How Asset-Based Finance Is Powering the AI Infrastructure Boom

Anu Rajakumar: Picture the AI buildout, and you probably think of chips and code, but underneath all the silicon and software sits an extraordinary amount of physical infrastructure: data centers, fiber routes, and the power plants to run it all. Hyperscalers and neo-clouds have guided more than $750 billion per year in AI and data center spending in both 2026 and 2027. The scale of the buildout has created huge demand for equity and debt financing. Increasingly, a portion of that financing looks less like traditional corporate loans and more like lending against the assets themselves.

The interesting question for a lender is how to earn an attractive return that's backed by real tangible collateral without having to bet on which technology ends up winning. My name is Anu Rajakumar, and today I'm joined by Sean Hinze of Neuberger's Specialty Finance team. We'll talk about how asset-based lenders participate in the AI buildout, where the value is and where the risks lie. Sean, welcome to the show.

Sean Hinze: Thanks, Anu. Great to be here.

Anu Rajakumar: Sean, let's start at the top. When most people hear AI investment, I'm sure many of them are thinking of the NVIDIAs and the OpenAIs of the world, but you are coming at this from a different angle, the financing side. Set the table for us. We're talking about hundreds of billions of dollars a year flowing into infrastructure. Why is this, at its core, a lending story?

Sean Hinze: Sure. First, maybe we'll take a step back and talk about why in the world are they spending $750 billion a year on this stuff?

Anu Rajakumar: It's a good place to start. Yes, I agree. [laughs]

Sean Hinze: Look, Google remembers what happened to Yahoo. Meta remembers what happened to MySpace and Friendster. Microsoft remembers what almost happened to it. Search is over. It's AI for Google. From social engagement, similar for Meta, and Microsoft knows they have to be all over this with their productivity suite or someone's going to take it from them. They cannot be last. They see this as an existential battle to remain one of the Fab Seven or whatever moniker we want to use. For them, it's existential, so they're going to spend three-quarters of a trillion dollars trying to make sure they stay on top of the pack. That's what's driving this.

Now, directly to your question, why the heck don't they finance these on their balance sheets? It's $750 billion, and it's going to happen again and again and again. They don't want to fund that with equity, and even though they're investment grade, they don't want to fund a lot of that with debt.

They know it's more efficient to finance these off-balance sheet, and there's financiers who would be happy to finance this at a lower cost of capital, because if it's in its own special purpose vehicle and for some reason they don't want to pay, we can just take that asset and put it somewhere else. We'll talk about what asset-based finance is, but these assets, their value is not necessarily tied to who's operating them. We can take these assets, resell them. They can be used all throughout the economy.

Anu Rajakumar: Actually, maybe on that note, why don't you give a quick explanation of what asset-based finance is? Specifically in this context, when we're talking about physical and financeable assets, what does that actually mean in this space?

Sean Hinze: Sure. Pre-crisis, the banks just financed everything. They financed the vanilla ice cream and chocolate with sprinkles. They did everything under the sun, and they were levered 30 times. We all found out that didn't work. Post all of the Dodd-Frank and new regulations, the banks are in the moving business, not in the storage business anymore. They can't finance all of these assets anymore.

At the same time, the direct lending really started up. That's a $2 trillion market. That's just lending basically to sponsor-backed companies. Basically, over the last 10 years, private credit markets have realized, Hey, we can finance all of these assets that consumers and companies run into on a daily basis more efficiently on our balance sheet and price that risk appropriately while also giving the borrowers the flexibility they need to complete their projects and do what they're trying to do. That can be airplanes, it could be infrastructure, it can be cars, it can be credit cards. Anything that you touch in your daily life, that's basically asset-based finance, and that market's probably about a $20-trillion-plus market.

Anu Rajakumar: Great. What we're speaking about today, you're talking about chips and power and fiber, et cetera, right?

Sean Hinze: Right. On the digital infra side, that includes GPUs, that includes data centers, that includes the power and the electrons that run those data centers. It includes the fiber that connects all that data throughout the world, as well as cell towers, et cetera. Anything that runs the digital economy, that's what we're interested in financing.

Now, look, that traditionally has been cell towers, fiber, and data centers. When it was just data centers, it was just like high-performance compute. We're going to move all our files off of the co-located server to the cloud, but now that has been completely eclipsed by the advent of AI and these large training models.

Anu Rajakumar: I do want to make it tangible for our listeners here. Without naming names, can you walk us through what one of these deals actually looks like in practice and talk us through the structure, collateral, and the return expectations?

Sean Hinze: Yes. We were part of a large GPU financing deal where we financed 55,000 Blackwell NVIDIA chips for a large neo-cloud. The thing that made that deal unique and so attractive was the offtake. Look, at the end of the day, I can underwrite the value of a GPU, but NVIDIA is cranking out new GPU models every two years, and honestly, if they could go faster, they would. It's really hard for me to sit here and put a price on an NVIDIA GPU and have confidence in that enough to risk my investor's capital. What I can underwrite is multi-year contracts to strong counterparties. That's something I can underwrite over time.

With this deal specifically, we had about 80% of the offtake contracts were multi-year contracts to investment-grade, high-quality hyperscale partners. Then the other 20% was, I would think of it as kind of a BB high-yield AI lab. The underwrite on that is actually pretty easy. You just assume the non-investment grade company goes bankrupt immediately, jumps right to default, and does the loan pay off? I think that was really attractive for 5% over the relative benchmark in terms of return, getting effectively investment-grade risk for 500 basis points of spread, and that's great real val for my investors.

Anu Rajakumar: Absolutely. That's very helpful. Sean, power keeps coming up in a lot of the conversations around AI. You can have the best chips and the models, but really, without reliable and abundant power, frankly, the data centers don't operate. From a lender's perspective, why does power matter so much, and what makes it interesting as collateral?

Sean Hinze: Sure. As this gets built out over the next 5 to 10 years, the bottleneck is shifting. Initially, the bottleneck was chips. I can't get NVIDIA chips. That's more or less been solved, but now the new bottleneck has moved on to electrons. Where can I get enough power? You have to have power to run these chips.

What's happened is the utilities were a little bit asleep at the wheel. Elon Musk said this about two years. He's like, "Guys, you got to produce way more power," and they were like, "What are you talking about?" Turns out he was correct. The US grid is effectively tapped out. There's very little power that's going to be able to come from the grid. Also, the lead times for utilities to build new power generation is like multi-years.

Now, the bottleneck and the solution is to actually build power, what we call "behind the meter." These data centers and the developers and the hyperscalers are scouring the planet looking for generation assets to build effectively their own power plants. We're not talking like a few megawatts. We're talking about gigawatt-size power plants. To put that in perspective, the island of Manhattan requires, as a baseload, about 3 to 4 gigawatts. There are multiple sites out here that are planning 1, 2, 5 gigawatts of power generation just for these AI assets.

Why do we want to finance those assets? Look, again, we talked about GPUs. It's hard to know what a GPU is worth in one or two years, given the product reintroduction cycle. However, on the power asset side, the engineering is pretty known. If you're doing a combined cycle gas turbine, There's not going to be some new technology out there to generate electrons more efficiently. I'm happy to lend against that technology for multiple years.

Further, on top of that, you have a structural supply-demand mismatch. These are incredibly complex machines. This isn't stamping an iPhone together. What that does for me is that tells me the supply of these assets is not just going to flood the market and destroy the value of these assets over time. We have a massive demand function, but the supply function is able to grow only so fast. That tells me you're going to have relative price stability over the next 18, 24, 36 months.

Anu Rajakumar: I was going to say, and isn't there a projected shortfall in the number of gigawatts? Will these assets be able to be built quickly enough to be able to avoid the shortfall?

Sean Hinze: Sure. When you look at it, I think it's 68 gigawatts short in the US alone in terms of actual demand, and that's with all these projects that are supposed to be getting built that are already online. That just tells me we're structurally short a ton of electrons, and that means there's just so much demand for these assets that we're going to be very safe lending on these.

In addition, there is no secondary market for these assets. People are literally pulling turbines out of Thailand that had been in storage for years and then shipping them over the ocean, which costs tens of millions of dollars to put them back into operation. You can't even get used equipment easily. I don't know where that lands in five years, but I think I can make a loan that's pretty attractive over the next two to three years.

Anu Rajakumar: That's super helpful. Thank you. Now, I think a lot of listeners are probably wondering, are we financing something durable, or are we in a bubble? What happens if someone cracks the efficiency problem, and suddenly that ends up being somehow that you actually need far less infrastructure to run these models? How do you respond to those kind of concerns?

Sean Hinze: That's a great question. We actually saw a hint of that with the DeepSeek movement that we saw about a year and a half ago. Look, DeepSeek came out, and it offered basically productivity gains of a little bit less than 2x. The problem with that theory is the effective compute demand is still growing at four to five times per year, so it's massively outdoing any gains from any huge one-time blips that we saw. We could have multiple DeepSeeks just keep happening in terms of innovation, and it's still not really going to put a massive dent in the overall growth of the compute demand out there.

For example, the models that we're running are still requiring 3 to 10 times more data each iteration every 6 to 12 months, and scaling laws are holding. In other words, what's happening is the more data we put in these models and the more parameters we put on them, they keep getting better each time, materially, and that hasn't slowed down yet. If you look at what the hyperscalers are planning in terms of their roadmaps, these gains are all largely embedded in '26 to '28 roadmaps, so we know the spending is still going to come over the next 3 years.

Then some people like to say, "What about massive improvements on the chips in terms of electrical efficiency?" Look, I would point to Jevons paradox from the late 1800s. That's static analysis. What actually happens here is as the price to run some mechanism or some technology comes down over time, you would think, "Oh, okay, then the price for that will just go down, and we won't need as much." That's what they thought with steam engines and coal in the 1860s in Britain.

Turns out when you get that price down, it stimulates demand on its own, and so even more demand and more compute is required. That's what we saw with coal, that's what we saw with electricity, that's what we're seeing here now with compute. Last year, we were talking about just running chat queries, and now it's agentic workloads, and it's inference, and it's only growing from there, and it's building on top of itself.

Anu Rajakumar: Now, it does seem like there's a fairly wide spectrum here. There's the fiber, the chips, the data center construction, the power equipment. Where are you seeing the strongest value today, and are there areas they're actually feeling a bit more cautious about?

Sean Hinze: Look, I think it's interesting. You can always find complexity and unique spots where you're going to find specific alpha, but I think broadly right now you're still seeing pretty good value in the chips, and really it shouldn't just be called chips, it's really chip contract financing. You're able to get very strong counterparties with multi-year contracts that have to pay no matter what, with contracts at a data center that I can step in if I have to repossess, and I could definitely sell those chips if I had to, or that compute. I think that's still at an attractive level.

Where I think that's going is, look, the amount of financing that needs to happen is actually going to overwhelm the banks. The banks are in the moving business now. They're not in the storage business anymore. Even with what they can do, private credit is going to have to be a solution. Just the absolute supply of financing need is so much that we're probably going to see pricing stabilize because there's just so much to do, so we can hold firm on price there.

Then other areas I think that are attractive is also power equipment, which takes-- It's actually similar to investing in aircraft. It takes specialized knowledge, takes specialized structuring, but they are economically essential assets even away from the AI buildout. Look, if I have a generator, I can run electrons obviously for data centers for AI, but I can run it for oil and gas, I can run it for utilities, I can run it for various different things, and there is no secondary market for that right now. I think there's good rel val when you can find a dual use for something outside of the AI buildout itself.

Anu Rajakumar: Absolutely. Now, as we wrap up here, Sean, I often like to ask practitioners this: for an investor evaluating this specific space, what do you think is the real edge, and what questions should a prospective investor be asking before they write a check?

Sean Hinze: I think there's a lot of hype and there's a lot of people just jumping in holding their nose without looking at what they're doing. I think structuring expertise, also knowledge and domain expertise in the space, and knowing who the right players are, is very important. I must have seen a dime a dozen of chip GPU financing deals, and a lot of investors are just trying to do as much as they can, and they're just going to give investors beta, but it's going to include all of that lower quality type of offtake.

I think being able to be choosy, be patient, pay attention to the structures that make the most sense, and also just knowing who the right players are in terms of credit quality and offtake is going to be very important. I think there's a lot of people running around just trying to get some exposure here, and there's always going to be projects that are going to be built on spec and not on demand, and those are the guys that are going to get exposed in the end.

Anu Rajakumar: Absolutely. We were saying with this entire AI rollout, there will be winners and there'll be losers, and the key is selectivity. Sean, thank you so much for all of those comments today. I can't let you go without a quick bonus question. We are recording this episode right in the middle of the 2026 FIFA World Cup, which is very special for me as a big football or soccer fan. I'd love to know, do you have a core sports memory that you can share? It could be attached to a favorite sports team or a personal story if you are or were an athlete, but what's a core sports-related memory that you have?

Sean Hinze: Sure. That's an easy one. I grew up in Texas. I went to school in Texas. I was an engineer in Texas. Watching the University of Texas get beat mercilessly by the USC Trojans, and then Vince Young single-handedly run into the end zone like a gazelle at the end of the fourth quarter was absolute magic. That's definitely the top.

Anu Rajakumar: Great. Lovely. Thank you so much for sharing. We had a great conversation today. I'm going to briefly try to summarize a few of the key themes that we chatted about. You articulated that the AI buildout is enormous. I think you said $750 billion a year and counting, and much of it is, in fact, physical, financeable hardware. We talked about how, for asset-based lenders, the goal isn't necessarily about picking the technology winners, it's lending against real contracts assets and collateral.

We spoke about how power is particularly important given the 68-gigawatt shortfall that we discussed. You also mentioned that your team sees the strongest value in short-term collateral-backed loans to high-quality customers, particularly in chip contract financing and power equipment. Finally, you talked about how there's a lot of hype in the space. It's really critical to find partners that have domain expertise, and as an investor, be selective, be patient, and really focus on high credit quality to avoid some of the potential landmines out there. Sean, I hope I adequately summarized some of your comments today. I just want to say thank you so much for joining me.

Sean Hinze: Thanks, Anu. Thanks for having me.

Anu Rajakumar: To our listeners, if you've enjoyed what you've heard today on Disruptive Forces, you can subscribe to the show from wherever you listen to your podcasts, or you can visit our website at nb.com/disruptiveforces, where you can find previous episodes as well as more information about our firm and offerings.