Banking on AI Infrastructure: Risks and Rewards of Data Center Lending

As AI drives unprecedented demand for data centers, financial institutions face a complex lending opportunity. Edwin Anderson of Oliver Wyman explores the unique risks of financing hyperscale facilities, from construction and power constraints to lease structures, community opinion, securitization, and long-term uncertainty surrounding the future direction of AI.

Subscribe to ProSight Podcasts: Apple, Spotify, Amazon Music, YouTube

Don’t miss ProSight’s Annual Conference in November, which will feature the session “Data Center Financing: Lending Considerations in the Age of AI Infrastructure.

TRANSCRIPT:

Frank Devlin: This is the ProSight Banking Strategies Podcast. We’re here to inform you on the top trends, challenges, and opportunities in banking today. ProSight is a leading non-lobbying connector of people and information with deep expertise in risk, fraud, compliance, and retail and commercial banking. Our purpose is to empower financial services leaders to strengthen and advance our industry through training and insights, as well as tools and resources like this podcast.

Hello and welcome to this ProSight Banking Strategies podcast. I’m Frank Devlin, a senior editor at ProSight. Data center lending is emerging as a major growth opportunity for financial institutions as billions of dollars flow into AI infrastructure. Yet these projects often look different from traditional commercial real estate investments, bringing unique considerations around construction, technology, power, and long-term demand. This episode features a conversation with Edwin Anderson of Oliver Wyman, a partner in OW’s banking and financial services practice. He’s going to discuss the particular challenges of assessing and managing the risks of financing hyperscale data centers. Let’s get started.

So Edwin, big picture, how are the risks and opportunities of lending into these new hyperscaler projects different from financing previous data centers from years past?

Edwin Anderson: I think it’s useful to think about data centers across a spectrum, ranging from… When I first studied data centers, it was almost 20 years ago, and many of the projects were fundamentally small. They had no meaningful water or really power requirements compared to other more power-intensive commercial real estate. And they were fundamentally drawn across many, many different users and in some cases even specifically separate tenants. So in a way, they acted more like traditional commercial real estate or, even kind of a close analog, cellphone towers where people often rent the bays.

Now we’ve moved, and of course there are intermediaries between this, but at the far end of the spectrum we have single-tenant hyperscalers and they are specially built buildings, unlike many of the older buildings that could easily be converted back to warehouse or office or many other uses. These are incredibly specialized buildings requiring specialized power and water infrastructure, often their own substations, and they have a great deal of correlated risk.

The odds that a bridal company and a craft store that are both using space and an old-school traditional data center all vanish at the same time and hit the usage is very different from cases where we have large companies making bets on a very particular AI outcome. In addition, these large data centers put stress on power and water and  local municipalities in a way that almost no old centers did and raise kind of fundamental questions about power delivery and in some cases strain on water systems, especially for smaller markets. And finally, you get to construction, the specialized labor, the sheer volume of specialized equipment and power comes back. And then finally, these are often being financed in a number of unusual ways. So there’s an enormous, enormous amount of change. I just want to emphasize those technical portions just require a level of specialized knowledge to underwrite them appropriately.

There’s a broader issue that is worth thinking about in that the U.S. has its largest companies leaning pretty heavily into what I would call the brute force answer for AI where fairly conventional modeling approaches are being used at more and more intensity to attempt to improve the product and that the question of whether alternate answers become more appealing in the long term. So for example, much of our AI expansion is being done under enormous subsidies from equity. And as products move back to true prices, if the world suddenly wakes up and says, “I don’t want to pay that much for a frontier product, I get a good enough answer from a public domain product or a lower-end product, aka the DeepSeeks of the world,” that’s bad for data centers. Those can both be done with immensely less resources. There’s even kind of bets on the perpetual growth and improvement of chips and hardware.

One of the biggest issues with some of the interesting alternative models for data centers, things like putting them under the ocean, were maintenance issues in a world in which, and part of that was turnover in a world in which chips and equipment are constantly improving, you don’t want a low-access model. In a world of stability, those are economically immensely more appealing. And look, I wouldn’t put large percentages on any of the cases I’m giving, but there are many of these other cases, different AI futures we could have, and they raise many questions about exactly what bet is being made. I would not bet against AI. I think that’s a very bad idea, but that’s often confused with betting on essentially the US-centric brute force approach.

Devlin: You’ve compared lending into a big data center in a previous discussion we had to lending into a toll road project. Can you explain how that’s so and what are the special risks and opportunities make this sort of analogy?

Anderson: The context there is that a lot of early lending on toll roads was done like they were simple cash-flow assets when there was an immense amount of complexity and technical understanding needed. Some of the early lending at institutions was done by handing it to people who didn’t have this specialized background. It’s understandable this wasn’t a large area, but often it did not involve the ramp-up needed in complexity. You can simply look at these and go, “Look, this is a lease from this large company.” I know how to underwrite a lease from a large company but understanding things about the kinds of bottlenecks that could happen in construction, understanding the sorts of clauses in leases that allow those companies to exit them, which are quite different from those that an office occupant has for being able to terminate their office. I mean, they’re overlapping ones, but there’s plenty of difference.

It’s also in understanding things like the construction plans and the pipelines to them, even subtle issues about appraisals, there’s an enormous amount of depth. I think that there’s reason to be cautious about getting caught off sides if your personnel aren’t adequately upskilled. I think toll roads are an interesting comparison because there was a lot of underwriting of them on as simple cash flow, but in a way they were a simpler case because often you really had to understand the underlying traffic model.

And if you didn’t understand the underlying traffic model, a few of the toll roads went very badly and it was often tied to essentially a misunderstanding of these technical documents, how to think about them and how much to trust them, and it led to some toll roads turning out very badly. Now to be clear, some of them have turned out great because there aren’t always problems in those technical documents.

But of course, when you’re in credit, when you’re in equity, you can kind of play the averages a certain amount, but when you’re in debt, when you’re in lending, you have to be very mindful of your understanding of that technical complexity.

Devlin: Talk about the toll roads a little bit more if I could. Did they not do well because the pricing was not right, because they overestimated people using different alternate routes, that sort of thing? What went wrong there?

Anderson: In some of the more famous cases, it came from traffic studies, I think were funded in intriguing ways where the writers of the traffic studies had great motivations to give very positive answers and it created biases in the system and there was a huge agency problem because the people organizing these deals were not the people who were long-term going to hold them. Most of them passed through to bonds or in some cases even became problems of the municipality. And so there was a lack of alignment and there was an agency issue, which is I think rearing its head here as well.

So the agency issues that showed up there and the technical understanding issues, they’ve gotten better over the last couple of years at almost all institutions, but a lot of underwriting was done at a fair number of institutions where the gap was excessively large to start and the closing of that gap has been uneven across the industry.

Devlin: The big companies are now preferring to lease. Maybe in days past that wasn’t the case. Do you agree with this or is that too simplistic of a view? And if so, why is this changing? Is it just the fact that like any kind of business, you’d rather not have your money in real estate, you’d rather have access to the liquidity to use in different ways? What has changed?

Anderson: I think that there has been some change, but of course all these companies are individual entities that are acting in their own ways, and so I’m always a little cautious about painting with too broad a brush, but there’s definitely been a big shift and I think it’s partially on the heels of a broader shift that’s happened over a longer timescale. And so there’s been a move to asset-light models and clearly companies that are innately asset light have a certain strength to them. They have a certain flexibility. We don’t care what office we’re in, we can move. Owning their real estate may be a terrible idea. We don’t need assets to produce our goods. We’re a law firm, we’re writing code.

Those are all true asset-light. We’ve started getting proxy asset-light. The companies that are incredibly dependent on their assets but choose to be asset light because I think of flaws in how the markets often analyze these companies because a company that needs its real estate in the long term, that needs its assets in the long term, if they’re truly a long-term need, shifting them off balance sheet creates a weakness, not a strength.

And some of the moves to asset-light models, if you remember some of the PropCo/OpCo splits went very badly because in the end it was gaming numbers that the markets look at rather than asking fundamental questions about the health of the company. So now we’ve reached a point where companies that have traditionally been fairly asset light that are about writing code and about essentially producing intellectual assets now are deeply dependent on hard assets for major components of their business. And that creates a tension where there are both. There’s certainly some practical benefits if you feel like the owner or operator of your business is value-add in how much they know about how to run data centers. But a decent number of these are triple-net-lease arrangements where the owners may not even be allowed to come to the data center.

And so the tech company is still operating the data center, so they’re not benefiting from that. They’ve just made themselves asset light. Now, if we look at, certainly I mentioned about how the street analyzes it, I think that in my more cynical moments, I look at some of the other industries that are asset heavy but have moved asset light by proxy, and we’ve all seen certain airlines and even certain companies that heavily leverage real estate build structures that allow them to essentially work toward flexing future versions of default and termination options.

If you think about an airline, they’ve got the airplanes they own and the airplanes they lease and the airlines they securitize, they buy different options with each of those. And if these companies are just trying to make themselves asset light to make Wall Street happy, that’s the lower risk answer. If some of them are building out default options in some sense or termination options because many of these leases have very interesting exits, then you’ve got to ask some tougher questions about your confidence in the credit.

Devlin: That leads directly to the next question I wanted to ask, which was, I would presume that there would be much less risk if a lender were lending to an asset light company like one of these big tech firms and they were going to use that money to build a data center versus if they were lending into one of those companies that specialize in data centers but are not the tech giants like a Digital Realty. Is that the case or you mentioned something about exit strategies. Are you necessarily more protected lending to one of these giants that could easily build the project themselves, but they’d rather use their money in a different way?

Anderson: The details matter here because whether you have direct lend to someone with a diversified cashflow, a Google or a Microsoft that could pay back the loan through other means, again, I think this is an absurd possibility, but if AI evaporated tomorrow, could they dig themselves back out? And clearly a loan that directly calls back to that parent is a beautiful thing, but the pure play large AI companies or some of the REITs have a little more dependency on the particular path we take going forward. And how comfortable you are has to be driven by, for example, the REITs, you really do care about their operational skill because in the end that’s one of the things that’s going to be important.

You’re going to care about their siting skill, like whether they’ve picked places that are going to in the long run be appealing, whether that’s to something you’ve heard before. If you’re doing inference, do you have the right position for good latency to users? Are you actually in a position where the power grid has enough excess that there may not be some clawback on your capacity? Et cetera. You want to look to see if they understand those things. I think there have been some interesting structures where the lessee was a subsidiary and the subsidiary may have a very strong guarantee back to the parent, may have sunsetting or requiring renewal guarantees or limited guarantees.

Suddenly what looks like a lease to a big great firm is way less appealing if there are certain kinds of exits. And traditional underwriters know to look for these things, but I think there’s been some bypassing of these in the kind of gold rush feeling we’ve had over the past couple of years that these things will sort out.

And just as a funny little aside, I for fun asked this question to a couple of AIs and they put much higher risk on lending to people away from the core AI companies or much higher than my perception. They use labels like high and very high risk on doing that, which I think is at minimum amusing.

Devlin: So speaking of exits and maybe for different reasons, we’re seeing quite a few data center projects being withdrawn due to community opposition. I’ve heard some people even say that that’s sort of part of this baked in with companies who know that there will be opposition, so they might have twice as many sites lined up. But my question is, what is the risk of a lender lending to a project that never gets off the ground? Is there any or are lenders very careful not to lend into a project without the permitting and sense that the community approves?

Anderson: Yeah, I think there’s some nuance here. There’s this question of whether it’s legally permissible and has been approved by government, and I don’t know of any lenders and any originators who aren’t good and careful and check those things, but those are completely different from whether there’s community acceptance. A bank or a private-credit lender or whoever’s doing the construction loan can get into a place where if they don’t do an enormous amount of legwork, they may be stepping into a situation in which, yeah, the forms have been checked, but there’s massive community opposition.

I think that this is complicated because first on just a systemic basis, the level of pushback first cuts very strongly across both parties. We saw people kind of say, “Well, this is a thing on the left and things like New York’s restrictions.” When Texas’s restrictions came through and Pennsylvania’s restrictions came through, the normal NIMBY case that we face is I want to build a corner store. I want to build a small apartment building in a neighborhood of houses. Those are things that by most studies increase property value and increase quality of life, but it may not fit people’s conception of their community. Those places still pay taxes in a normal way. They’re going to provide housing and/or jobs in a normal way.

With these data centers, this upwelling is coming from if you’ve bypassed the process and then you put a facility that may deliver short-term construction jobs, but in fact, those may even create complications of boomtown kinds of problems, but then they often deliver negligible jobs. I mean, if someone involved in the financing of say a center that runs temporary power for a long time and therefore leads to air pollution and additional noise pollution and even the long-term noise pollution from many of these facilities and to the people close to them, we need to recognize these are real objections.

Devlin: So say that all of that is fine and it all goes smoothly, the project goes through. Are we in the clear when we’re building or what can actually shut down or derail a project that’s already underway? Could there be a shortage in materials? I know that there might be some highly specialized technical products that maybe you can’t get access to, especially if we are in a boom and maybe there’s a hundred other data centers that need the same materials. What about labor supplies? Are there risks there as well?

Anderson: You’ve hit on two that have really affected people. A lot of the stuff in building a data center, there’s probably not a subtle concrete shortage in your area, but when you get to things like switching equipment, when you get to things like some of the equipment needed for that substation, when you get to the specialized labor needed to set these things up, there can be a lot of shortfalls, and especially when it comes to the equipment inside the building. I mean, the moment you start to touch the chips, the PDUs, there’s so much room for there to be shortages of labor and of staff in addition to equipment. I mean, that’s been a big issue so far, but you have to add to that. We spoke earlier about what are being called moratoriums. They vary in how restrictive they really are, but the public swing in the other direction is creating lawsuits.

It’s creating these higher-level moratoriums, and so various structures for essentially fighting the data centers can definitely happen. So I do think there are a number of ways things can go off the rails. Now, I do think that a lot of construction loans are designed so that the folks on the development side are putting forward a lot of the early money, and it can become obvious quite early on that you have problems. So I think in many of the cases, it fortunately might be kind of an embarrassment and loss of future revenue, but the actual severity compared to the loan commitment may be very small and maybe even not material to the institution if they’re structured well. I do think it gets more complicated if, for example, you’re well along in the process before a moratorium or lawsuit or other forms of community objection, for example, stop a center where you’ve already managed to spend all the money on the expense of hardware and the labor and everything else.

Devlin: Right. How much of a relief valve could be the securitization of some of these loans? So they get originated and maybe they’re risky, but then they can be sold on to other markets. We recently had some SEC guidance, I believe, on securitizing data center loans. Does this affect your analysis at all?

Anderson: The first one’s a major one. The risk sits with the person who ends up holding it. And so for example, if you’re originating loans and immediately getting them into the market, you’re not actually dependent on how AI turns out. You’re dependent upon a consistent short-term view. In fact, some of the construction loans that got done quickly and smoothly didn’t ever have to worry about the direction of AI because they already got built, done, and handed off to somebody else for the permanent loan. And of course, the final holders of that risk need to understand what they’re holding and the implications of it. But if you are in the business of passing those loans through, you should know you’re in that business, and it’s an enormous de-risking for them, of course. I think the SEC rule is probably right in principle. These are not normal securitizations in most cases.

I think its outcomes are a little more troubling and suggest that there may be gaps because there’s a lot of transparency drop here. There are restrictions like Reg AB, parts of Rule 15 and so on, that people don’t have to be as transparent if these things aren’t conventional ABS anymore. And one thing we’ve certainly seen is that transparency is lacking in a lot of places in this space. As somebody who spent years on bond trading desks, I am a firm believer that markets operate more efficiently and more fairly in a world of transparency where people can be making the right judgments about the right risks for them. And look, I’m also a believer that there’s kind of no bad risk, there’s just a bad price for that risk. Anyone who wants to make even some of the bets that my comments might seem to push against, I think there’s a right price for that.

I think the real question is people need to, am I making the bet I intend to? Am I getting paid well enough for that? And for those that have more pivotal roles in our economy, banks, insurers, pension funds, et cetera, especially, do they have an understanding of their reasonable stress-test downside consequences that fit the roles that they have in society?

Devlin: Edwin, I think you answered quite a bit of my final question. Maybe you want to expand on it a bit, and that was just how can lenders protect themselves against the various risks? You just described a philosophy that makes a lot of sense. Are there any things that you’d like to summarize that we’ve chatted about? Such a complex issue, it seems kind of odd to say, can you give me a takeaway in a minute or two, but what could you say if that were the question?

Anderson: I think the takeaways are that maybe I’d start with a meta point, which is, and not about the company, Meta, but meta in the sense of broadly cutting across things, that betting on data centers is not betting on AI. It’s betting on a very particular set of outcomes and ways that things play out. Some may think that that is the overwhelmingly likely one, but I think that’s a little less obvious than it might seem, and I think people need to make sure that they understand that they’re making a bet that’s further than that. A second thing I would follow with is to recognize that huge tech changes, and I know there’s kind of a fight of whether to call this a bubble or not. We’ve called every other one of a bubble, including the one that didn’t burst. Only one tech change ever didn’t create a crash afterwards, and it was the first British railway bubble, and they still call it the first British railway bubble, but it never popped because the economic gain was astoundingly higher than even the optimists thought.

Maybe that’s this, but otherwise, there’s probably going to be some chaos and people need to be realistic about how that plays out or whether they want to bet that this is the second time in, as far as I’m aware, recorded history, that a tech change didn’t have a downswing on the back. Second, if you think about the early comments about the underwriting, you need to understand this underwriting, and there needs to be a lot of scrutiny and understanding of the real legislative limits and legal constraints around power. There needs to be a deep understanding of appraisals and how things like, it’s called FF&E, the furnitures, the fixtures, et cetera, are included, not included, how they were treated in your appraisals. There’s a lot of nuance in these documents that is not very important, a lot of it. Do you have a real firm grip on that?

Another is smarter limits. A number of big institutions are already doing this, and again, I found it great to hear that they are, and I think that people need to be rethinking the way they set their frameworks because there’s a huge correlated risk here on the kind of AI bet being made, and people need to be thoughtful about that. All this wraps back to, I brought up the issues of knowing the risk you’re making and the agency issues. Organizations need to be honest about what their goals are, what they’re motivating people to do, and which bets they want to make. I think it’s entirely reasonable to lean all in on this sector if that’s the bet you intend to make, and if you have the things aligned to be able to pull on the best data centers, the ones that have the best leases, the best siting, have the most community approval.

You can be really smart about this, but I think people that are just following the pack may inadvertently be getting the worst of the loans, may be not ending up quite where they think they’re ending up, eyes-wide-open process. I also want to acknowledge this is hard, and the criticality of doing these things quickly, efficiently, and well is important.

Devlin: It’s an exciting opportunity, but one that has to be taken very seriously and a lot of thought and sounds like a lot of hard work has to go into this to get those deals done the right way.

Anderson: I would maybe highlight, there are a lot of different dynamics here, and in fact, Oliver Wyman’s done some work with the World Economic Forum on how to do some of the components of this lending in more responsible and resilient ways. It can be done well.

Devlin: That concludes our podcast. The ProSight team would like to thank Edwin for sharing his perspectives with our Banking Strategies podcast audience today. To our listeners, thanks for spending your valuable time with us. If you liked it, please spread the word and don’t miss ProSight’s annual conference in November, which will feature the session, Data Center Financing: Lending Considerations in the Age of AI Infrastructure. Thanks again. I’m Frank Devlin.

The views expressed by the speakers are the speakers’ own and do not reflect the views of ProSight Financial Association, BAI, or RMA. The views expressed and information shared are of a general nature and are not intended to address the circumstances of any particular individual or entity. No one should act upon any such views or information shared during this podcast without appropriate professional advice after a thorough examination of the particular situation.