Anke Richter: «The market doesn't know yet how to price data centers»
Data center bonds have become a major new asset class at record speed: 120 billion US dollars in just ten months. An analysis by credit strategist Anke Richter has drawn wide attention. In this interview, she explains where the market is mispricing risk – and why Europe will be left behind.
The market for data center bonds has emerged from nothing in the space of ten months. Since last October, developers and operators in the US have raised roughly 120 billion dollars – in structures that have little in common with a classic corporate bond: special purpose vehicles secured against land, buildings and lease agreements, rated in some cases under project finance methodology and in others under real estate methodology.
For investors who have spent years analyzing the same issuers, this is uncharted territory.
Anke Richter, credit strategist at the ISP Group, published one of the first comprehensive studies of this market in August: 22 densely packed pages of quantitative analysis, complete with her own valuation framework for the new bond class.
The report has brought her numerous investor conversations as far afield as Asia and Australia. And two bond issues that came to market shortly afterward provided the first test of her valuation method.
Ms. Richter, since October roughly 120 billion dollars in data center bonds have come to market. Have you ever seen anything like this?
No. And almost more interesting than the volume is the size of the individual bonds. The smallest in this universe is 715 million dollars, but we have a great many bonds in the two, three, four, five billion range. The largest is the Beignet deal, the Meta-backed project in Louisiana: 27 billion dollars. Those are orders of magnitude you don't really see in the traditional corporate bond market, and they are concentrated among a handful of larger issuers.
You've been in the credit market for thirty years. What does this remind you of?
I compare it to the time twenty-five years ago, when the internet boom hit. Back then, the early mantra was the same: everybody is a winner. People valued «eyeballs» and everything else along with it. And some of it later turned out to be bad investments. It will be the same this time. An enormous amount of money is being poured into AI, but some of it is being thrown out the window – we just need to figure out where, so we can avoid investing there. That's the challenge.
And the structures themselves?
I always come back to AT1 bonds and hybrid bonds. When those first appeared, everyone asked: how does this work? What's the risk? And nobody was sure they really understood it. There's always an education phase at the start. Which also means: the market is not priced to perfection. And that, in my view, is exactly why there are pockets of opportunity.
«The basic idea is project finance, the whole thing is ring-fenced.»
What is an investor actually buying here: corporate risk or project risk?
The basic idea is project finance, the whole thing is ring-fenced. These are huge sums raised in the market, then you build for a year or eighteen months, and at some point the cash flow starts coming in. Because it's ring-fenced, all things equal, you can support a higher debt capacity. That said, there are also structures with weaker covenants that can allow the project to re-lever again later. It took me a while to notice this: rating agencies rate some of these projects under a project finance methodology and others under the classic real estate rating methodology from corporate finance.
Why real estate?
Because when we talk about data centers, we're ultimately talking about very sophisticated warehouses. We're not talking about the technology – the tenant brings that in. This is brick and mortar, a building.
The riskiest phase is construction. What are the pitfalls?
It's hard for a non-specialist to assess, and yes, things can go wrong, especially on the energy side. But I've myself followed the oil and gas sector closely. If you compare it with other project financings – drilling for oil in the North Sea, or a highly complex chemical plant – building a data center isn't technically very challenging. The challenge is one of project management, of timing. Right now, everyone is trying to do exactly the same thing at once. Apparently it's very difficult to find an electrician in America; an electrician can reportedly earn 300,000 dollars a year these days. So construction is the riskiest phase. That said, compared with a mining project or an oil project, where suddenly there turns out to be less oil than expected, this is actually on the lower-risk side.
«Your big risk then is that the tenant remains solvent.»
And once it's built?
Then you have a long-term tenant, the rent rises each year by a fixed amount, and depending on the contract, you have almost no costs. Your big risk then is that the tenant remains solvent. These leases can effectively not be terminated; anyone who wants out has to pay the remainder of the rent. The shortest run ten years, the longest in this group twenty. And with a handful of exceptions, the project can be amortized within the first lease term. So it's not a case of the lease ending after five years while you still have enormous debt outstanding and have to find a new tenant. That risk doesn't exist.
These bonds yield significantly more than comparably rated corporate bonds. Is that a genuine market inefficiency, or are the rating agencies not fully capturing the risk?
Talking about market efficiency would presuppose that the market is efficient. I'm not sure it's always as efficient as we learned at university. A lot of this is simply very new. Except for one project, everything is still in the construction phase, and that construction risk brings uncertainty – we're still in a learning phase about everything that can go wrong. You can read a certain amount about it, but at the end of the day: how risky is it when something is being built in Nevada? Most of us aren't experts in that field. On top of that, there are the project finance structures, where people first have to work out what to look for. And the market still hasn't found a framework for how to price the new data center bonds and their risks.
You've developed such a model. Has it proven itself?
After I published the report, two larger deals came to market. One was Zenith Arc, a project in Oklahoma. Primary market price talk was in the high eight-percent range. I applied my framework and said: at least 9.30 to 9.40 percent. The bond was ultimately priced at 9 percent and then rose to 9.60 to 9.70 percent. So I was pretty much on the mark there. Then came QTS, where I had called 6.20 to 6.30 percent as fair value. The bond was issued above 7 percent but is now trading around 6.90 percent, even though rates are higher. The framework has held up, though the market is still uncertain about how best to price these bonds.
«In my view, something like this cannot trade 100 or 150 basis points away from the tenant.»
Where do you see the value?
Once I'm past the construction phase of the data center, I have a bond where my tenant is Amazon or Meta or some other extremely highly rated company – and on top of that I have the security of the data center itself. In my view, something like this cannot trade 100 or 150 basis points away from the tenant. I've suggested 40 to 50 basis points. That's where the compression potential lies. I would be more cautious with bonds that have a weaker tenant. Take Oracle: a low triple-B. Once such a project is out of the construction phase, it will also be, at best, a low triple-B. Then both are at the same level, and how much compression potential do I really have there? Projects with CoreWeave as tenant yield more, yes – but that's for someone with a higher risk appetite, since there's a risk the tenant defaults and you may have to find a new one. In such cases, it's important to look at the location of the data center, since some facilities will likely be easier to re-lease than others.
At Zenith Arc, the tenant is Jane Street, a proprietary trading firm. That's a type of tenant that doesn't appear at all in your framework of hyperscalers, Oracle and neoclouds. How do you assess such a counterparty over a term of ten years or more?
As a starting point, we take a bond of comparable maturity from Jane Street. Then we add 40 basis points – our standard premium for data center bonds, since we have no direct exposure to Jane Street. Because this project is still in the construction phase, we think investors should get a risk premium of around 120 basis points. On top of that, we add another 40 basis points, since Jane Street is a weaker credit than Google or Microsoft and we think these bonds are inherently more volatile. Adding all that up, such a bond should yield about 2 percentage points more than a Jane Street bond. Of course, once the construction phase is over, the 120-basis-point risk premium should disappear.
Will Europe replicate this market?
The problem in Europe is energy costs. These AI facilities need enormous amounts of power. And you can't just put a new Nvidia rack into an ordinary data center, because the power lines, the cooling, everything has to be different. And in many parts of Europe – I'm German myself – energy prices are such that you simply wouldn't build there. I recently saw that a data center is being built somewhere near Frankfurt, but we're talking about fairly small projects. The largest project in America, that Beignet bond in Louisiana, will consume twice as much energy as the city of New Orleans. Just think about that for a moment.
«You can't just say: Louis Vuitton or Heineken, I know those.»
So nothing at all?
In Europe, it makes sense where energy is cheap – generally the same places where you'd also produce aluminum or steel. Definitely not Germany. The Nordics are more promising, and there are already projects in France. When you're planning a data center, that's the first question regardless: where do I have access to power? You tell the utility how big the campus will be and how much power you'll need, and ask: can you even deliver that? Only once that's secured do you develop the project. That's the real bottleneck in America, and in many other countries too. There will certainly be the odd bond in Europe, but definitely not at the scale we're seeing in America.
In the recent general debate in the German Bundestag, there was discussion of how Germany could become a leader in AI.
If you don't have power, not much is going to happen. So that's more political rhetoric than reality.
What does all this mean for a portfolio manager?
It's part of the journey every investor is going through right now: where do I want to be invested in the AI ecosystem? Do I want exposure to chips, to hyperscalers, to data centers? What's an equity risk, what's a bond risk? With Oracle, for instance, I think: that's probably more of an equity risk. Right now we're in a phase where a lot of people need to go back to their desks and do the hard work. You can't just say: Louis Vuitton or Heineken, I know those. And a lot of investors simply don't have that kind of time. There's always this belief that everyone has huge teams analyzing round the clock. That's not how it is. It's a phase of greater uncertainty, where people are genuinely out of their depth in places. And that, of course, always brings opportunities too.
Anke Richter has been credit strategist at the ISP Group since January 2022, where she covers the USD and EUR credit markets. Previously, she was an associate managing director at Moody's Investors Service, responsible for around 100 ratings of European oil, gas, chemicals and mining companies. Before that, she headed European credit research at insurance asset manager Conning and worked as a credit strategist and analyst at Mizuho International, Conduit Capital Markets, Calyon, Deutsche Bank and JP Morgan.
