Taisu AMA: AI Infrastructure: Power, Data Centers & The Real Bottlenecks- Investor Perspective

Hosted by @Exascale Labs · 2026-07-03 · Tags: EXASCALE

TLDR

Exascale Labs and lead investor Taisu argued that power, cooling, deployment speed, and operational execution—not GPUs alone—are the primary constraints on AI growth. They presented modular data centers and an asset-light orchestration model as faster, more capital-efficient ways to meet expanding global compute demand, while acknowledging contract conversion, permitting, deployment, and scaling risks.

Speakers

Notable quotes

Transcript

Speaker 1: Hello. Testing. Can you guys hear me? Loud and clear. All right. Thank you.

Mark: We're going to get started in just a couple of minutes. Thank you all for attending. All right. Hello, everyone. We're going to get started. Today, Exascale Labs, as you all know, is going to be hosting a series of these AMAs. This is the second one. You can go back on our timeline and look at what we do with our CEO. We're on a mission to educate people as to what is actually powering all of the AI that you are using in your LLMs and even separate applications, because the LLMs and the prompting and the AI agents, the agentic AI, those are all sexy. We realize that. But what's really needed, and especially now, because there's like a 10X level of AI that's needed in the next five years, is what we're going to talk about today, which is the AI infrastructure. And joining us is Taisu, who's our lead investor, actually. And they've got a wealth of knowledge around investing in this space, as well as some insights as to things that you should be aware of. But before we begin, we want to be clear that Taisu is an investor in Exascale Labs, our lead investor, top investor. Our comments today are based on our investment perspective and publicly available information. We will not be discussing non-public financials, valuation, IPO pricing, allocation, offering terms, or giving investment advice today. But I do want to encourage you to ask questions, to follow us, especially Tai Sue and Exascale Labs, and to subscribe if you like what we're saying. And I think you will today. So our guest is Rafael from Taisu. Rafael brings over 10 years of experience in banking, startups, and fund management with a focus on alternative investments in the last five years and is actively involved in Web3 since 2021. As an operator, he has driven a company from inception to profitability, leveraging skills in technology, product development, and fund management. and that he's acquired from roles at Credit Suisse, a fintech startup in a boutique asset management firm previously. Rafael holds a Bachelor of Science in Information System Management from Singapore Management University and a master's degree in Computing Infocomm Security from the National University of Singapore. That's quite a lot. And I'd like to ask, you know, Rafael, Tai Xu's perspective as an investor, What first made Exascale stand out in the AI infrastructure market?

Speaker 1: Thank you. Thank you for the question and introduction, Mark. So I would say two things. First, the layer that you are operating in. When we were looking at the market, so everyone was looking, talking about chips and models, but the physical infrastructure, the physical infrastructure underneath it was being underbuilt. relative to where demand was heading. And second, the model itself. So you're not trying to own data centers and compete with hyperscalers on CapEx. So you orchestrate capacity. You sit between supply and demand and making the whole thing work operationally. So we find that more capital efficient, more capital efficient way to capture the infrastructure opportunity.

Mark: Yeah, I mean, that's, it took us a while. We found the same exact thing because it's very, very competitive at the AI level in terms of, you know, the LLMs, that top level of AI agentic computing. And we completely agree with you on the capital efficiency. of finding that infrastructure opportunity and taking advantage of it. And we do know, especially when robots come on, the demands are going to skyrocket. They already are. I mean, we're already behind. The industry is behind on delivering these AI infrastructure solutions. So, you know, I know a lot of investors are focused on chips and GPUs. Why do you think the bigger bottleneck may Maybe the data center's power, cooling, and infrastructure. I hinted at it. Why do you think that is?

Speaker 1: I would say because GPUs, they are only part of the equation. So even if you have access to the hardware, the chips, you still need somewhere to put it. You need the power to run it and also the cooling to keep it operational. Now, those three things, traditionally, they would take years to build. And we're talking like two to five-year timelines. You can't just software engineer your way out of a power constraint, for example. Now, what's exciting about what you guys, Exascale, is doing is that you are actually attacking that bottleneck directly. So there's the modular data center approach. and that changes the deployment equation. So instead of waiting years for a traditional build, you can stand up capacity in a fraction of the time in locations where the demand actually is. Now, I think that's a meaningful structural advantage in a market where speed of deployment is itself a competitive mode, in my opinion. What about Yeah. What about from Axascale's perspective on modlenecks?

Mark: Yeah, and thank you for that. Putting me on the spot. But I'm prepared to answer this question. But first, I want to define what a modular data center is. So these are data centers shipped in containers, like shipping containers. and they're fully ready to go. You just plug and play, essentially. I'm simplifying it. So please, if you're an expert out there, I know it's a little bit more difficult. You've got to hook up the cooling and all that. But a lot of this is self-contained. And instead of waiting two to three years, which now it could be even longer to build a data center, these modular data centers can be dropped in literally and set up in four to six months, maybe sooner if you've got everything cleared and ready to go. So imagine having to wait two to three years and, you know, who knows what bottlenecks are going to exist in the future, especially with regulation now. People are complaining about power and cooling. Now you've got a modular data center approach that you can get done in four to six months. You get their projects, you could start immediately working on your project. It's a huge innovation. And if you're looking at data centers, this is something you should probably look at. But I have to say you nailed it. Again, GPUs get all the attention because they're the easiest part of the story, right? Everyone knows about Nvidia. Nvidia's did an amazing job, most valuable company in the world. That's how good they've done it. And everyone who wants to talk about the chips, I get it. Chips are very sexy again. It's A sexier part of the story. But the GPU itself is just an incredibly expensive paperweight if you don't have the infrastructure around it. So from Exascale's perspective, the real bottleneck is the physical layer. The questions that we ask, and anyone looking at this has to ask, and what investors are asking now, because they're finally getting a clue, is can you get enough power? Can you bring it online fast enough? Can the facility handle 100, 150, 200 kilowatts per rack and beyond? I mean, we're going to get up to 800 with the Vera Rubin coming out next year. the birirubin chips. Can you cool it? And is it environmentally friendly? We're hearing a lot more of that. Can you operate it reliably? Can you deploy it close enough to where the demand is? And can you do all of that without waiting three years and blowing up the economics around the area? For example, in California, it's very expensive to operate a data center. And other parts of the country like Texas, it's not. So you've got to be very careful where you deploy these things and to make sure it's close enough to where you need them. It's a very complex formula figuring how to do it. So that's where the market is getting constrained. There was a recent research report that came out that said that 50% of these data centers are either delayed or have been canceled. And that's a huge number. And the AI world moves at software speed, but the infrastructure world has Historically moved at utility and construction speed, which you all know is a lot slower. And with AI, you know, speeds up software quite a bit. It doesn't speed up the hardware side, unfortunately. Not yet anyway. And so for us, that mismatch is the opportunity. Enterprises that want AI capacity now, but are waiting on traditional data center models that wasn't designed for this level of demand, frankly. So our view is pretty simple. The next phase of AI is who's got the best model and who can turn on the power, the cooling, and make that GPU capacity into live, usable AI infrastructure quickly, four to six months, not two to three years. And that's why we're focused on modular data centers. We do data centers, traditional data centers as well, but we're really focused on modular data centers right now because that's where we see the demand and the advanced power architecture that's within the data centers and obviously operational execution, which is very important because, you know, in this market, speed and capacity is actually the moat right now as investors or as anyone wanting this type of data center will tell you it is speed to market. That is the critical factor in most of these decisions. But let me let me now turn it over to Taishu again, because you've got deep experience in digital assets. What parallels do you see between digital asset infrastructure and AI infrastructure?

Speaker 1: I would say this is actually more than what most people would expect here. So when you think about crypto mining, so actually at its core, it's a pretty compute intensive workload here. So what is that? What does it need? So you need like cheap power, reliable cooling, you need 24/7 uptime to be economically viable. And the companies that want, they weren't necessarily the one with the best token thesis, the best tokenomics, the ones actually that could procure power cheapest, they build the capacity fastest, and they keep machines running longer. So you can see that finally there's that parallel.

Mark: Yeah, I'm glad you drew that parallel because crypto mining, we used to do that and we used to understand how that works. And that's where we got a lot of our knowledge, frankly. But, you know, a lot of what you do in crypto mining, believe it or not, because we're using the same chips, is what we're doing in AI. Obviously, it's a lot more advanced than that. And we've we've really had to put in different types of infrastructure. How you architect it, how you build it's different, but the knowledge is fairly similar. So that was a good evolution, we feel, in our eyes. So for for people who understand AI apps like ChatGPT and Claude, but not the infrastructure beneath them. I don't blame them. Nobody thinks about the infrastructure underneath all this. How would you explain Exascale's role in the AI stack?

Speaker 1: I would use the slightly cliche picks and shovels analogy. Now, every AI application today that people interact with daily, they need all these needs enormous. want to compute, to train and to serve all the users in real time. Now, that compute, that has to live somewhere, and that's basically physical data centers. These data centers, they need our cooling and basically round-the-clock operational management. Now, that's the layer that Exascale is in. Now, what I think is distinctive about your model, Exascale's model, is that you guys are not owning the rest data. You are orchestrating it. You are connecting GPU capacity supply to enterprise AI demand on long-term contracts. Now, that's capital light, and that's, in my opinion, sitting in a critical part of the stack. So, but regarding this, What about you, Mark? How would you, in your terms, explain that access skills role?

Mark: Well, I do like that analogy, depiction shovels. It's exactly right. But from my perspective, most people think AI is the app. They see ChatGPT, Claude, Midjourney, whatever is sitting on their phone or laptop. But that real power is happening underneath that. It's like, you know, you go to a restaurant and you order a meal. and you have the most amazing meal of your life, but you're not thinking about how they cooked it, every ingredient, every, you know, what they heated it up to, what they added. None of that is, you're not thinking about, you're just thinking how great the meal is. But, you know, we're the ones in AI that's serving that or creating it. So, you know, AI isn't just floating around in the cloud like some invisible genie. It runs on GPUs. Those GPUs sit inside of data centers. Those data centers need incredibly massive amounts of power. And next year, it's a heck of a lot more. In fact, we have to change the whole infrastructure of these data centers and modular data centers in order to incorporate what's coming next to with the Vera Rubin chipset from Nvidia. That's just the next generation of chipsets. Sorry, I didn't define that earlier. Those data centers also need high density cooling. And that cooling is more and more got to be environmentally friendly. And I think when you look at the types of cooling that's coming out, they are a lot more. You've got to also make sure that the networking, the uptime, and you need an operator that can actually manage that whole thing without turning it into some giant science experiment. You need experience. And that's where we fit in. You know, we're not trying to be another chatbot or LLM. We're not trying to compete with OpenAI or Anthropic. What we're doing is building and orchestrating, like you said, orchestrating is a key word, but physical AI infrastructure, those companies and enterprises adopting AI need in order to scale their operations. So we're the chef in this story. The simple way I can explain it is probably this. I like the chef analogy, but there's another analogy I like. If AI apps are the race cars, right? That's what you're seeing. It's going around the track. Exascale is helping build the track, the fuel systems, the pit crew in the garage. And without that layer, none of AI is excited about how it actually works and runs. They're not thinking about that, but that's as important, if not more important. You can't have a race car sitting in the middle of a field, right? You've got to have all of that infrastructure around it. So what makes You know, our model different is we're focusing on being capital efficient. We don't need to own every piece of land or every building. We just basically connect the GPU supply power, data center capacity and enterprise demand into a long-term contracted infrastructure. So we're in the part of the stack, the AI stack where demand is exploding, but where execution is really hard. I mean, guys, I was in software my entire life. Hardware is 10 times harder. And these are real physical hardware pieces that need you need to procure from around the world. You can't just update a piece of hardware on the fly. You can't change it on the fly. Everything's got to be physically done until we have robots doing it. It's incredibly difficult. But I'm going to go back to Taishu now. And just from an investor's point of view, what should people look for when evaluating AI infrastructure companies?

Speaker 1: I would say a few things. So power access, is it secure? Has the power been secure? And if you can get more power, and that's really the fundamental input. No power, no data center, no, yeah, nothing. The geography positioning, where are the clusters relative to where demand is growing? Major listed new clouds. They haven't built much in Asia yet. And yet, demand from Asian AI companies externalizing their compute, it's real. And it's accelerating, which is exactly where you guys Exascale is positioned. Next would be business model capital efficiency. Are you owning the assets or orchestrating them? And also in terms of contract quality. Long-term off-take agreements with prepaid terms are obviously a very different risk profile from spot market exposure.

Mark: Now, what's interesting is you're in Asia, and so you're seeing what's going on in Asia. My exposure has been a little bit more limited, but when I went to Taipei just a few weeks ago, it was massive. They had buildings around the city that were hosting This massive event and the enthusiasm level was off the charts So i'm just genuinely shocked about how on top of things asia is and what they're investing in and the robots that are coming out the ai infrastructure uh is starting to be delivered whether it's from american companies like ours or other companies um, but I I know you said that there's not a lot of neoclouds that have been built in Asia yet, but my, it's going to be coming in a big way. I was really genuinely impressed by what I saw there, just a massive show, a massive event. But let's, Rafael, let's get over to, you know, us personally, you know, because I know a lot of people are interested in what we're doing and how we're doing it. What gave Taisu confidence in ExaSkill's team and execution capability?

Speaker 1: Well, in my opinion, it's the depth of operational thinking. In this space, I believe that having a compelling market narrative, that's like table stakes, right? But what's... What's really rare? It's a team that's thought through the hard parts. You mentioned that hardware, it's really very hard and difficult there. How you manage all this uptime at scale, how you handle GPU procurement lead times, how you convert this large pipeline into sign contracts under real-time pressure. When we spend time with you guys, a team, we can tell that Exascale really understands that execution in AI infrared is really opportunity intensive in a way that most tech investments simply aren't. And you guys have built for that.

Mark: Yeah, not without a lot of work, but yes, you're exactly right. Love that perspective. And I'm glad you've invested in us. And just to everybody listening in, a reminder to like and subscribe and comment, ask questions. This is a very important future that not only us, there's other companies obviously that are investing in this, but this will be, this is the single biggest bottleneck right now in AI. So how do you think about the IPO as a milestone? for Exascale beyond just the transaction itself.

Speaker 1: I would say it's primarily about what comes after that, actually. So going public, so this introduces this reporting cadence and this level of transparency that changes how a company operates. You know, you have this quarterly accountability, governance standards, and of course, a broader shareholder base than now. So for an AI infrastructure company like Exascale, so I believe that this discipline can actually strengthen your enterprise customer relationships. Like large customers, you want to work with companies that have public financial reporting. Also, The capital markets across that comes with the access that comes with the NASDAQ listing. So it allows you to finance this infrastructure build up. So at a scale that, well, just private capital alone can match.

Mark: Okay. And so when you look out, I'd say five years from now, Let me ask you another question before that. Do you believe AI infrastructure is more than a short-term trend, or is it just short-term because we need to figure it out?

Speaker 1: I would say it's more than a short-term trend, mainly because demand itself isn't just coming from one place as many. foundation model training, that's inference at scale, that's automation, enterprise AI deployment, et cetera. So every one of these, they are adding incrementally to this total GPU hours that's demanded globally. And every time there's this new model capability released, there's a new one coming up. Sonet, Opus, Vivo, etc., all this, it just creates more demand, not less. And all these compounds. Also, there's also mentioned previously as well, the geographic angle matters as well. So Asian markets, I believe they're still early. The demand from Asia, AI, hyperscalers, externalizing compute to foreign data centers, really just beginning to build here. And there's this market that you guys Exascale is in before the major Western competitors are showing up, which is really where you are nicely positioned to be. So, yeah, not a short-term trend.

Mark: Yeah, my view is it's not a short-term trend for two main reasons. One, all predictions are from every global tech leader out there, whether they have an interest in AI infrastructure or not. is that the demand for compute is going to go up 10x. So imagine that. We can't even keep up with demand today. So imagine 10x in the next five years. Think about when robots come online. Think about when everything is either enhanced or taken over by AI. In fact, you know, Elon has said that he's going to put data centers in space. Now, why does he need to put them in space? Because he sees the issue with scaling these types of of AI infrastructure. It's not easy. It's going to be difficult. So it's definitely not a short term trend. If anything, we're underestimating in my belief what we're going to need going forward. But going forward, Rafael and looking five years out, where do you see exascale?

Speaker 1: Oh, I like this looking forward question. So five years out, I would see Exascale as one of the defining names in Asian AI infrastructure. So also mentioned previously, I saw that demand here for compute in this region is really in its early innings. Asian AI hyperscalers, I also mentioned just now, they are just beginning to externalize the workloads to foreign data centers. So I'll say in particular in Southeast Asia, so we don't see that much yet. And this window won't stay open forever. So I'll say if you guys continue to execute on the pipeline that you guys have, continue scaling the SLI model, five years from now, I expect Exascale to be in a scaled, profitable, public listed, infra operator with a genuine first mover position in this fastest growing AI computer market in the world.

Mark: Yeah, it's always nice to hear from third parties looking in. When you're inside the company, you know, obviously it's harder to be objective. So I appreciate that answer. What do you think? And maybe we'll both answer this. What do you think people should understand about the difficulty of building real AI infrastructure?

Speaker 1: Well, for this one, I believe you'll be better positioned to answer this.

Mark: Okay. Yeah, we want to keep it brief here for everyone listening. But I do want to comment and say, if you're listening to this now, you're already way ahead of anyone else that's either looking to Invest in the space or use data centers in the space or anything having to do with AI infrastructure. Nobody talks about this. Not a lot of people, right? It's not the sexiest part of the equation, but it's the most important right now. So the thing people, I think, need to understand is that real AI infrastructure isn't just, hey, I'm going to buy an Nvidia chip and plug it into my laptop. You know, that's not how it works. I know that's the easy version people imagine. And I'll be honest with you, before I got into this, that's what I thought I didn't know anything about. Hey, you need quite a bit of power cooling and big data centers in order to operate the AI that I'm, the little AI that I'm using. So the reality is much harder. You need access to the GPUs, which are constrained sometimes, sometimes they're not, it just depends. But what is constrained is you need a data center that can actually handle That density of the GPUs, it's only going to get worse next year with the Vera Rubin chipset from Nvidia. The power requirements are different. The cooling part are different. And a lot of companies and data centers aren't prepared for it. So especially around the power, the power itself is going to double the requirements. So then you need high density cooling. These things get hot. And now the cooling's integrated within the chip itself. Can you imagine? This is how complicated this stuff is. When I went through and saw exactly what has to happen in order to make this work next year, even now, really, these special racks that need to have cooling infused within it, and then it's got to flow through the chips. When you look at the next generation chipset, it's quite complicated, quite amazing. And again, it's not software, it's not something you just update on the fly. So, and then the other piece is you need the networking, the redundancy, meaning you need backup batteries. Power isn't always reliable around the country and around the world. So it needs to be, you need to understand the uptime, the security, the monitoring, and you need to have an operations team, whether it's yours locally or somebody like Exascale that knows how to keep all this running 24/7. It's a big operation. And so, Every one of those pieces has a lead time. Transformers, switch gear, power contracts, permitting, interconnections between the thing, construction, the cooling systems. None of this moves that software speed, unfortunately. And that's really a big bottleneck, as we talked about in AI right now. The demand side obviously is moving incredibly fast. I set it a 10X figure. I stand behind that. You look at all the industry execs and people The researchers and analysts, they all say the same thing. And everyone wants AI capacity now, but the physical infrastructure side still moves in months or years. If you build it to traditional way, it's going to get longer as more and more regulators jump in on this. And that's where we came in and we said, hey, we have a solution for this. And we're trying to compress that timeline we have with these modular data centers. We're committing to four to six months as everything's lined up. So we bring all that together, the compute, the power, the modular data center design, the high density cooling and operations into a much more integrated model. It's basically plug and play. Don't quote me on that, but it just for heuristics, it's plug and play. So when people ask, you know, how hard is this? I say, well, AI feels like software, but it is now one of the most physically demanding infrastructure businesses in the world. To make it look that easy, you got to have quite a lot behind it. And those companies that understand that and that put that together are the ones that I believe are going to be the winners in the future. So I will end on that as far as my explanation. I could go on and on and on. But I do want to ask a final question of Ty Sue, because there are challenges in this, you know, this business. And by the way, we're going to ask questions from the audience after this. So if you have any questions, put them in now. But my final question for Ty Sue is, you know, what are the main challenges of Exascale as a business? It can't be all, you know, roses and champagne, right? What are some of those challenges that you see?

Speaker 1: I would say from the challenges perspective, I would say any serious infrastructure business at this stage, they will face really the same core challenges, such as converting pipeline to sign contracts, managing deployment timelines, scaling operations fast enough to meet demand and yet, you know, not compromising quality. Now, granted, These aren't unique to you guys, Exascale, per se. It's just that these are the table stakes of really building in this space. So what gave us conviction here that Exascale can meet these challenges. And we're confident in that, of course. That's why our investor is that the team truly understands that these challenges are at a deep operational level.

Mark: OK, thank you for that. All right, now, everyone, we're going to take audience questions. We have quite a few, so I have to be kind of careful about how I pick and choose these. But the first one is from Ox Apon. From an investor lens, what quantitative metrics or red flags, such as energy cost per flop, water usage intensity or stranded asset exposure, separate viable long-term AI infrastructure bets from those likely to repeat the boom or bust pattern seen in certain crypto mining operations? Wow. Well, The biggest red flag for me, from an investor point of view, is power. And that is, okay, they're building these modular, they're building big data centers. And I'm not just saying Exascale, I'm saying any company, Google, Amazon. So what government or what town or what city is supporting? Are they really being supported or are they facing lawsuits left and right? Because we know with crypto mining, they were being shut down left and right. They were taking, they were taking, they're passing new laws outlawing these things. Are they going to do the same thing with AI data centers? Now, everyone knows AI data centers are much more important than crypto mining operations. Sorry, crypto, but it's true. Because everyone could use this. Not everyone's using crypto. I know that you might think they need to, or, you know, it's the financial system of the future, fine. But For AI, it's being integrated into things that everyone needs. So that, and I'd say second would be cooling, because there's going to be some new innovative cooling features that are coming out. We're integrating into our next generation designs, because the chips itself at the chip level need to be cooled. You should see, they look like little capillaries throughout these chipsets. It's amazing. But all that's got to happen, and it's going to be somewhat environmentally friendly. So that's, from an investor's lens, that's what I'm looking at. That's what I would be interested in exploring and making sure that whatever investment that you're making has those features, or at least they're thinking about it and know how to answer that question. Okay, next one's from Gator Nation, down in Florida, I'm betting. We've seen a few mentions of MOUs. Have you seen that demand strong for modular data centers? And do you think growth will happen faster in your niche versus traditional data centers? Okay, yes, we've had quite a few MOUs. There'll be more coming, more contracts signed as a result of those MOUs. And have we seen, yes, very strong for modular data centers. Obviously, we're educating the market right now on what's the power of using a data center over a traditional data center. We do both though. So if you want a data center, we can do that. If you want a modular data center, we could drop ship one in, set it all up and operate it if need be. And so do I think the growth will happen faster in our niche? Absolutely. The challenges are manifold with traditional data centers. As we've mentioned, these are big, massive buildings that require a lot of power now and probably twice the amount of power when the Vera Rubin chipset comes in. So you can imagine the surrounding area. You can imagine citizens, regular citizens in those areas. A lot of people are overhyping the challenges of data centers, especially around pollution. And, you know, power is probably true, but the rest of it around pollution, these are very clean types of operation. There's a lot of people talking about modular nuclear power stations, and there's going to be environmental concerns around that. So do I think growth will happen faster in RNH? Absolutely. It's going to because it's cleaner, it's easier, and we could deploy it much faster. So yes. All right. And Rafael, feel free to jump in on any of these if you want.

Speaker 1: No problem. On MOUs, yeah, I have to defer to you from MOUs perspective.

Mark: All right, next question is from Turan, who's Bayer, and if I'm not pronouncing your name, Correctly, I apologize. You highlighted Exascale's asset-light orchestration model as more capital efficient. How do you think this approach positions the company relative to traditional data center owners in terms of scalability and risk management? Rafael, you could take that or you could pass it back over to me.

Speaker 1: On this one, I would say that for traditional data centers, so you need to commit your capital upfront. You need to wait years for the build and then you hope that the demand materializes at a scale that you plan for. So there's potentially this mismatch because if it doesn't, then you're already sitting on this stranded CapEx. So the orchestration model, what Exascale is doing, they flip that. So in this case, Exascale scales with demand rather than ahead of it and you're really not carrying this hardware depreciation risks with the different generations of GPUs. So you can deploy into new geographies without the full large capital commitment of a grown up view. Mark, if you could add more to that.

Mark: Yeah, I don't have much more to add to that except for our model. You know, we kind of leverage other owners of these assets because sometimes we'll just come in and take over and manage it more efficiently. Other times, you're committing upfront capital, but it's nothing like a major data center, a modular data center that costs a fraction of what these other data centers cost. And you could just drop ship them right in and plug and play. Remember, everyone that's come in now and listening to this, it's not that simple, but four to six months versus two to three years for a data center, I'll do that every day. All right, next question. I'm going to throw this back over to you, Rafael. You noted Tyler Altons has asked a question. You noted strong potential in Asia markets, particularly Southeast Asia. What macro trends or demand signals are you seeing from Asian AI companies that make this region particularly attractive right now?

Speaker 1: I'll say that for these in Southeast Asia in this case, and what we're broadly seeing, it's there's the different Southeast Asian countries, their own domestic AI adoption, it starts accelerating from enterprise AI deployment, government digitization initiatives. I think you've seen in the news as well, regional tech ecosystems, and they are all, we can see them maturing here. So that to me would be a long-term demand layer. So of this building underneath the whole hyperscaler externalization story. But for Exascale specifically being positioned in this region before the incumbent Western operators, they arrive. Yeah. So as I mentioned, this is really the first moving dynamic that we are really looking for. And I also mentioned just now that this window, it doesn't stay open indefinitely.

Mark: Yeah, for sure, especially now. This one coming from Nora. You mentioned the importance of long-term off-take agreements with prepaid terms. How do these contracts change the risk profile and predictability of revenue in this sector? Well, for me, I see it as more predictable. It's almost like a software subscription, if you think about it, because these long-term agreements come with maintenance and service, and it's very hard to swap teams out in this field, especially having the hardware and infrastructure knowledge that a company like Exascale Labs has. I think the risk is greatly reduced. We're not just a construction company going in and building and exiting stage left. We're actually sticking around to run it. And so the predictability of revenue is, I think, much easier. Think of it like a software subscription, not in all cases, but in a lot of cases. And then obviously, as chips like the Verirubin chip come out and what's the next generation after that? What's that going to require? It requires a lot of planning, a lot of expertise to make sure that you can run those chips because those most data centers, I'd say 90% of data centers today can't run the next generation of Verirubin chips reliably or most effectively or most efficiently. And then they're not going to be able to run the second-half, or without getting too technical, the higher level of birubin chips that require much, much more power. So all these things factor into a maintenance and services agreement that will last years and years, which make for predictable revenue. Okay, I like this question from Arif Tomur. For people new to infrastructure investing, what's one key mental model or framework you'd recommend they adopt when thinking about AI infrastructure opportunities? I'll pass it over to Raphael if you want.

Speaker 1: All right, I'll try. So in my thinking, I would say that the mental model I'll use, it's back to the AI. The title of the AMA would be to really think about where the bottleneck is, to find the bottleneck and own the bottleneck. So in the early internet days, it was bandwidth and then cloud computing, it was data center capacity. And now with AI, it's the physical infrastructure. Again, it's the power, the cooling and the compute. it's important to really control this access to the scarce layer. And having the control of this, it will allow you to really capture the disproportionate value to this area. So when thinking about really any infrastructure opportunity, I always ask myself these two things. First, if they are sitting on this bottleneck, number one, and number two would be How long? How sustainable is it? How long does that cost? Now, of course, if both answers are yes and there we go, that's where I want to be.

Mark: Excellent. I do want to note as a follow-up, Gator Nation brought this up and reminded me to talk about this. They asked, Do you plan on only focusing in Asia or do you see your service going global? No, the service is global. We're primarily in the US. We are expanding into Asia. We do have Asian customers. We see a big potential in Asia as to everybody else, Europe as well, the Middle East. Now we're open for business everywhere because Taisu is based in Southeast Asia, but they too invest in US corporations like ours. We are bringing up some of the Asian side, but certainly global, most of our businesses here in the US. But thanks for reminding me to talk about that. Last question, and this one I am going to throw towards Taisu, because Blankman is asking you directly. Taisu emphasized the team's operational depth. I'm sure he means exascale. So what specific capabilities or mindset do you look for in other founders and operators when evaluating AI infrastructure companies?

Speaker 1: So thank you for this question. I would say that Two things, they are non-negotiable for me personally in infrastructure. So first one is to be really honest about the constraints. So if the founder can only talk about the upside, that would be quite concerning. But the ones who can clearly articulate what can go wrong, how they're managing it, I think these are the ones that I can trust that they will execute well. Second would be the operational obsession versus over like the narrative. So in infrastructure, there's always this great pitch that we always hear, but versus that with a great execution, sometimes there's this wide gap here. So what I would want to see from founders would be those who think about uptime, common lead times, contract conversion, and not just talk about market size, it's huge, it's growing and all that. And when we find that, find these both together, that's when we get the conviction and also, well, why we invested in the exascale.

Mark: All right. Well, with that, we're at the end of our time today. I want to thank everyone who attended. Thank you for the questions. Great questions, by the way. I want to thank Rafael from Taisu. They are one of the leading investment firms in South Asia. And I hope you guys were educated as to what the real bottleneck is out there in AI and listening to a different AI perspective. We're all seeing the sexy things AI is doing. But, you know, a lot of us, including me, did not understand what powers it all and why it's the most critical layer in AI. So again, like, follow, please follow Taisu and Exascale Labs if you're interested in this continuing saga. And follow our timeline. We have a lot more updates for you, a lot more things to come. Raphael, any final words?

Speaker 1: No. Thank you. Thank you for your time. Thank you, everyone, for listening. Yeah. Thank you for the opportunity for this AME and also for the opportunity to invest in you guys.

Mark: My pleasure. Thanks, everyone. Until next time, we're trying to host a lot more of these going forward. Thank you, guys.

Speaker 1: Thank you.