Research Matters by Rothschild & Co Redburn

George Karamanos:
Welcome to the first Rothschild & Co Global Advisory Podcast.

My name is George Karamanos. I’m the Chief Strategist of Rothschild & Co Redburn, and we will be focusing on global themes and how they affect the capital allocation decisions of management.

We have with us Alex Haissl, who sits in equity research and heads the global tech team.

Alex, thank you very much for joining us today. You wrote a fantastic 148-page report on AI infrastructure. From my understanding, AI infrastructure captures the data centres - everything that is needed to actually have AI.

You talked about more than 200 gigawatts of announced capacity, but my understanding is only half of that is actually real and likely in the coming years. Can you run us through the work that was done and what the key conclusions were on the magnitude of this buildout?

Magnitude and timing of the buildout

Alex Haissl:
Yes, thanks for having me on this very first broadcast.

The way we think about this global race in infrastructure buildout is that it’s really a funding cycle. As long as the funding cycle keeps going, we can build more and more of these data centres.

The way we’re thinking about these two components is that there’s maybe around 100 gigawatts of capacity being built by the hyperscalers—the likes of Amazon, Microsoft and Google. They are real; we can see these projects really materialising.

But then there’s also the other half of this infrastructure buildout, around 100 gigawatts of capacity. This is where it’s getting a little bit more difficult, mainly because of funding, because the hyperscalers provided the balance sheet for this cohort of the buildout as well. So it’s not getting easier from here.

George Karamanos:
By when do you believe this first 100 gigawatts can be built?

Alex Haissl:
We estimate it will take a couple more years to really get to this 100 gigawatts, because it’s not straightforward.

Often, as analysts, we focus on just volumes and when they can hit. But the reality is that a lot of these data centres are really delayed. You need a lot of permitting.

It’s more the norm that these data centres are delayed. Over 50% of data centres are delayed by at least three months, and what we’ve seen is most likely being pushed to the right even further.

George Karamanos:
If we were to split the use of these data centres between training and inference, how should we be thinking about this difference and the mix? What are the implications for whether there is sufficient capacity or not at the end of the day?

Alex Haissl:
Training is basically building and developing the intelligence. This is when the model labs really train these models.

Then inferencing is basically the use of these models. If you use a chatbot, that would be inferencing. If you have an agent and it’s doing some tasks, that would be inferencing as well.

We estimate that around 70% of the capacity being used right now is for training. That’s a little bit more controversial, because the industry sees it the other way around.

The tricky part really is: if training were to hit the wall in terms of performance, or if, for safety reasons, training is paused, then all of a sudden we could end up in a situation where we’re already oversupplied right now, if you really stop pushing on the training side.

George Karamanos:
Where do you see training moving towards? Because there is robotics, there is sovereign AI. Will that also need training, and therefore do we still have another five to ten years of training needed?

Alex Haissl:
When we talk about training, it’s very important to distinguish between the final training runs and the research component.

The bigger part is not the final training run. It’s really the research, trying to work out what’s going to happen.

I think there’s a lot of room to optimise on the training side, which raises the question of how much we actually really need in the long run in terms of capacity.

Funding and refinancing

George Karamanos:
I guess, then, if we were to move to funding: bubbles burst because there is a lack of liquidity rather than valuations or capacity or anything like that.

There is a big funding wall coming up in the venture debt market, on the private credit side, in 2027. When do you see the big refinancing coming through in this part of the market?

Alex Haissl:
Funding is already getting more difficult right now. What we said in our report as well is that the credit market is flashing some red lights, while the equity market is still outright green.

Initially, it was really the hyperscalers that used up their massive operating cash flows. They issued debt; everything was still fine.

But what we’re seeing now is that, to build this second part—the hundred gigawatts of capacity—you need new financing structures. You potentially also need chip companies like NVIDIA and Broadcom providing some backstops, because the true leverage of hyperscalers is actually higher than it would look on the surface.

In terms of your question on refinancing, a lot of debt has to be refinanced around 2030, 2031.

Even for data centre buildouts, when you look at some of these former crypto-mining companies, they have only issued short-duration debt on long-term contracts. That’s another layer of complexity when it really comes to the funding side of things.

George Karamanos:
What do you think can derail the appetite for this debt over the next two years?

Because we have US presidential elections in 2028. We have private credit rising at the moment. The bond markets and the equity markets are not giving the same view.

Can you give us some extra thoughts on that?

Alex Haissl:
Most of the initial buildout was really investment-grade debt—when Amazon and Google issued debt. But then they also have companies that build data centres on their behalf, which is also kind of an extension of the investment-grade footprint.

What is changing now is that there’s debt being issued for non-investment-grade companies. So I’m not just talking about OpenAI and Anthropic, but also neolabs that pay loads of money for compute, but the credit market doesn’t really like it so much. I think that’s one part.

Then the other complexity is obviously delays in data centres, and really trying to work out: if a data centre is delayed by half a year or a year, what are the liabilities? What are the costs? Are the economics actually changing?

I think all of this may just mean that there’s less and less appetite, in particular for certain parts of the infrastructure buildout.

Off-balance-sheet commitments

George Karamanos:
Can you talk to us a little bit about off-balance-sheet financing?

Just to put it in perspective: we’re on 100 gigawatts at $50 billion per gigawatt. We’re talking $5 trillion.

The free cash flow of the hyperscalers is roughly $700 billion a year. So we’re talking three years of free cash flow gives you $2 trillion, which means that you really need another $3 trillion of debt.

My understanding is that the US financial system has $2 trillion abundantly available. Beyond that, it starts getting a bit tight.

But you’ve been talking about another trillion of off-balance-sheet items. Can you talk to us a little bit more about what exactly that means and how that affects the numbers?

Alex Haissl:
Yes. We estimate that the hyperscalers have off-balance-sheet lease commitments in excess of $1 trillion.

I’ll give you a very simple example. If you’re Microsoft, you have two options. You can build a data centre yourself. You pay for the data centre; you have everything on balance sheet. Or you have another company building the data centre on your behalf, but then you just have a 15-year lease, maybe with some extensions as well.

In that scenario, it’s off balance sheet, but you still have all of these liabilities.

What we’ve seen is that the hyperscalers have shifted more off balance sheet, which means that when you look at the KPIs on their own balance sheets, it looks better than it really is.

To some extent, we believe it’s also just de-risking their own balance sheets already. That’s what I said earlier: the true leverage is already higher across the system, because it’s much more difficult to really trace it down.

Interest rates, credit spreads and funding costs

George Karamanos:
At the moment, we have the ten-year, the five-year and the thirty-year bond yields actually spiking.

What do you think will end up being the sensitivity of these moves, if they stay, on the available funding, the cost and the economics of these data centres?

Alex Haissl:
Firstly, the spike in yields is not only an issue for long-duration debt, but also short-duration debt, because the spreads have really started to widen.

Let’s say you’re a neocloud and you’ve issued debt for a contract with Microsoft or Meta. It was an extension of an investment-grade customer.

But if you now want to issue debt for a neolab—a company that wants to compete with OpenAI and Anthropic—they may pay you more, but the credit market doesn’t really like it so much.

We’ve actually seen spreads on the shorter end widen for specific contracts.

The sensitivity is high, because there is already a debate around returns. If your funding costs go up, you get squeezed unless you want to pass on this cost to the final end user.

That turns out to be more difficult, because that’s why we have this debate around token maxing and open source in the first place. So there will be lower returns as a result across the board.

Profitability and the distinction between paper and cash returns

George Karamanos:
Which, I guess, brings us to the third component, which is profitability.

We’ve talked about magnitude. We’ve talked about funding.

On the profitability side, and the implications for valuation, what kind of returns are you seeing from these investments? What are the expectations? What’s the mark-to-market? Where do you think this goes?

Alex Haissl:
In terms of the infrastructure buildout and the returns: let’s say you build a data centre that lasts for 15 to 20 years, and then you put in GPUs that last for six years.

Typically, these long-term contracts would need mid-teens returns, but that obviously assumes that this data centre is utilised for the full duration, which we don’t know. So there’s always a downside risk here.

The reason why most of these deals look very, very similar is quite straightforward, because these are agreed deals, right? This is a deal where Amazon would invest in OpenAI or Anthropic, so they know exactly what kind of returns they really get.

But just to be very, very clear: these are on-paper returns. These are not real cash returns from these projects.

The best way to think about this is that around 70% of the hyperscaler infrastructure buildout is for two companies. What normally comes alongside a compute deal is an equity injection.

Amazon has signed two deals with OpenAI—to give you a number, $138 billion. But at the same time, it also invests $50 billion in cash in the same business.

On a gross-profit payback, it’s maybe a one-for-one. What you end up having is a 5% or 6% stake in OpenAI.

So the deals don’t look terrible. But from a cash perspective, it looks very, very different.

George Karamanos:
On that, what you’re saying, if I understand it correctly, is that from a cash flow perspective, this is cash flow neutral. There’s a little bit of circularity. It’s liquidity funding.

But the valuation uplift for Amazon comes from the equity valuation of OpenAI. Doesn’t that create a certain circularity in the system?

Alex Haissl:
It does. That’s what the equity market has already started to price in.

If you had told me last year that Amazon’s cloud business was growing close to 40%, with 40% operating margins, no one would have believed it. But it’s even more unbelievable if the stock is flat.

What is happening is that the market is discounting all of these earnings because they’re non-cash. We’re seeing a much, much lower multiple as a result.

George Karamanos:
If you were to mark to market at the moment—the increase in cost, the reduction in token pricing and, in general, the pricing that we’re seeing today—what kind of returns would you be looking at?

Alex Haissl:
I mean, if—

George Karamanos:
Are we looking closer to 8% or something?

Alex Haissl:
It would most likely still be above the cost of capital for the big tech companies, unless they have price escalation clauses in the contracts where they could easily pass it on. It will still be covering the cost of capital.

I think the big debate for all of these returns is that, if you have an overbuild, the second and third hardware cycles could be negative.

Customer concentration and enterprise adoption

George Karamanos:
In that respect, you had mentioned in your report that the key customer base at the moment is not old-economy companies that are the true buyers of these services, but the AI labs.

Can you talk to us a little bit more? Because that also comes back to your earlier point that 70% of the usage is actually training.

Alex Haissl:
Yes. We estimated that, for the hyperscalers, around 70% of the capacity buildout is for OpenAI and Anthropic. That’s what you can see in the order books.

Now, incrementally, there’s capacity being built for neolabs—companies that really want to compete with OpenAI and Anthropic. Obviously, they don’t have a high revenue base. It’s also training-related.

If you really want to boil it down, true enterprise spending on generative AI is quite small.

There are two ways you can think about it. One is: what is the revenue contribution to OpenAI and Anthropic’s revenue bases? We think they are geared towards the tech side.

Then you can also analyse the order backlog for the hyperscalers, and you can see that enterprises haven’t really accelerated their spending.

The reason this is concerning is because that’s the external money you need to really fund all of this infrastructure buildout.

George Karamanos:
We were looking at an earlier piece that looked at consumer-facing companies, like the Walmarts of the world, the L’Oréals of the world.

The early adopters were only 7% of the sample we were looking at, and they were allocating 4% of sales to AI—effectively doubling their IT costs. That was equivalent to 20% of staff.

I guess this is not a decision management will make over the next budget cycle or two budget cycles, but over a longer period of time.

Do you think the AI labs will continue maintaining their share of the capacity that is available for at least five years, in order to give these old-economy companies time to catch up? Or are we going to start seeing some gaps there?

Alex Haissl:
The share has to go down from here. The 70% that I gave you is not sustainable.

That’s where you can see hyperscaler spending for non-labs in terms of the infrastructure buildout. So there’s a pre-build for old-economy companies.

Maybe just to zoom out and give some perspective on the numbers we’re talking about: we started our conversation on 200 gigawatts of capacity.

For the audience, if you say one gigawatt costs you $50 billion to build the data centre and put the chip in, we’re talking about $10 trillion of invested capital.

If you want to make a 10% return, we’re talking about a trillion. If you then compare this with the income from S&P 500 companies, excluding the Mag Seven, that’s a big chunk.

The shift may happen, but it may just take longer, because the truth is that most enterprises are simply not ready to use AI at scale at this point in time.

George Karamanos:
I think we had looked at that amount: if you translate it into revenues, it’s equivalent to the top three and a half thousand companies at 100% penetration, utilising AI at 4% of sales, or the whole SME market.

So we’re quite far away from that.

I think the average time it takes for innovation to come through is 17 years, and we’re, what, in year four?

If you were to take a step back, though, and look at the supply chain of AI infrastructure, you have energy, semiconductors, optical equipment, the actual shells where the data centres are sitting, and then the hyperscalers, the AI labs, et cetera.

Where do you think pricing power will be strongest?

Pricing power and open-source competition

Alex Haissl:
Right now, what we see is that the big labs—OpenAI and Anthropic—still have pricing power over the hyperscalers, because they have secured the bulk of capacity at really, really good economics for them, which obviously means not so much for the hyperscalers.

But then the other debate is how much pricing power is left for the AI labs with the end user—you and I, or companies using these tools.

So far, what we’ve seen is that they had pricing power because their models have been more expensive. But with the rise of open-source models, they also get more price pressure.

It does feel like you get deflationary pressure across the stack over time. That’s what we saw during the cloud era as well.

George Karamanos:
In your report, you talked a little bit about Chinese competition and their ability to be a lot more efficient.

From a geopolitical point of view, do you see those Chinese models—the open-source ones—being more utilised by the Western world, or do you see an equivalent building?

Alex Haissl:
In terms of open-source dynamics right now, it’s really dominated by the Chinese labs, which is a little bit of a problem for the US, because they obviously want to have a US open-source ecosystem.

NVIDIA is trying to resemble, basically, a US open-source competitor. They’ve acquired Hugging Face, which is a platform.

In terms of usage, we’re seeing that open source, volume-wise, has really taken off and is gaining share. That doesn’t necessarily mean it’s the bulk of the revenue component.

I think the outcome is that, if people use open source really cheaply, you may get pressure on some of the token prices for frontier labs. That’s what we’ve seen.

It also then comes back to how much you can actually reinvest into training.

It is quite a fine balance, because if the frontier labs get pressure on gross margins, there’s less money left for training. But training is supporting all of the infrastructure buildout.

So it’s quite a fragile time for the entire ecosystem.

What investors should watch

George Karamanos:
I guess this fragility that we can see from the equity side is offset at the moment by the momentum and the push that the ETF market is giving to the space, and they maintain valuations at quite a high level.

But if we’re to summarise what you are effectively saying: you’re talking about $5 trillion to $10 trillion of capacity buildup. One hundred to two hundred gigawatts is where you see the market settling.

But there is a big question mark over usage and whether this is an oversupply. I guess what you’re saying is that they probably need a little bit less than what they’re building.

When do you think the market will start focusing on this?

Do you think that will be two years before that financing wall hits the market in 2030? So, from 2028—is 2028 the crucial year when we will start seeing these fears becoming more accentuated in the market? Or do you think that happens now?

Alex Haissl:
Watch the credit market. I think, for equity investors, the focus has to shift to dynamics in the credit market, because it’s a debt buildout. That’s more relevant to start with.

I think the first real test for equity markets will be the IPO market. I’m not talking just about the frontier labs, but also a lot of infrastructure players, data centre companies want to go public, and it needs to be seen how much appetite there really is.

The scepticism has definitely grown over the past few weeks, despite amazing headline numbers in terms of deal sizes.

George Karamanos:
Thank you very much for joining us today. This has been fantastic.

Thank you very much for listening to us today, and we hope you join us next time.