Can Infrastructure Keep Pace With AI?

Good afternoon. Thank you very much for joining us in this conversation. I know you've just heard a lot about artificial intelligence, right? I mean, just us here — we haven't been able to resist the temptation to use ChatGPT, Gemini, or something like that last week, right? So, are we talking? Yes, we're talking enough about AI, the demand, the deployment. Are we talking enough about the physical infrastructure needed to keep up with what's happening with AI, the compute capacity, and what will be needed in the future? So this is a perfect combination of people because we can talk about everything that's really needed in that infrastructure chain — from deployment, data centers, energy, capital, talent, everything. So let's start by getting straight to the point. The bottlenecks right now. I'd like you to start because you're also a major data center operator worldwide — 150 facilities in 20 countries, and the largest operator here in India. Where do you see the friction?

Well, first of all, it's a pleasure to be here. Regarding your specific question, let me take a step back and explain where we are in the cycle. I think we're barely scratching the surface when it comes to demand, right? So if we look at current token consumption, we have 30 million software developers, and a fraction of that is where it's mainly concentrated, along with some research work. We have a billion knowledge workers. We're barely scratching the surface. We have about 2.5 billion frontline workers, like nurses, factory workers, etc. And all of that will be positive in terms of AI, and we haven't even talked about consumers yet, right? As for what consumers will do, the new business models, etc. So to frame it in terms of where we see the demand equation, I think there's a lasting secular trend in what's the art of the possible here. Now, the biggest obstacle ends up being the data center capacity you can sustain, right? And the long pole in the tent, of course, is energy. But beyond energy, it's a systems problem, because as a data center operator you need to bring local power if there's no grid connection. As an example, grid capacity in the United States has remained flat over the last three decades. China has doubled every decade over the last three decades, right? So grid capacity is a problem. Then you enter a data center site and we lose a lot of energy doing an AC to DC conversion. And then the mechanical and electrical equipment installed in data centers — generally for us, we're one of the largest in the world — and even so it's running on about 42 weeks of lead time, right? So there are obstacles everywhere, and then the biggest obstacle nobody talks about is the hidden one: talent. So if you think about electricians, skilled technicians, design, construction — I mean, an electrician currently earns about $200,000 in the United States. That's more than many office jobs, right? So there are many players who can bring that whole ecosystem and orchestrate it all — that's where I think the recipe is. Capital is not a problem. I think we can talk about that.

We couldn't agree more. In short, capital is not a problem, and neither is demand. The bottleneck is who can deliver these projects. Perhaps the biggest narrative we see around the world, and which is somewhat removed from what we're seeing on the ground, is that somehow AI demand for compute capacity, for energy, for the supply chain that backs AI — is this demand somehow fragile? You know, we hear about a more efficient model and people worry about the demand outlook. We hear about a data center being released and people worry about the demand outlook.

Demand is outpacing supply so much right now that there will be a supply-demand imbalance until the end of this decade and probably well beyond. And what we're really seeing around the world with that truly durable demand profile and the growing scale is that it's not about who has the capital. It's exactly as was said — it's about who can combine capital at scale with operational and industry-specific expertise at scale to deliver these projects.

And as Connor said about demand, current frontier models cost around $500 million per run. Before, if you had to run a three-year-old model, today it would cost four or five million dollars. That's how fast they're scaling. And innovation must continue. Abhijit and I were chatting — we're probably not even in kindergarten yet when it comes to where demand is heading. And this is what he was talking about regarding knowledge workers before physical AI — we're in a very embryonic phase of its development. Demand is going to be stratospheric from where we are today.

So how do you think that will affect Helman, for example? Because basically it's building the pipelines of AI.

Look, I think this is what Connor said — keep building. Demand is not a problem. There will be stumbles. And the stumbles will happen because there will be areas, locations, and specific aspects where people might build ahead of demand. It's happened in every infrastructure project. I was talking to someone about how if you build an eight-lane highway for a town of 50 people, it won't be used. That kind of construction, that kind of investment, will get stuck. But if you're sensible about investing, demand will be far ahead of supply for the coming decades.

I want to ask you how you differentiate that, but before I do, Sandeep — because Sanjeev keeps talking about demand and, of course, he's also providing AI deployment to his enterprise clients. What are you seeing in terms of how people are actually adopting artificial intelligence?

Does that affect infrastructure building?

Yes, it expands.

Yes, I think it goes in one direction. In the same way that users are becoming more efficient in using the right model for their use case, infrastructure is also becoming more efficient. You talked about how the grid isn't going to keep up with the amount of generation that needs to come online and transmission. We're seeing much greater construction and acceleration of behind-the-meter solutions to address energy demand needs and AI infrastructure demand needs because the market isn't going to wait for the grid. So infrastructure building will become more fit-for-purpose in the same way that model usage will become more fit-for-purpose. And, by the way, ironically, AI is also helping with that. So you now have demand-response mechanisms deployed in utilities that can manage not only AI workloads in local data centers but also local community demand, right? So especially when you're doing inference, because remember, training follows power, inference follows consumption, follows users. So inference needs to be near urban areas, for example. Oh, sorry, inference, yes. So now we're seeing a lot of innovation across the whole stack, right? So whether it's off-grid, fuel cells, I think we'll also see nuclear power.

Yes. If I can jump in, because in that last sentence you stole half of my favorite line — because the energy demand environment is so significant right now that it will require any or all of the above solutions. Every form of generation has demand right now, but we see four forms of generation capturing a disproportionate share of future demand growth because each has a characteristic, a right to win, and something the market needs. Fuel cells because they're the fastest to deploy, renewables because they're the cheapest, batteries because they stabilize an increasingly volatile grid, and nuclear power because it provides baseload energy security. Everything will grow, but we believe those four will grow the most. I think by 2030, our forecast suggests that at least a third of data center capacity will have on-site, off-grid generation. Much of the work in the U.S. is behind the meter. The predominant part is behind-the-meter product. And another thing we're seeing is that demand is arriving — and then I'll come back to the energy question — really at the edge. And this is just getting started. But I can say that in the last quarter or two, edge demand didn't exist at all. Two quarters ago, we're only now seeing orders to build 1 MW and 2 MW data centers at the edge, and we never thought those days had ended. But I think much of the debate centers on data centers. I think the biggest debate and the biggest challenge is energy. Yes. And energy infrastructure — whether it's transformers or the grid. In most cases, I think going forward, excluding China, you have to write off the grid because you shouldn't rely on it almost anywhere in the world. India is a bit unique because it has a national grid. But places like Europe, like the UK, which just announced a major grid investment, the United States and Germany have a generation problem — they simply don't generate enough electricity. France has enough, but its grid is old. So in every country, I think you have to write off the grid, as the existing grid is part of the solution.

So the alternatives — do we have enough right now? I mean, you've listed them all, but for example, even nuclear power, there are a lot of mixed feelings about what that means in many countries. How do you overcome that?

I think your question... I'm sure we all ask ourselves some version of the question: will there be enough? What is enough? There certainly won't be enough supply to meet demand. I think that's clear. I mean, we have about 200 gigawatts of data center capacity today. If that tripled to that number tomorrow by magic, we'd use it all. And we were talking with Conor in the green room about where the real risk lies, right? If, for example, you have excess capacity, if for some reason there are stranded assets. I mean, we saw the dot-com bubble, for example. And you were saying, 'No, there's enough demand out there.' But yes, someone has to pay at the end of the day.

So there are two things. I think you're absolutely right. There will be stumbles. We all come here and say demand far exceeds supply. A very simple answer to that would be to simply invest all the money possible because everything is going to be productive. That's not true. I'll tell you from Brookfield's perspective — we see this enormous investment opportunity as a situation where we can be incredibly selective and focus only on the best assets, in the best markets, with the best contracts and the best corporate credit counterparties. And we can still invest as much capital as we want in opportunities with very attractive risk-adjusted returns. But you've highlighted an important point. One thing that's sometimes underestimated about this generational AI build-out is that it's like any other infrastructure build-out over time, except that it's at a larger scale and happening faster. The other thing that makes it very, very different is that the vast majority of the infrastructure being built is being built on long-term take-or-pay contracts with the best corporate credit counterparties in the world. If they're the best corporate credit counterparties in the world today, it probably means they're the best corporate credit counterparties in history. The biggest drivers of this infrastructure build-out are the major hyperscalers from both the United States and Asia. And in places where organic hyperscaler demand isn't driving infrastructure build-out, we're actually seeing the sovereign step in and backstop. So unlike what everyone compares it to — fiber in the dot-com bubble — much of this build-out is backstopped by long-term take-or-pay contracts with the largest corporate credit counterparties in the world today. There's one more difference, if I may. Data center capacity was already in short supply before the ChatGPT moment. Because remember, we've been building data centers and our capacity has grown at more than 15% per year over the last decade. You know, streaming, payments, public cloud, etc. So now AI infrastructure is being layered on top. I mean, that's more than double, right? So that challenge becomes more acute.

Let me ask Sanjeev — is that also what you saw when we were talking earlier about differentiating between projects that work and those that don't? Contract structures, the sponsors.

But let me go back to what Connor was saying. We were present when the fiber build-out was happening in the U.S. We supplied, but we also provided engineering or project management on some of the largest fiber builds. The fiber build-out was happening in anticipation of demand. I used to have debates with the guys building Qwest and Level 3. They were all my peers. I still know many of them. The basic thesis was: let's lay the fiber and demand will come. Yes. Now it's very different. The demand is there, but it's not being met. Let's get data center capacity out. The biggest advantage, and what really makes it economic, is that the other side has to be right. Where the challenges arise is not so much in cost, but in lead time. And that's where the supply chain breaks. He says it's 42 months to get that transformer — 42 weeks to get the transformers. He's lucky. Because he doesn't have... you know, he's not talking about... yes, and he's not talking about full averages, he's talking about... But there are projects where it takes a year and a half to finish all of that. That project goes. If it's over budget and doesn't meet client demand. And that's the challenge. The supply chain — at the same time as that, I think one of the points is that we've taken at least one global supply chain and split it into two parts. One is focused on the U.S. and the other is focused on China.

I was going to say, how does geopolitics add to all this friction and imbalance in the world right now?

Completely — you have two supply chains. There are ongoing wars, you have rising costs, that's another set of problems. Before addressing the war problems, you have to manage a supply chain built over six or seven decades. Breaking it apart and saying I have chips I can only use for Chinese hyperscalers, chips I can only use for the other hyperscalers. Every part of the supply chain is different.

Go ahead, sorry. I want to give the floor to Lior in a second, but you asked how geopolitics and conflicts are not only affecting supply chains, but are also reinforcing a demand dynamic in which — I don't want to oversimplify — the world is becoming increasingly digital. Nobody disputes that. Data is now the world's fastest-growing product and holds national strategic importance. It must be stored, processed, and transported at a level never before seen. And it's increasingly important for companies, but particularly for governments, to have security over their own supply chain to manage and store their data. This is a national security interest, and therefore we're seeing demand environments across all regions of the world. Every country wants a certain level of its own AI infrastructure. And this brings us to the point that if it's not going to be provided by corporate demand, we're seeing governments step in.

Perfect, because we're seeing that also developing in India.