this on this podcast.
Asbed Bedrossian: Well, so far it has resulted in the enmity of some of our neighbors.
Hovik Manucharyan: This is more of a political issue. The current government says we’re in an era of peace, whereas our neighbor is saying that you are the enemy. We’ve taken much of your land in 2021. We’ll take some more.
So there’s a dissonance over there. But let’s just say you were advising the Armenian government today on this project,
Hovik Manucharyan: on the use of Armenia’s AI strategy. Its current scale and model, what would you say and what would you focus on?
Richard Campbell: Well, please don’t treat this as just a data center. This is an infrastructure project. This is a foundation for expanding the technological ecosystem of the whole nation. New power sources, new education, you know, GPUs are going to depreciate, but talent doesn’t.
If you grow the people around it, that’s what makes the transformation. The investment ultimately comes to the people, not to the physical assets, right? The physical assets ultimately are only valuable if they benefit the people. So if the government is keeping the people in mind and growing those things in every way, that’s where the benefit comes in.
This can be a strengthener of Armenia, both internal to the country and with its relationship with its neighbors. If you get both those things together, the talent only flourishes further. I know I’m a techno-optimist. I see the potential in this.
I really appreciate your skepticism. These are the questions to be asked. This goes badly when it’s just consuming your power and resources to make compute for elsewhere. This becomes incredibly valuable when you grow an ecosystem around it to benefit yourself and everyone around you.
Asbed Bedrossian: Yeah, the skepticism is on the political and geopolitical effects and our lack of trust in the government doing this. But on a technology scale, I think it’s a wonderful thing. Frankly, I wanted to have a high-performance cluster in Armenia 25 years ago, but those things were not possible at that time. Richard Campbell, thank you so much for joining us today.
We have enjoyed this conversation immensely. We hope that you will join us now and then.
Richard Campbell: Happy to. Thank you so much for having me on.
Hovik Manucharyan: It was a pleasure, Richard. Thank you.
Asbed Bedrossian: That’s our show today. This episode was recorded on August 11, 2026.
Asbed Bedrossian: We’ve been talking with Richard Campbell, who has been playing with computers since the 1970s. He has done almost every job you can think of in computing. Including manufacturing, sales, software development, helping to scale websites during the dot-com boom, and teaching software development all over the world. There is a lot more to his bio, and you should go to our episode page, podcasts.groong.org/episode-number.
Check it out. Click on the bio.
Hovik Manucharyan: We hope you liked this discussion, folks, and thank you to everyone who has supported us so far. podcasts.groong.org/donate We’ll get you to the page where you can give us a monthly contribution or even a one-time contribution, and that all helps us do things more efficiently. And if you can’t support us financially, what can people do, Asbed?
Asbed Bedrossian: They can share, they can like, they can hype, and they can comment on our shows and keep in touch with us.
Hovik Manucharyan: And most of all, make sure you’re subscribed, folks. We probably have a lot of visitors who just got here from searching for Richard Campbell. So if you like geopolitics in Armenia with an occasional geek out session like we did today with Richard, then we are the channel for you. Thank you very much.
And I’m Hovik Manucharyan.
Asbed Bedrossian: And I’m Asbed Bedrossian, we’ll talk to you soon.
Hovik Manucharyan: Take care.