← Shift Key with Robinson Meyer
Shift Key with Robinson Meyer
How Clean Energy Could Prepare for an AI Bubble
01:19:13
Rob and Jesse talk data center finance with the Center for Public Enterprise’s Advait Arun.
Loading summary
Transcript138 lines
- [00:01]
A
You are listening to ShiftKey Heatmaps weekly podcast about decarbonization and the shift away from fossil fuels.
- [00:08]
B
Get excited because on this week's show.
- [00:10]
A
We are talking about project finance specifically.
- [00:14]
B
For data centers with Advait Arun, a.
- [00:16]
A
Senior Associate for Capital Markets at the.
- [00:18]
B
Center for Public Enterprise.
- [00:19]
A
Advait is the author of a report on the AI buildout called Bubble or Nothing, which, you know, hint hint is.
- [00:26]
C
Kind of the vibe we're talking about today.
- [00:28]
A
We talked with Advait about the bond market and asset depreciation and systemic risk.
- [00:33]
B
And mini perm loans.
- [00:35]
A
Hold onto your hats.
- [00:36]
B
It's all coming up after this.
- [00:39]
C
America's future depends on reliable power provided where and when it's needed. It depends, in other words, on long duration energy storage. Hydrostor's advanced compressed air energy storage technology is helping build the grid of tomorrow with secure, reliable power and thousands of American jobs. With bipartisan support and a flexible supply chain, long duration energy storage is the missing puzzle piece to scale energy independence. Learn more about Hydrostor's Willow Rock project and the future of energy storage at Hydrostor ca. Electricity demand is growing faster than ever, creating an unprecedented opportunity to meet load growth with customer owned distributed energy resources. Uplight is a clean energy technology company that unlocks grid capacity by activating energy customers and their devices to generate, shift and save energy. The uplight demand stack integrates energy efficiency rates and virtual power plant programs into a cohesive portfolio to meet surging demand while improving grid resilience, accelerating decarbonization and reducing costs for energy providers and customers. Learn how Uplight can help your utility harness existing customer assets to meet the moment@uplight.com Heatmap. Hi, I'm Robinson Meyer, the founding Executive editor of Heat Map News.
- [02:14]
A
And I'm Jesse Jenkins, a professor of Energy Systems Engineering at Princeton University.
- [02:18]
C
And you are listening to ShiftKey Heat Maps weekly podcast about decarbonization and the shift away from fossil fuels. On this week's episode of Shift Key, we are having a very wonky conversation about the AI boom, about how it is being financed, who is financing it, and what all of that could mean for clean energy companies. Our guest today is Advait Arun. He's a Senior Associate for Capital Markets at the center for Public Enterprise, which is a think tank that studies how the public sector can guide economic development. Advait is the author of a new report called Bubble or Nothing, which is a study of how data centers are financed and what that means for clean energy and power hungry infrastructure. Over the past 18 months as you know, the AI boom and data center related electricity demand have become entangled with the clean energy industry and with the products of cutting edge decarbonization. And I would say electrotech startups, tech companies are now willing to pay a lot for electricity, especially reliable 24, 70 carbon electricity. They're also willing to pay a lot to build data centers. And some people are worried those investments may not work out. So what does all of this mean for the long term viability of these decarbonization companies and for the long term health of decarbonization? And if you think the AI boom is maybe a little fragile, which you may or may not, I think we kind of come out in both ways that you'll hear on the show. What should policymakers be thinking about now in order to protect decarbonization and clean energy from the deflation of a bubble? Well, we're talking about all of that today. All that and more. Advai Arun, welcome to Shift Key.
- [03:52]
B
Nice to be here. Yeah, thank you so much for having me. I know we've talked a lot together about what data centers are doing for the energy system, and I'm excited to talk about what they're doing to the broader economy too.
- [04:02]
C
Well, you've written a fantastic report that we'll stick in show notes that I recommend to anyone who's listening to this show. If you are a Shift Key listener and you appreciate the wonky depths of any policy area that's in the news right now, Advai's report, Bubble or Nothing, from the center for Public Enterprise. It's a fantastic piece of research about how all the different parts of the AI and data center ecosystem fit together and what those kind of interconnections and relationships mean not only for data centers and AI and all the industries dependent on them, which increasingly means the energy, the clean energy industry, but also the broader economy and kind of gives a sense of where things might go out of this current economic froth. So I wanted to start by just asking you. You start the report by just describing the four big types of actors in the AI boom, who they are and what they do. Can you give us a sense of the four big categories of actor that are playing a role in this boom?
- [05:05]
B
Sure. I think first and foremost the ones that we're all the most familiar with are the AI service providers themselves. And that includes the tech companies that I think we've all gotten to know very well over the last 20 years. That's Google, that's Meta, that's Amazon. That's also to some degree, Apple, although the spending is a little lower. We're talking about the big tech companies. We're also talking about a more, I think, nimble, smaller set of actors. They're called Neo Clouds or I think Compute as a service companies like Core, Weave like Nebias. And we have a bunch of service providers in that ecosystem, Cursor Replit. They're all in the business now, whether they're already doing work on the Internet or whether they're just providing AI specific services. That's the first category. These are the tenants of data centers, but what they do is provide services to all of us to use in the form of large language models and the like.
- [05:51]
C
Are these tenants, are they the ones who are often called hyperscalers, or is that a different category?
- [05:57]
B
So there's been a shift where we started to call the tech giants and some of the AI giants just hyperscalers by virtue of the scale of their data centers. I'm not actually sure what the origin of this term is. I'd be curious to go do a deeper dive. But we've just started calling them hyperscalers by virtue of how many billions of dollars they've started spending on the physical infrastructure required to provide these AI services. So that's the first actor. Just the companies themselves. And they are tenants in data centers, which are a piece of real estate infrastructure that has a lot of, I think, commercial technology property. It's a weird asset class, but that's the land and the facilities that these tenants lease out. And these data centers are built by companies like Equinix, like Stack Infrastructure, Digital Realty. There's a whole host of names here that we might not actually all be as familiar with, but they are doing the brick and mortar picks and shovels like steel, concrete, silicon of this transition to a new economic model idea, as some of the supporters and proponents of this technology might argue. This is. That's the second actor. The third actor would be the debt providers. This would be chiefly private credit and private equity. You have firms like Blue Owl and Apollo providing a lot of debt and equity into these data centers, into the tenants ideas for how they want to structure their GPUs, their server racks. They're also working with lenders in the bond market. There's underwriters like Morgan Stan, underwriters in investment banks like Morgan Stanley, Goldman Sachs playing in the space, but also bond giants like Pimco. And they are long established financial institutions working in the fixed income capital markets to draw down billions of dollars of financing into these fiscal infrastructures. And last but not least, the fourth actor is Nvidia amd. The GPU providers, the GPU producers, really, they're the ones making the chips and designing and making the chips that ultimately Apple, Microsoft, Amazon, Alphabet, they'll all be using.
- [07:44]
C
I think why this was helpful to hear is that often, especially in energy conversations we treat. First of all, I appreciate you calling out the hyperscaler thing because this was a transition that I had not fully clocked, frankly, because in my head I think of the hyperscalers as like I knew they were a class of rapidly growing data center users, but I was not totally clear on who exactly they were. And I think understanding that part of what's happened here is that we've just started calling the big tech companies, whether it's faang, whatever they are, Google, Microsoft, meta hyperscalers. And that as a category kind of also includes the OpenAI and Anthropics of the world, was not something that I had necessarily fully understood. And I think at the same time what I also hadn't understood is that the those companies are often different than the actual data center companies.
- [08:36]
B
Yeah, you're right. It's all a little bit confusing. And what I will say is that where data centers themselves are concerned, there are different kinds of data centers. We've had them for a long time. They're the physical infrastructure behind the Internet that we're using right now to record this. And those would be cloud focused data centers. Now you have compute and AI focused data centers. They're still similarly in these steel shells around the country that we're building at a rapid clip. But the data centers, the ones that serve the compute needs of the hyperscalers, those are actually designed by the hyperscalers themselves. It's just the hyperscalers don't own them. Usually the hyperscalers step back and lease out the data center for themselves, even though they were the ones to design the transaction for that infrastructure in the first place. A different company will usually own and operate it when the data center is actually up and running.
- [09:23]
C
So you know, when we talk about data center companies, we're actually talking about something closer to like a construction company or a real estate developer that is putting together these big powered shells that then are what actually is purchasing the power off the grid and actually transacting at the physical level with the built environment to create the infrastructure that is needed for AI models. But they are not the same companies as the metas or the Googles that are then running software inside Those powered shells.
- [09:54]
B
Yeah, that's exactly right. These tech companies, they don't want to be taking on the burden of having these long lived capital assets on their balance sheets. And even if Meta, for example, is talking to the utility themselves to understand the kinds of power they need to draw the load requirements, it's that data center company that's really working on the real estate for the transaction. They bring in the contractors, they bring in the suppliers, they bring, they put the steel and the concrete down in the ground. They're the ones handling all of this.
- [10:19]
C
So Advait, your report draws out all the different interconnections between these actors. Financial interconnections, to be clear. Not great interconnections. And I want to get to this bubble question in a second, but first, can you just describe a little bit how these different actors are connected in the financial system? And then when you looked at the whole rat's nest here of financial relationships, what did you take away as the biggest risk to the AI boom? We're going to talk about whether it's a bubble in a second, but I wanted to start by asking like, what's the biggest risk to the entire superstructure of financial relationships that have been created here?
- [10:58]
B
So I'll walk through exactly why this works the way it does, piece by piece in a second. But I think upfront the biggest risk is just the lack of demonstrable cash flows. The way the sector is built, it's a bunch of very competing inference service providers offering essentially what seem to be more or less identical services. And that puts price pressure on all of them to sort of undercut the other expand market share. And as others have described it in the market, it's a very existential race for a lot of them now, as price can't, as prices are sort of suppressed and revenues aren't really materializing very quickly. You're also in a situation where refinancing deadlines are coming up, refinancing deadlines on getting GPUs in your servers, on refinance, on paying your lease to a data center. These refinancing deadlines are, I think, the sufficient condition for us to be very concerned. It's not about what stock prices are doing right now. The valuation represents what a lot of investors are thinking the earnings potential of this technology is. But in reality it's just the lack of cash flow and the asset liability mismatches throughout the sector create these refinancing risks where first and foremost, the value of the physical infrastructure underlying the AI boom is really uncertain. Those GPUs have a value that is hard to determine. Volatile fluctuates, especially as Nvidia and AMD release new products. All of a sudden you have no collateral base and not a stable collateral base against which to borrow. At the same time as cash flows are suppressed due to competition for market share, all that happens as data center tenants, the hyperscalers and a lot of the retail tenants alike are they're all undertaking multiple cycles of really expensive, hard to finance capital expenditure while they have to pay their lease. They're going to be doing multiple cycles of that within one lease term.
- [12:35]
C
And this is because the GPUs keep getting updated. So basically you have to go in. So you have your, you're a tenant in a data center, you got to go in there periodically and like buy all new GPUs and replace the GPUs in the physical data center with new ones just to stay on the cutting edge. Right? That's kind of the mechanism here.
- [12:55]
B
There's two things going on. There's the physical depreciation, there's the physical deterioration of the gpu. To use it really quickly, especially if demand rises for the service you're providing, that will wear the technology out and you got to get a new one. But also it seems like there's a pretty solid demand, especially among the tech companies and the hyperscale and some of the neo clouds like Core Weave for the cutting edge GPUs, they want to lock up that capacity first. And in that case, if Nvidia is releasing on a yearly cycle of AMD is designing its chips with respect to what the hyperscalers are looking for, I think we're in a race to sort of hoover up a lot of the new high end cutting edge capacity out there.
- [13:29]
A
And that's both a strategic effort to get some degree of product differentiation by having the fastest, best chips in the market compared to your competitors, but also to block out your competitors from having them because they know there's a limited supply each year of these chips. So there's a lot of this. That's the part of the reason it's so frothy, it feels like is everybody's racing to get ahead of each other because at some point there's a big pot of money at the end of the rainbow where they consolidate market position and are able to become the next Google or the next Amazon or one of them gets even bigger cornering this market. But in the near term they're all just like continually outlaying cash to try to stay ahead of the competition. Not necessarily because there's demand for the product.
- [14:04]
B
Right now there are weird parts of this market where it actually looks a little less competitive than I think this story will describe it as. For example, the hyperscalers such as Google are all paying for the services of Neo Clouds. These Neo Clouds, like Corweave Nabius, they're buying up GPUs to provide compute as a service to whoever is asking for it, for training for inference. But it turns out that some of their biggest customers are also, technically speaking, their competitors, like Google, Microsoft, Meta and these large tech companies. The parts of the MAG7 that are publicly traded, like Google, all the rest, they are building up their own data centers, buying their own GPUs, but they're also renting the services of supposedly their competitors.
- [14:40]
A
And that's just to expand their their bandwidth right now. Right, as they struggle to bring more capacity online.
- [14:45]
B
Right? I think, yeah, there are two explanations. One, it's to get ahead of incoming demand that they're all projecting. They want this extra capacity just on hand, a way to make sure that GPU couldn't be used by a competitor if they rent that out or have that contract available. And then the other thing is, we've all heard these stories and seen all these very, very fun, intricate diagrams of roundabouting or circular financing where everybody's growth projections are therefore just dependent on everyone else's growth projections.
- [15:11]
C
Can you give an example of the roundabouting here? Because I think people will understand that there's some that like Nvidia is investing in Meta, which was then investing in the neoclouds and. But can you just give an example of the extent of the financial interrelationships in the sector?
- [15:28]
B
Sure. There's a lot going on. I think Nvidia is investing in OpenAI, I think up to $100 billion. I think OpenAI has put $300 billion into a deal with Oracle. Oracle is buying chips from Nvidia and using its data centers to lease out to OpenAI. Nvidia bought, I think, $6.3 billion worth of cloud services from Core Weave. Core Weave is one of these Neo Clouds. And Also I think Nvidia's taking back the GPUs that core we buys from it. If Core Weave doesn't use enough of them, OpenAI is paying Core Weave, I think, $22.4 billion. OpenAI is working with AMD to deploy AMD chips. Oracle spending 40 billion on Nvidia, it gets kind of worrying. Like Core Weave becomes a major player at the center of this, OpenAI is a major player. Oracle is becoming a major player in building out all this data center space to lease out to the providers. And Nvidia is backstopping a lot of this market. Nvidia's over half of Nvidia's revenue, I think just comes from Core Weave, Microsoft, OpenAI, if nowhere else.
- [16:21]
C
You describe this dynamic. We're going to get into kind of discussion in a second, but I just want to kind of let you download information into the chat, into the podcast, so to speak. First, you just described this dynamic in the report of like this whole boom is driven first of all on the anticipation of consumer revenues that haven't showed up yet. People and companies are not spending the amount of money on AI that this whole boom is premised on them eventually needing to spend. So that's like one dynamic. But then the second dynamic is everyone, as Jesse was saying, everyone wants to be at the frontier of AI development and that means they need the newest chips from Nvidia or amd. But because Nvidia and AMD are releasing chips every year, nobody wants to own these things. So everyone like wants to use Nvidia and AMD chips, but nobody actually wants Nvidia or AMD chips to sit on their balance sheet, which is kind of mind blowing. Can you just describe why that is? Why does everyone want these things but nobody wants to own them?
- [17:28]
B
I'm really glad you asked because I think this is going to be the key pin on which the sector is going to turn over the next few months. We're seeing like this rush for GPUs right manifest as not only a rush to access GPU services, but the desire to separate that access from ownership almost entirely. Companies like Core weave contract the GPUs they own out to customers. Fixed price contracts. If you want to Access Core Weave GPUs for an inference task you want to run, you will just pay a lump sum for a certain amount of capacity and not Pay by usage. CoreWeave is therefore like it's a take or pay contract. You pay upfront, you take as much as you need and you've already paid. Coreweave has de risked its own GPU holdings by offloading that risk of using the GPU onto you. Any depreciation it faces will ideally be compensated for by the fact that you've already paid upfront for everything you want to use out of that GPU cluster you just bought. That's one way to avoid the risk of GPU'S depreciating on your balance sheet. As Nvidia continues to make GPUs, the other way to do it is just to rent from Nvidia. I think the XAI special Purpose Vehicle that that company has set up rents GPUs from Nvidia. That whole thing is financed just to get the GPUs but not have it on its on its balance sheet. And in actual accounting world, you will mark the value of the lease on your balance sheet and say, this is what it costs. Here's what the value is. But that probably is a little more flexible in how you can arrange that accounting. Rather than having a GPU that you need to find some accounting standard for, mark that value to market, figure out how to resell it in case something happens. And I think, well, what I will say is that in the GPU rental space, if you're trying to rent GPUs rather than buy them from Nvidia, Nvidia therefore is the one eating all of the GPU risk of having those GPUs on its balance sheet. Every time Nvidia makes a new GPU, its existing holdings of GPUs that it's renting out to other customers like Core Weave, that depreciates. But the thing is, this bet makes total sense for Nvidia because if you're Nvidia making the GPUs, why would you not bet on the value of your new GPUs like you are going long on yourself in financial parlance, you're going long on Nvidia, is going long on Nvidia. This bet fails if the demand for GPUs collapses, if the demand for infrared services collapses to a degree that the value of GPUs also collapses. But unless both of those things happen at once, it makes total sense for Nvidia to be okay with everyone renting out their capacity because they're taking a bet that they can eat their own depreciation cost. They're, they're the actor that gets saddled with these. At the end of the day.
- [19:45]
C
Can you just like explain that a little bit more? Because basically what you're saying is that Nvidia has ownership or has some kind of rights to such a large number of its own GPUs that when it releases a new class of GPU, which it is doing now every year, that automatically bumps down in value all those GPUs that were once on the cutting edge and have now become old. And that kind of mechanically means that Nvidia next class of GPU always needs to have a value greater than the loss in value incurred by the depreciation of its current, its current GPU fleet. Right.
- [20:28]
B
That's I think, the right way to approach the question. I think in a world where we're assuming first of all that demand for GPU use doesn't rise to the levels that these companies are projecting, if it does, if cash flow materializes, this isn't actually a problem. But it's worth noting that I think some AI providers like Anthropic are pivoting very heavily towards enterprise AI integrated development interfaces. AI coding related services like Cursor and Replit are also here, whereas I think services like OpenAI are more so on the large language model, daily use kind of structure where it's actually not clear given how inference continues to cost more and more per I think inference services cost more and more and the cash flows don't seem to be matching them, at least currently. It does seem like trying to charge users for how much they use rather than giving them subsidized access to these services. If that continues to be the case, then we continue to see the prisoner's dilemma where everyone who's offering a service, an inference service, has to compete with one another to expand market share rather than, I think, price at cost. So I think that's the other way to look at it.
- [21:23]
A
Let's talk about how this sort of falls apart. I think we're seeing clearly is a race to capture future market potential. And in any such circumstance, I think it's probably realistic to assume that not everybody's going to win that race. Right? Like some people are going to take big write downs because they're not going to make it.
- [21:38]
C
Right?
- [21:39]
A
They're going to stumble or they're going to make a mistake or they're ultimately going to get bought out by their competitors when they get devalued. And so, you know, in a capitalist system that's normally fine, right? The investors in those companies take a bath and regroup and we move on. And in the long term the sector as a whole might be quite healthy after that kind of shakeout happens where a number of long term players emerge that are able to offer services where they don't have to keep spending every year on the brand new chip because they don't have to be constantly racing to the edge and maybe they can get to a more sustainable cost of goods sold or where the revenue base grows enough that the actors can take advantage of that and maybe have a little Bit of pricing power in that context. But there's a way this falls apart as well, which is what you're I think, concerned about. And I want you to sort of spell out like how does the stack of cards crumble here? Where there's a more systematic revaluation of the market here that causes a broader collapse, where even those companies that are healthier, that are actually moving into a more competitive position, might, might get pulled down with that.
- [22:37]
B
Sure. So the way that I approach that question in the report is using what's called T accounts. They're not very common in policymaking world or even in a lot of public, I think, financial documentation, but we use it all the time to simplify and communicate the risks that are shared between different market actors. And through that kind of abstraction we can build that like scaffolding around what risks, when and how do they affect other market actors. So the way that I see this waterfall, this cascade of risks materializing. Well, I think first things first, it's far more instructive to learn that a lot of the construction loans on data centers are what's called mini perm loans with four to six year terms. That is far more instructive to me as a point in time to watch out for than for example, whatever the stock price movements this morning have been. Narratives can change on a dime. But loan terms actually represent, I think, how capital expenditure is being expensed.
- [23:24]
C
And just to kind of explain exactly what this is, because I had to look it up. Basically if you're building a data center, you take out a construction loan first to build the data center and you get a loan for that, that's like a few years and then that gets you to the place where you have your most of the way to a powered shell. You're most of the way to turning this thing on. And then eventually you will take out a permanent mortgage which kind of takes into account, it's a 20 year loan, 15, 20, 30 year loan. It takes into account who your tenants are and what they're going to pay. And that's kind of when you have 100% or very high tenancy in your business, you can get a clear eye on what your cash flow is going to look like. That's when you go to a bank and take out a permanent mortgage for a commercial entity. Like a, like a, like a data center. Right.
- [24:12]
A
That's how it normally works for commercial real estate.
- [24:14]
C
Yeah.
- [24:15]
A
Is that how it's working for data centers right now?
- [24:17]
C
Yeah. But then in this case there's this thing that companies are getting in the middle, which you are calling a mini perm or like mini permanent.
- [24:24]
B
The mini perm loan actually is the construction loan. Mini perm is the term used in commercial real estate where you construct the property and let it sit around for one to two years to generate that initial cash flow that you help use to repay the debt. And normally the loan is structured where it's interest only until the point where the property starts operating. And then you start paying down principal and then you refinance the whole mini perm loan at maturity for that permanent loan or that term loan like you mentioned, between 10 to 30 years long. That's so almost essentially correct. It's just the mini perm loan is also the construction loan. That's the first thing. And I think when we're talking about the waterfall of risks, well, these will be fine unless there's no revenue. And let's start there. If there's no, if we're not spending enough on inference services and there's no revenue or the contracts that core reeva signing with regular customers start falling through, then after that revenue starts deteriorating, we have GPU depreciation and GPU rentals. Rental costs start biting. A lot of the tenants of the data centers, their inability to, I think, treat their collateral value as stable, their inability to pay off, for example, their leases, even though leases come with a lot of security. There's extra lease terms, there's extra months of rent. So on that, I think data centers are charging their tenants despite that, if leases start failing, then all of a sudden data centers are at risk of not having lease income from their tenants. And that tenant churn. If there's no new tenants to take that over, if there's a demand, if there's demand shock or a market correction, then all of a sudden you're left with data centers being unable to pay those, whether it's the mini perm loans or whether the term loans, they're unable to pay for the cost of building that powered shell and operating it over the long term. And at that point, what you have left are data centers that are not really. They don't have tenants, they're not really paying any more for power. You have GPU producers like Nvidia sitting on the sidelines, not being able to sell or rent out their assets. And you don't really have hyperscalers in the business except for the ones that happen to survive, consolidate, so on. I don't mean to stress that like, this will be a dead sector. I think, Jesse, you're absolutely right. Then in a world of a market correction, we might see some competition, consolidation, some change in business models. This happened after all the previous booms. Whether it was the dot com bubble, Amazon actually ended up using a lot of that dark fiber and building out AWS really successfully. We had the fracking boom and then the bust and then fracking started consolidating and earning a profit. After the, after the pandemic in particular, we'll see this consolidation. So I don't mean to say that it's over, but this is how the risks will waterfall for the sector up to a point where you will have stranded assets and you will have a lot of extra GPUs that aren't being used if these risks do materialize.
- [26:50]
C
This episode of Shift Key is brought to you by Hydrostor. The US economy is growing fast and so is its demand for reliable energy. But generation alone isn't enough. Long duration energy storage is the key to building a resilient, secure grid that can meet tomorrow's needs. Hydrostor's advanced compressed air energy storage technology stores clean energy for hours at a time and then delivers baseload power when it's needed most. Hydrostor's Willow Rock energy Storage center in California will provide 500 megawatts of storage, enough to power a city of half a million people for more than eight hours and employ over 6,500 people during construction. Backed by bipartisan support at all levels of government, Hydrostor is leading the charge in utility scale storage with projects underway across America. We're helping communities unlock economic growth and achieve energy independence. Learn more about Hydrostor's unique technology and how it's powering America's Future@Hydrostor ca. So when you talk about a four to six year timeframe for these data center MIDI perm loans, when did that clock start? Just if I'm thinking as a journalist in an editorial sense, are we looking at 2027? Are we looking did some 2028, 2029?
- [28:11]
B
I think, yeah, because I think if chat, GPT and the boom in the narrative around the data centers was all at the beginning of 2023, the end of 2022, then I think you're going to actually start seeing a lot. I think we do start seeing a large uptick in construction or construction loan financing that year and the year after.
- [28:27]
A
Yeah, 2024 was really the biggest jump where it moved from a couple hundred billion a year to almost 300 billion a year globally and now is expected to jump another couple couple hundred billion this year. Right Closing in on half a trillion a year. Yeah, the money's moving for sure. So again, I'm trying to, like, I want to break this down a little bit. Like there's a world where, you know, that you're describing, where there's a market contraction, which leads to a lack of new tenants, which means that we have overcapacity, and data centers go dark. They don't have the ability to secure additional leases, the leases run away or cancel contracts, and then you've got this asset that you can't generate enough revenue to cover your loans. That's one sort of extreme. And I think that could certainly happen in some circumstances. The question is how much of the market is likely to face that at once. But there's another risk to the economy, I think more broadly, maybe not quite as damaging, which is just that as soon as that happens, new construction stops. Right. So how much of our current energy sector, our current construction sector, and our current GDP is tied up in this expectation that we're just going to keep building more and more and more data centers every year? And how much of a risk is that, even if we actually still use all the stuff we built or do within a couple years that get bought up and lease out their capacity, we're no longer going to be spending half a trillion dollars a year on new construction. And that has a pretty big impact across the economy as well.
- [29:48]
B
Yeah, it's a good question. And I want to highlight, I think, some of the ways that risk will manifest that I think policymakers ought to be really attuned to. I think first we should establish the baseline. I think to the best of my understanding, AI investment is more than 40% of US GDP growth this year. And I think one economist had mentioned that I think it contributes more to US growth in the first half of 2025 than all of consumer spending. This, functionally, is the motor of the economy right now. And if we look at previous bubbles, like the end of the housing bubble and the Great Recession, that actually destroyed a lot of contractors, business models across the country. And I think the rise in housing prices and the lack of quick building is in part attributable to the destruction of the housing sector after that bubble. And I think the capacity that was built up in that sector sort of disappearing. So in the absence of other motors of growth in the economy, any kind of construction here could actually take a while to work out. From when it comes to the electricians, the labor, the steel, the concrete, these sectors that are right now being pulled into the Data center world. We don't necessarily have a quantitative measurement of what that looks like. But when it comes to breaking relationships, you'll see the impacts over the near term and the medium term as prices rise, as delay times also rise in the building of infrastructure that also uses those inputs. That's, I think, one way to look at this risk. And I think that's also why if like this is something that we're, I think, really thinking through more holistically, like with every previous recession, I think it's incumbent on policymakers to understand that there's something they can still do about the potential to use these stranded assets. There's something policymakers can do to make sure that there's a cushion to the shock for the other industries that are now putting their picks and shovels towards these particular tracts of land that might in certain cases go dark.
- [31:26]
C
What should policymakers do? Because I want to let me just step back and say what's really cool about the report you've written is that it's kind of classic sell side research. It's like really a really interesting look at the financial stresses and relationships that this one sector faces. But what's so useful about it and why I would recommend it to the shift key listener is that who it's really targeted at is policymakers and saying, okay, well here's this set of risks that are happening in the economy. They're going to happen right then it's going to, there's going to be a fallout from it. We need to start thinking now about how to mitigate the downside of those risks to the broader economy and specifically to the clean energy economy that we care about most. So what, how should policymakers kind of think through cushioning that downside to the broader economy or to the clean energy as they look at the possible deflation of this boom?
- [32:14]
B
Yeah, now you can call this sell side research for policymakers, if you will. I think that's actually what we're going for here. We have three different kinds of risk that we identify. The first is to just people's wealth. I won't get into that one, I think because it's. I think what's more relevant is I think risks to state and public budgets as well as to the energy system. The data center build out is also forcing in real time where public budgets are concerned. There's a lot of tax exemptions for data centers that were instituted for the old cloud kind of data centers. A lot smaller, a lot more dispersed around the country. These tax exemptions on property Taxes and sales taxes have kind of been grandfathered into supporting tax exemptions for these large hyperscale data centers. It's not really clear that a lot of city councils, local governments, state governments, regulation, really have a good awareness of the kinds of projects they're now dealing with. There's a data center in New Mexico or a set of four data centers that's worth $165 billion where the developers are only paying about $300 million in lieu of taxes for I think the kind of the deal that they made with the county, they're paying less than 1% of the total value of the project in lieu of taxes. And I think that's not necessarily like it is in my mind. I think an oversight on policymakers part to, I think, not reconsider some of these rules around how to manage this growth.
- [33:23]
C
Because.
- [33:23]
B
Because if public budgets are too dependent on the potential growth of these data centers, on the jobs they might create, and it might not all be that many. If public budgets are also correlated to the growth of the tech sector more broadly, a market correction here could severely hurt public finances. So we have to make sure that we're watching out for this. When it comes to the state's capacity to actually support other public services, livelihoods, the health of their budgets, so on.
- [33:45]
A
Doesn't that actually limit the amount of local revenue that's tied to these? If they're giving substantial tax abatements, then they're not actually tying their budgets to revenues from these data centers.
- [33:57]
B
That that's fair. It goes both ways. I think if you're giving tax exemptions to data centers, then you're not tying your public budgets directly to property taxes or sales taxes from there. But you might still be counting on the growth in corporate taxes, sales taxes from workers who are now at the site. And those might not actually justify the kind of property tax, sales tax you're losing out on the site itself.
- [34:17]
A
But that's independent of the systemic risk to the sector. Right. I mean, the risk of the sector drying up is that there won't be any more of those mispriced deals to be made. Right. What you're saying is that local governments need to be careful that they're not giving away too much in tax exemptions, that it isn't paid back by the construction and employment income taxes and whatever that's generated. But that's sort of separate from will construction stop or not.
- [34:39]
C
Right?
- [34:40]
B
Well, think of what utilities have started doing. Utilities as well are also wisening up to the risk of for Example, phantom load interconnection. And I think a lot of data centers are branching around, around the country, using LLCs to sort of buy up land, ask for interconnection rights. And utilities have started asking for gating and interconnection fees. They're working on large load interconnection procedures to really make sure that they know what they're dealing with. I think there's also an assumption, and I think it is correct in many cases, that more large load can help actually support the stability of the utility system and the cost that ratepayers face across time. But I think without these gating fees, without transparency, I think this is something that policymakers ought to be thinking a little bit more seriously about outside of the utility sector as well.
- [35:19]
C
I would imagine in certain states the public budget problem becomes more pronounced, like in California, which is so dependent on kind of the top few percentiles of earners. And those earners in California are quite dependent on the wealth effect from rising tech stocks. Like I could imagine that following through quite significantly. And of course, California, that would have implications for what kind.
- [35:43]
B
Right. I think there are some states in California in particular where I think the high income earners in tech and the fact that the tech industry is still very concentrated there, a market correction would certainly affect that state's budget in particular. And we just have to be attuned to what these correlations actually are, California in particular.
- [35:58]
C
But I think, yeah, so talk through the assets, because I think that you said the wealth effect is significant, the public budget effect is significant, and the third was the actual energy assets themselves, because that's right now, data centers are basically buying a lot of offtake from clean energy development and energy development more broadly across the country.
- [36:19]
B
Right. And I think this is something that a lot of folks have been attuned to more recently. The fact that data center boom is also leading to a construction of new energy resources of all kinds. What's I think very cool and promising about the data center boom is the kind of innovation we've seen on flexibility, on demand response, on batteries, flow batteries, in particular, supercapacitors, uninterruptible power systems, cooling systems. There's all kinds of really interesting energy work happening under the umbrella of this particular boom. And I think insofar as I'm interested in actually capturing the benefits of this innovation, this is a part of the bubble that I think I really find valuable. And I hope the policymakers who are also concerned about the energy crisis, affordability crisis, whatnot, might also find I think valuable to preserve. I think what's really, I think worrying is that if a data center market correction does happen, we might see a lot of these energy assets, perhaps stranded, underutilized, the developers might go broke. Depending on who's who. That's something we should be mapping out. The report doesn't super get into that map, but that's I think the next step here. But also data center sites themselves have something really valuable, interconnection rights, especially once they get them. If there's a market correction, utilities and I think grid planners might end up seeing a lot of these dark data center sites with surplus interconnection left over. And what do you do with that? It might be too far away from other uses of existing load if that data center is not actually a load site. And there might be even behind the meter generation or power co located with that data center that could just be potentially stranded in the event that there's no load to take that power insofar as data centers really are pushing out that load curve. So we have to. I think the recommendation that we're making very strongly is for policymakers to consider ways to I think, intervene in the interconnection process, pick up stranded energy technologies in particular, and really make sure that the technical benefits and the interconnection benefits of the data center boom don't necessarily get lost in the event of a market correction.
- [38:08]
C
What would a policy program for that look like?
- [38:14]
B
So first and foremost, I think figuring out what to do with interconnection, that's I think the big thing.
- [38:18]
C
And when you say interconnection, like what's the benefit there that you could eventually stick a factory there, for instance?
- [38:23]
B
Yeah. Our argument is that this is land you could be using for something else like that. I think the challenge of that is, and I think this is where this gets interesting, is that data centers are willing to pay top dollar for energy. They're willing to pay top dollar for that interconnection. They want to make sure that they get built. You have to find a way to refinance that interconnection price or the energy price, the power purchase agreement at a lower price point. I think it's not clear that any other industry is able to pay the kind of top dollar that data centers are able to. And that becomes the challenge. I think the second challenge is that the interconnection queue process itself is fairly chaotic. The phantom load queue, similar to the whole phantom generation problem. It forces utilities to make interconnection decisions based on the Interconnection needs at everything above a particular project in the queue, anything in the queue that came before the particular project and consideration matters for that project. And if data centers drop out, if there's a market correction, parts of that queue disappear. Utilities have a lot more work on their hands redoing a lot of their interconnection studies and analyses. And that's something I think that we're going to have to fix up in the event of a market correction that ends up spilling off in the load itself too.
- [39:25]
A
I mean, this is why the utilities are starting to require these bring your own capacity constructs or tariffs in many cases where they're really quite substantial collateral or upfront payment requirements now being charged. Because, you know, I think we have to continually remind ourselves of the scale of these data centers. This is why I actually have a lot of skepticism about the ability to like reuse these sites. I mean, a gigawatt scale data center, there is literally no load out there that will fill up a gigawatt demand. Unless we're going to like start a new like virgin aluminum industry right from scratch, which we're not. I mean there's just, there are no other energy consumers anywhere close to the power density of data centers, period. They are the most power dense consumers of electricity in world history. And so you slap a gigawatt scale data center down, or 2 gigawatt cell data center down, or you build the infrastructure for that and then it never nobody shows up to plug in their GPUs. Like that's like building the infrastructure for another city for like Pittsburgh Times Two or Chicago. There is just nothing that will step up and fill that gap. And so this is why the utilities are saying, look, if we're going to build this infrastructure for you, like we are not going to our ratepayers or the regulars are making them say this too. Our ratepayers are not going to be stuck holding the bag for this. If you don't show up, you're going to end up paying up a substantial amount upfront for this. And so far what that has led to is a substantial contraction in the number of large load requests they're getting. Because now you got to be real serious about it. And that's I think exactly the right approach here. Because if you think about sort of classic risk management, right, you want to stick the risk with the person who's best able to manage that risk. And the utility has no way to know whether this load request or that load request is real. I mean they can make some guesses. But the people who really do know are the ones developing these projects. And so sticking the risk management challenge onto the developers of the data centers is the right place to put it because they're the ones able to actually say, all right, now I'm at a point where I know this project is suitably de risked that I'm willing to put up that cash to build out that infrastructure, to build the supply, to build the transmission upgrades. And if they're not willing to do that, then you know that it's not real. Right? It's not a real load. So I do think that's having a pretty profound effect on this. The problem right now is that that is not universally applied right there. There are a number, I mean, many, many places that are developing new large load tariffs, but there is no universal tariff and there are a handful of places that are out in front of others. So I think there's also a bit of just shopping around still where data center developers are looking for the places where they can continue to get away with the old practices. Well, they can. And that is where I worry about the biggest risk is some of these utilities that don't move quickly enough to implement this kind of approach are still potentially putting themselves and their customers at risk if the data centers that say they want a gigawatt of power don't show up in the end.
- [42:03]
B
No, it really speaks to the need for a policy modernization kind of push on the power side, on the state side, on the budgeting side of all of these entities that don't want to actually bear the risk of building a whole new city's worth of, of infrastructure. I think it's a good way to put it. Yeah.
- [42:15]
A
And I'm curious to see how this plays out too, in terms of again, other parties in this market becoming increasingly risk averse as well in terms of requiring more and more upfront or collateral collateralization of contracts with the hyperscalers with the data centers. I think the idea that you're doing business with Google, it's got one of the deepest balance sheets in the world, right, Or Microsoft, that's seen as a pretty darn safe bet. But if there is a broad market correction, I guess this is where the systemic risk comes in is if there are a number of, you know, many different contracts with carrier for, you know, H Vac equipment or with utility for transmission service that are all premised on the fact that these are deep pocketed hyperscaler customers. But then there's a broad market correction that means that Their equity valuations are no longer going up and they can no longer continue to finance these sorts of projects, then it's not just them that gets hit. It might be others who misprice that risk in their contracting.
- [43:05]
B
I think that's the right way to look at the risk landscape. And right again, I don't mean to say that this market correction will necessarily happen in the way that, like if there is one, it'll happen in the way that we're describing. I don't want to say that it's necessarily going to. If it does, I think what we will see is partial consolidation of the sector for sure. That happens in a lot of previous crashes. It doesn't mean that the data center boom is necessarily over. It'll actually have validated a lot of utilities and perhaps even states decisions to sort of gatekeep everyone's access to new power through these fees and all the rest of. But you're right, I think we actually do need to look towards, I think downstream to see, right, the H VAC contractors, the electricians, a lot of the construction labor on the site. The fact that this is the motor of the US Economy and that might seem a correction. Where else do they go? Is another policy question that we don't answer in our report. But I think it's worth, I think it's worth paying attention to.
- [43:51]
A
We could build housing.
- [43:53]
B
Yeah, that would be nice. Build housing, build some trains, get some high speed rail aid over the country.
- [43:58]
C
But this is kind of what you wind up suggesting in the, after the 2008 crash. I mean, first of all, there's a wonderful moment in the report where you basically say we think about the Minsky moment. Hyman Minsky, famous financial economist, kind of scholar of instability in the financial system, has this idea of a Minsky moment, which is like often seen as the moment when you realize the emperor has no clothes and like your cash flows can't finance your existing credit load. Is that correct?
- [44:25]
B
The conventional understanding of like when the market is actually unstable and the Minsky moment is about to be triggered is when market actors don't have the ability to refinance their debt or sell their assets at a value that allows them to pay even their interest payments. That's, I think, where this all starts to break down from a financial perspective. But I mean, we're sort of drawing from a school of political economy here that argues that the real Minsky moment is those very political decisions about who loses their shirt when that happens.
- [44:50]
C
Basically, if this, yeah, if this were to play out then some, then politicians are going to have to pick winners and losers with the public's attention focused on them.
- [44:59]
B
That's right. And I think it's a good thing that I think the public is already paying a lot more attention to the energy electricity affordability crisis because I think everyone is already primed to understand that like this is where a lot of the risks of the sector's growth and the market correction might lie. We also want to provide through the report and understanding that these are other pathways that policymakers ought to be paying attention to what happens to the ecosystem of developers and energy assets around it, what happens to public budgets or might happen. And then of course the wealth effect that American's wealth is a lot of it is in the stock market, what happens there. And the banking sector is also a little bit involved through a indirect, in an indirect way. But we sort of want to make sure that these are all mapped out to consider.
- [45:39]
C
This episode of Shift Key is brought to you by Uplight Power demand is growing at unprecedented rates, creating a critical moment to meet load growth with customer owned distributed energy resources to ensure reliable, affordable and sustainable energy future. Uplight is a clean energy technology company that unlocks grid capacity by activating customers and their connected devices to generate SHIFT and save energy. Uplight combines personalized customer experiences with an AI driven flexible load management platform that harnesses customer ders at scale to maximize program value. Uplight integrates energy efficiency, time varying rates and virtual power plant programs into a single cohesive portfolio. These solutions help meet surging demand while improving grid resilience, accelerating decarbonization and reducing costs for energy providers and their customers. Uplite serves over 80 energy providers, including eight of the 10 largest North American utilities. Learn how Uplight solutions can help your utility meet the moment with existing customer assets@uplight.com heatmap that's uplight.com heatmap or look in the show notes.
- [46:52]
A
We haven't really talked too much about the role of private debt in this space. I mean because of the financial crisis and some of the reforms afterwards, public banks are not playing quite the same role in this sort of financing boom as they might have. But there's large basically private banks that are not as transparent, don't have to be because they're not publicly traded and not reporting, they're not commercial banks, they're not reporting under the same kind of reporting rules to the SEC or to the banking Commission. They're financing a large portion of this debt. So can you talk a little bit about is There a potential risk that is a systemic exposure point for the broader economy that could trigger something like a financial crisis. Or is this still something that's much more contained than say the challenges around the housing crisis in 2007, 2008?
- [47:32]
B
I lean more towards being pessimistic about this. I want to say upfront debt is not a bad thing. Debt is how things get built and it's not something to be afraid of just because it exists. And I think we have to, we ought to not be so alarmist about debt itself being a problem. The fact that a lot of hyperscalers are using what's called sale leaseback transactions to keep debt off of their balance sheets, especially over the long term, this makes sense. That's how we build renewable energy too, for what it's worth, sale leaseback transactions, so on. Now all this being said, the reason that private credit is taking on much of this market and it's not a very transparent sector, that's a consequence of post 2008 financial regulations and I think changes in how banks want to interact with the market versus how what used to be called shadow banks or non bank financial institutions want to interact with the market. Banks are sort of unable to take on really risky kinds of loans, particularly construction. Which is why the way banks are actually playing in this market is in the very end of this process. At the asset backed security, the commercial mortgage backed security, commercial real estate side of things. We actually don't have a sense of how exposed banks are to data center commercial real estate. I assume it's growing because banks exposure to commercial real estate is already particularly high. But private credit, these are firms like Blue Owl, Apollo. This is becoming more and more of a systemic risk insofar as it's just not transparent. And that could be. Well, it just doesn't let us know where our money is going. That's what the transparency problem is. But also it's a place to park capital for university endowments, workers, pension funds. A lot of our savings as well flow through the stock market into real estate. Investment trusts, private credit for insurance companies invest a lot in private credit. There's an increasing amount of capital that investment managers in the economy manage on a lot of people's behalf. That is now in the private credit sector. And that sector, it has a lot of use cases, it invests a lot in assets that banks couldn't invest in. It provides capital on flexible kind of creatively structured terms in ways that allow firms to grow in ways that allow markets to get standardized. It's on its own, inherently, I think this is not like a bad thing to do, but the amount of money that's going into it, the lack of transparency around it, the fact that private credit is also exposed to the GPU financing side of the sector and the long term like construction and operation side of the sector, leads me to sort of believe that if there's a flashpoint, it's probably here. And the fact that private credit is the sort of driving force behind the increasing amount of debt being used for the sector, there's a lot of folks sort of arguing that debt is not a major part of this just yet. Hyperscalers have a lot of cash, they can plow it into the sector. But hyperscaler spending is actually what's driving a lot of their partners in the construction space, the rental space, to actually issue their own debt. And maybe not even that. Just I think two weeks ago, Meta, I think, promised to issue $30 billion worth of its own debt for this purpose.
- [50:08]
C
I think Amazon announced a huge, big debt offering literally today on the day we're recording this, on Monday, November 17th. Yeah, three hours ago, Amazon announced that it was going to raise, that sought to raise $12 billion through a U.S. bond sale. And can I ask one, when you say that our money is in this, would this happen in like the 10 to 20? I guess I'm assuming our listeners are relatively young, but the portion of people's retirement savings that are in bonds, is that kind of flowing through private credit now?
- [50:37]
B
The portion of our retirement savings? To some degree, yes. I don't actually know the exact number or exposure.
- [50:42]
A
Well, there's two channels. Yeah, there's, there's exposure to private debt through those bond markets that you might have in your retirement account. But there's also then an increasing role of index funds, right. ETFs that are indexed against broad market indicators. And now the market indicators are all overweighted to these large tech companies as well. So that's another, another channel where a lot of people. Yeah, like if you own the s and P500, great 500 companies, nice broad index. But actually that's like half is those five or six tech companies right there. And so so much of their equity valuations now are premised on their role in the future AI economy. Although we should say they are pretty diversified. I mean, the conventional Googles, Microsofts and Metas, they are fairly diversified in their revenue streams. And so while they might get revalued, it's not like they're going to have nothing left because the Only thing they.
- [51:30]
B
Do is this is not the end of the tech sector.
- [51:33]
C
Yeah, I did kind of walk away from your report thinking, you know, the conventional hyperscalers, the conventional big tech companies, Google, Meta, Microsoft, Amazon of the world, it is in their interest actually to pump this thing as big as they can get because they know they have the cash flows to walk in at the end of the boom and buy up all these assets that everyone else is betting on. And it's these small independents that are kind of being taken for a ride while Meta or Google or any of these, they have cash flows to count on. They're not solely dependent on the AI industry and they have long term debt. They're big existing companies that have been around for a long, long time and they're going to be around to do the kind of distressed asset purchases that the market might, might allow them to do in the event of a kind of bubble deflation or bubble popping.
- [52:25]
B
Yeah, I think that's correct. And again, like I said, a market correction is by no means the end of the tech sector. Right. Like they have cash flows in all kinds of, in all kinds of services in the cloud. Advertisement the use of, I think database platforms. We're not in a zero revenue environment, so to speak. It is interesting that it's kind of unclear and we have to piece together from public information how much of this revenue comes from AI related products, what that looks like. There's like annualized revenue metrics that aren't actually like properly accounting standard metrics. All this being said, right, these are some of the biggest companies in the world on the market that we already invest in for a bunch of other reasons. It's interesting that I think we can expect that where AI services are concerned, a consolidation of the sector might have those services better integrated into services in the Internet we already use. It might have, for example, OpenAI, which is a tie up with Microsoft, folded back into Microsoft. It might have the NEO cloud sort of being bought up by the likes of Google, which has already made some investments in companies like Terawulf and coreweave. So we're likely to see a kind of consolidation like this. And if it helps to provide a positive spin on this, the consolidation in the fracking sector took a few years, but after 2020, a bunch of the main oil and gas names that bought up a lot of those assets, they used capital. Discipline is what the term is in academic economics. And the sector turns a profit, at least it does now. So I think by no means is this the end of the tech sector. It just does mean that there's going to be a lot of extra assets sitting around the country not being spent on in the event that everything here rationalizes as everyone actually figures out what the enterprise business model of the sector looks like.
- [53:55]
C
Let's say you're a Democratic senator listening to this podcast. As I'm sure.
- [53:58]
B
As I'm sure senators have time to do.
- [54:01]
C
Yeah, multiple currently are exactly. One of the takeaways I took from your report is that like, if there were to be a moment where there was some market correction, there was a sense that there was a bailout needed, the kind of priority of federal policymakers should be these tech companies are going to be fine. The priority should be going in and making sure that it's those core innovations around electricity, around handling the scale of load the data centers are able to. Data centers have. And it's around the energy assets that data centers are indirectly financing that is like the priority for federal relief. That basically how policymakers should think about this is like we need a troubled asset relief program for the batteries, for the kind of raw stuff of the clean energy transition that we can salvage out of this otherwise massive moment of speculation. Is that a fair summary?
- [54:57]
B
Yeah. I want to make sure that this conversation around public distressed debt investing essentially remains in the forefront of policymakers heads. We've done it before during recessions. Again, the Minsky moment is really when you choose who loses their shirt, who doesn't. And I think right to Jesse's very legitimate concern that these are gigawatt scale data centers. They're extremely large. There's a lot of energy being built. It's not clear what could possibly, possibly go there. I think that's a. It's a legitimate threat that we actually might not have enough in the near term or medium term to actually fill that gap in the economy. Which is also why it's important for policymakers as they consider how to put together, I think, a troubled asset relief program or how to set up a public distressed investor, that they find ways to actually stretch the cash flows of these assets or stretch the earning potentials of these assets over the longer term, find ways to hold them in reserve like private equity firms already do for market opportunities later on. There's lots of creative ways for the public sector to invest in assets that otherwise, I think a lot of private sector assets might be unwilling to touch due to just a bad market environment or lack of risk tolerance or so on.
- [55:58]
C
Well, and what I was going to say is that when I think about the role of private credit in the market, and I think about who's buying solar and wind farms as they come up, as they reach completion. It's a lot of these same actors, right? Like, the risk is, if you totally obliterate private credit, is that you also obliterate the pipeline for renewable construction.
- [56:17]
A
I mean, it's definitely driving a lot of the additional expansion. Now, the sort of the headwinds from the loss of tax credits are in many ways being countervailed by the tailwinds from the massive sucking sound that is data centers gobbling up all the power they can get everywhere they can get it. So that certainly is a key dynamic. And if that ends, then it certainly could result in a slowdown in renewable energy construction or just energy construction more broadly. And so maybe that's a point where we might need to revisit interventions in the market, right. To think about advancing technologies that whatever policymakers prefer at the time want to advance.
- [56:51]
C
Right.
- [56:51]
A
I think that the big question is which technologies there is, because the gas turbine industry is also in this tough moment where they're basically enjoying nice margins and they're at full capacity and they're trying to figure out how much they can expand their manufacturing capacity without eroding those margins, without ending up out over their skis. If the bubble pops and they end up with a lot of excess capacity, right, as they did in the 2000s already, they'll also be hit, as will the coal plants that would have closed absent this large increase in electricity demand. If that demand doesn't materialize, then there'll be more pressure to retire old coal plants as well. So I guess at its core, I mean, if we lose that big increase in demand growth, that effectively will lead to a downshift in the forward price curves, right? The expectations of future prices in the power markets, that's driving a lot of that investment right now, driving a lot of the maintenance of these coal plants that would otherwise not be worth it, driving all that investment in new gas turbine capacity and the expansion upstream. So it would have pretty broad systemic impacts in the power sector, but not disproportionately for renewables.
- [57:49]
C
I don't think.
- [57:49]
A
I think it would be broadly across the market. And I'd be curious. It depends who's in the hot seat at the time in Congress or the White House as to which technologies they think are most worth bailing out in that context. But you could easily see very different ways in which that would play out in a kind of distressed moment.
- [58:05]
B
Well, there's one really Good example in the clean energy space that I think is worth consideration. The New York Power Authority, which was I think empowered by legislation in the governor's budget a few years ago to actually build renewable energy assets, is backstopping a lot of struggling assets, private assets, put them on its balance sheet and now has a 7 gigawatt supply, 7 gigawatt pipeline of renewable solar and battery, mostly in some wind assets that they're building. These are projects that they identified, but also a lot of projects that private partners brought to them and said we can't do this due to the changing market environment. Do you want to pick this up? And all of a sudden the New York Power Authority is contracting with existing clean energy developers to build their projects, except it's now under the New York Power Authority and not under the developer. That's I think one model for thinking through how the public sector can actually backstop investment in these renewable energy assets. That would have to continue, but developers might be losing their shirts over in the event of a market correction. More broadly, I think we should remember that electrification is happening anyways regardless of the data center boom. And I think we want to make sure that electrification related outcomes, replacing stuff with batteries with heat pumps, working on better demand response and flexibility, policymakers should still be attuned to preserving and growing those sectors and making sure they're not dependent on this one particular growth sector, which is tech. All the while I think utilities will adjust their IRP and we'll see a flight back to quality of electrons. For example, the quality, if you will, of a coal fired electron is a lot less for a utility or grid manager than even a CCG like gas turbine or a solar battery produced electron.
- [59:33]
A
Yeah, I think there was one maybe area where there's a particular exposure or risk here and that's for some of these emerging clean firm power technologies, particularly I think nuclear fission, the advanced fission companies, or maybe even some of the fusion companies like Commonwealth Fusion, that really seen a huge investment of private capital now due to the potential for them to serve as a key supplier, data centers, and a higher perceived willingness to pay for these emerging technologies than you might have assumed three years ago. Some of that is quite speculative. Right. And like in any moment would potentially be at risk. But I think one of the key challenges is going to be to see what happens in that moment. Will, you know, will promising technologies that could be key parts of the next phase of decarbonization or the next phase of American energy supplies go down with the ship.
- [60:19]
B
Right.
- [60:19]
A
If they are in A moment where they're trying to make substantial capital allies to say, get through the NRC processor, build their first demo reactor, and that's sort of financed on the back of equity expectations that they're going to have this big customer at the end of that process and then that expectation vanishes. That could put some of these companies in quite a pickle because they aren't well capitalized and their, their runways for cash are limited. And so all it takes is a misstep and they may run out of Runway. So I think that is another area where policymakers need to be careful. That isn't to say that every, every tech company building a new reactor deserves to succeed. I mean, that's certainly not the case. But we should make sure that if that does happen, that those advanced energy technologies are not all sort of systemically at risk of major setback in that kind of circumstance.
- [61:03]
B
I will say, and this is something that CPI has been working on for a little bit, especially where nuclear and geothermal are concerned, the returns that we're seeing on the surface stock market for tech companies are effectively the collateral that's bumping up the valuations of companies like Oklo or the Fermi reactor data center complex. The collateral for those valuations is the growth of the tech companies. And I will say that I think it's going to be challenging for companies in the startup space, in the clean firm startup space, to take a lot of their booming equity valuations and use that as kind of like apparent guarantee for project finance. That's a very challenging thing for startups to do when they're looking for like software level returns, even though it's like very much an infrastructure sector. But to reduce that equity valuation through a market correction is not good for them either. It really does hurt their ability to undertake the necessary project finance, which is a totally different kind of capital expenditure in the process. I want to make that distinction between the kinds of capital here clear and.
- [61:53]
A
Just without naming any particular names of companies, it is worth saying that some of these valuations are insane. There are potential for individual company market corrections absent this broader miss Broader boom.
- [62:03]
C
Well, and I was talking to a former Biden person the other day and they were saying they look at the Trump administration and how the Department of Energy is acting around the oklos of the world, these new providers of, of firm power that they would like to eventually make clean. And they're like, Trump is making all the same mistakes that the Biden administration and the 2020s clean tech and climate tech boom made in that they're loading up these companies equity valuations and they're capitalizing these companies quite well. And then the company gets to the point a few is going to get to the point a few years down the road where it's like, okay, time to build some projects and it's not going to want to build it with equity capital and it's going to be not have enough cash on hand to finance. It's going to need project finance like everyone else does. And it's going to be kind of stuck in the same position that we saw some of the climate tech companies get caught in recently.
- [62:57]
A
Yeah. And they have said that they'll step in with, you know, loan programs, office loan guarantees to try to play a major role in helping some of these advanced reactor companies get it over that challenge. But it is a big challenge. I mean, try to get a loan right. Project development loan for a new reactor technology with zero proven operating history.
- [63:14]
B
If I'm a credit investor, I'd be very, very scared of this.
- [63:17]
A
Yes. That's a pro forma that it'd be very difficult absent some very serious guarantees of government backstabbing.
- [63:24]
C
Advait. We've done a great job of kind of dancing around maybe the biggest question of the report. And I would say you do a very good job of playing coy about it. In the report where the report's titled Bubble or Nothing, you actually don't come out and say whether you think this is a bubble or not. And of course it's kind of a weird bubble too, because I think we've all. There hasn't been a moment where these leaps in equity valuations for the hyperscalers has happened where people haven't been like, boy, it looks kind of bubbly. And if you remember Back to the 2000 and tens too, people were worried about a tech bubble then too. And it turned out that it wasn't a tech bubble. It was just a rapidly growing and healthy part of the economy. And so what I wanted to ask you to was, number one, did you walk away from this report and from your conversations with investors and creditors, policymakers thinking it was a bubble, number one. And number two, is this like unusual that we have a bubble and we kind of can't stop talking about how bubbly it looks, or is this a new type of bubble where there's a bubble happening and we all know it's a bubble?
- [64:22]
B
I will not personally say whether or not I think this is a bubble. I do think though, that the fact that so much of Our attention is centralized around it. It testifies to a kind of new way of the media's relationship with the economy. And not even the media just in general, but the fact that the federal government is interested in this being the next industry of the future, the fact that I think we haven't had too much else to talk about in economic news due to the dominance of the hyperscalers and Mag7 in the market, the fact that they're the collateral for improvements in the energy system, and even some people are blaming them for the affordability crisis. I think it's very easy to get into a headspace where we're all paying rapt attention to the day to day stock movements of these companies. I do certainly think that I don't know what it was like necessarily to be following the news in the dot com bubble, but I do certainly think that the amount that we've all been talking about it at the same time is very striking to me. I think it's important as well to recognize that bubbles have psychological motivations, more so than just pure economic motivations. Of course, from the perspective of a policymaker and someone who's done credit analysis for stuff like, I obviously look at these firms and look at their lack of revenue and think this is dangerous, this could be getting over their skis. But a lot of companies have gone through this point and made it out. That's not to say that these companies will or won't, but the fact that so much of the market moves in response to the leading tech companies, there's the degree of what asset centrality and crowding and extremely high relative values relative to historical values. It makes me think that like there's something to watch out for, anchored by the fact that a lot of the people leading this investment boom, whether it's the federal government seeking to promote it or whether it's the leaders of these companies, the CEOs envisioning some kind of vastly different future for the the economy. There's a psychology to it. I think Keynes would call it like animal spirits that's pushing this investment boom the way that it's going. I won't say whether it's a bubble or not, but we have to look at the risks for what they are.
- [66:13]
A
I just think it's wild how quickly we went from like, oh, this data center thing seems like it's overhyped, which to be fair, we've had an episode on about a year and a half or two ago to oh my God, this is actually real at such a huge scale. To, well, this seems pretty frothy. There looks like a bubble going on here. When will it burst? And it's kind of amazing just how quickly those media narratives and kind of conventional wisdom cycles through. But it is remarkable that despite all the talk about the potential for a bubble, the investment seems to keep plowing ahead.
- [66:41]
C
And there is something different too, which I have to say I did not. I think when we did an episode early on, Jesse, that was kind of like, huh, is energy for computing really going to be this massive driver of energy growth? Like we've worried about this in the past, it didn't happen. What feels different about AI, which it's taken me kind of, which I understood but did not fully hear articulated until relatively recently, is that like if you're Google and you run Google Docs, the more people who use Google Docs, the more money you make. There are incredible economies of scale around energy use with Google Docs. Right? The marginal user of Google Docs does not cost you almost nothing, costs you almost nothing in energy terms. But the marginal user of Claude OR Gemini or ChatGPT, energy use scales linearly with that because these are. Because those products are so computation intensive and that is a relatively new feature of the entire energy landscape and the relationship of these technology companies to the energy industry. And it's something that obviously we've been talking about now for months, but like it is such a marked change from even what I think we kind of how we thought this might go a year ago.
- [67:50]
A
That seems like a good place to.
- [67:51]
C
Leave it, and we're going to have to leave it there. Adventurun, thank you so much for joining us on Shift Key.
- [67:55]
B
Thank you so much for having me.
- [67:57]
C
And keep us posted on your research. I'm curious what you thought.
- [68:11]
A
I guess I'm still struggling to figure out if there's a systemic risk here. I think it's pretty obvious that the investment in the sector is frothy, that people are racing to get ahead of competitors and to stay on the frontier and to corner market share, not racing to keep up with fundamental demand. And that's going to lead to a correction at some point that can't sustain itself forever. Of course, if one could predict exactly when that was, one would be an extremely wealthy person shorting that market at the right time. So the hard part is not will it burst? Or the question is when? But I'm still not convinced that necessarily and I'm willing to continue to explore this and be convinced that there is a broad systemic exposure here. You know, at the level of financial crisis territory. I do think the over indexing of the stock market and our current like GDP growth because the rest of the economy is actually not doing well thanks to tariffs and other things that there is a risk that we tip into a mild recession if construction spending slows and investment in GPU slows and new power construction slows like that is enough probably to tip us from a moderately growing economy into a moderately shrinking economy. The big question, is there a broad sust risk here that would be much worse than that, that would push us into something deeper, more like the 2008, 2009 financial crisis. And I still don't quite see the evidence for that yet, which isn't to say it's not out there. I'm not the world's best market analyst. That's certainly not my forte. But I didn't quite see it in this report. There are some concerning signs about the mismatches between assets and liabilities, but not necessarily the idea that this is going to be some big broad ripple effect that brings down the entire economy.
- [69:39]
C
No, and I think in some ways the report was helpful because it helped me understand that like this is not the same. We're not looking at the same setup right now. If this were to fall apart, this kind of engine, it'll do so differently than the last. It'll do so differently and in some ways it would do so in maybe like an early 90s recession way. I mean one dynamic we haven't talked about that we didn't talk about with the debate is like data center construction is exceptionally geographically concentrated. It's really, really for a engine of economic growth at the moment, it's like exceptionally limited to just a few places. Now interestingly, one of those places is a part of the country that has not seen a lot of like construction and physical capex driven growth lately. And that is the Northeast corridor. Like the northeast corridor is seeing data center. I think that's kind of interesting. I mean I live there, so that's interesting. North Georgia is seeing data center growth. West Texas is seeing data center construction.
- [70:32]
A
Parts of the it is starting to spread out now that the grid infrastructure is posing a constraint in the key tier one markets. They're now casting about to these other second tier, even emerging totally new markets just because that's where the power is. So that's interesting. It is driving a bit more geographic dispersion for better or worse.
- [70:47]
C
But I wonder if that's also kind of part of why there's this like interesting kind of two Track economy we see right now where like some parts of the country feel like they're doing fine and some parts of the country are just like utterly already in what seems like a recession, which is to be clear, a normal state of affairs, like most states are passing in and out of recession at one point or another. But it could explain other things we've seen in the news lately, like the collapse and economic polling for the president or something like that. Anyway, interesting to watch. One question. Okay, one last thing, which is that like, let's say that these big drivers of new load dropped off the grid like a few data centers didn't get built. When you think about the merit order for electricity, like is that good or bad for emissions? Right. Could it be good because you basically all your renewables go first and then what? You lose? Dirty stuff.
- [71:33]
A
Yeah, it's good.
- [71:34]
B
Yeah.
- [71:34]
A
I mean we should be very clear. Any electricity demand growth, especially large rapid electricity demand growth that is not helping decarbonize a major part of the economy like EVs or heat pumps, is not good for emissions. It just makes the challenge that much harder. It is giving a lifeline to a number of advanced energy technologies like advanced geothermal or demand flexibility, other things that maybe will in the long run kind of get enough juice that that'll pay off in terms of future emissions. But in the short run, what data centers are doing is shifting us from a world where wind and solar are meeting moderate or no demand growth and then eating into existing market shares from coal and gas and driving down emissions into a world where wind, solar and storage are racing to meet the growth in demand and then having nothing left to to drive out coal or gas. And so we still have just under a couple hundred gigawatts of coal fired power plants in the US that three years ago looked like they were on their way out, thanks to a combination of flat demand, growing renewables and EPA regulations, and at the moment now have a very new lease on life because power prices are up and they're being dispatched more often in a variety of markets. And capacity prices are now arguing for new entry and not just maintenance of existing capacity. So it's not great from an emissions perspective, at least in the near term.
- [72:47]
C
And with that, that's a perfect segue into upshift Downshift, our weekly look at climate and energy news where Jesse and I pick one item of interest. And if it's making us feel more upbeat about the energy transition, it's an upshift. It is making us feel more downbeat. It's a downshift. Jesse, on that note, what did you bring for us today?
- [73:03]
A
Well, it's actually quite good segue there because I have an upshift tourist, which is another report from Ember. We've talked about their data a lot. They've been cranking it out lately in very helpful ways. That solar and wind effectively covered all global electricity demand growth Q1 through Q3 of this year. So, you know, year to date, as of the end of September, total fossil fueled electricity generation was down modestly about 17 terawatt hours. Hydro was also down about 54 terawatt hours. All of the growth came from solar, wind, and then a small slug from new nuclear. And so we are seeing total emissions from the global power sector decline this year, along with the share of total global generation from fossil fuels, despite an increase in demand of about 2.7% globally in electricity consumption. So this, again is the kind of thing you have to see at the top of the plateau before you start to come down. The other side is that we have to grow renewables and nuclear and other clean electricity supplies fast enough not only to meet that demand growth, but then also to squeeze out the role of existing incumbent fossil generators. That's the part that actually drives down emissions as opposed to just avoids emissions growth. And of course, China and India are essential to that story and a big part of driving this shift globally. With fossil generation in China down 1.1% in the first three quarters of the year and actually down 3.3% in India, thanks to both record growth in solar and wind and mild, milder weather, which we should say is a global trend. Overall, there were some heat waves last year globally that pushed up electricity demand by about nearly 5%, 4.9% year on year in 2025 or 2024. This year so far, in the first three quarters, we've only seen 2.7% growth, and that's in part due to milder weather so far this year. We shall say that Ember expects through the end of the year that trend to continue with renewables covering all demand growth and fossil generation down slightly on the year globally.
- [74:53]
C
Love to hear it.
- [74:54]
A
Yeah, it's good to have some good news every once in a while.
- [74:56]
C
Yeah, exactly.
- [74:58]
A
What about you, Rob?
- [74:59]
C
I have a downshift. I have a downshift. It's a New Jersey downshift, too. Womp, womp. There you go. My colleague J.L. holtzman had a great scoop in the past week about how yet another offshore wind project on the coast of Jersey has been killed. I feel like this is kind of maybe anticipated, but this is now the nail in the coffin. It has happened. The leading light wind offshore project was going to be about 35 miles off the coast of New Jersey and generate 2.4 gigawatts of clean electricity, has filed a letter with the state regulator saying that it sees no way to complete construction and wants to cancel the project. This is driven partially by changes in the Trump administration that's changed. It's driven partially by the higher cost of offshore wind. It seems like offshore wind now is much more expensive than we thought it might be in the late 20 teens. Something to talk about in a future episode perhaps. But it means that at this point all the offshore wind projects off the coast of New Jersey are canceled and the only one on the northeast corridor is Empire Wind, which soldiers on seemingly due to a rumored deal between the governor of New York and the Trump administration to preserve New York's development offshore wind in exchange for allowing some gas pipelines to go through, which were then actually accordingly approved last week as well. There was a gas pipeline that is going to run from New Jersey into Brooklyn and expand capacity for the gas network in Brooklyn. And then there's another gas pipeline that I think is just on the verge of being approved that's going to run and extend a mainline natural gas pipeline into New York State and potentially then into New England, which is an interesting deal that if we ever get more reporting on it, we can discuss on a future episode of Shift Key. But I did want to note here first, my colleague JL's great reporting, which you can find at Heat Map and of course in the show notes, but second, that the dream of offshore wind in New Jersey I think is officially dead now and there's no more active projects. And maybe it's AP1000, maybe it's nuclear's time to shine.
- [76:53]
A
We'll see. Yeah, I mean it was, that was the vision was for the offshore wind to provide about a third of our future 100% carbon free electricity mix. And so we are very much right back at the drawing board now in New Jersey with a new governor, I should say elected recently as well, will be coming in, I think, with a mandate to think more about affordability and maybe revisit some of the policy priorities and strategies, but with a clear reset on offshore wind. That was one of the key pillars of our entire strategy in New Jersey that just has not materialized. And so what will replace that is a big open question, one we could talk about more in the future as well.
- [77:25]
C
One of the things along the way became exceptionally polarized. Became.
- [77:28]
A
I was going to say, so, yes, yes, in New Jersey, I mean. So one of the things I think would be interesting is the go back and rewind time and what if Leading Light Wind had been the first project approved, A project that was 35 miles offshore, far enough that you cannot see the turbines from shore, as opposed to Atlantic Shores, which was, I think, at its closest point only like six or seven miles offshore would have been very visible and I think fueled the huge amount of controversy as the first project. And I think this is what happens when you let the economics dictate which one goes first, because obviously the one that's in shallower water nearer to shore is going to be the cheaper one that will win the auction first in the first round, and then you'll auction off the more expensive ones that are further afield later. But from a public perception and acceptance perspective, perhaps we should have inverted that order of operations and might have been in a very different place. Now, of course, hindsight's 20 20, but. But I think it is one of those cases where when you're thinking about how to build out a nascent industry, social license is pretty darn important. Not just which project is the cheapest to build.
- [78:24]
C
We also should have included some kind of cost inflator in those contracts too, if there was going to be a.
- [78:31]
A
Once again, if you can time the market and you knew there was a giant macro inflationary cycle coming, then.
- [78:36]
C
Exactly. Well, that will do it for today's episode of Shift Key. Thank you so much for listening. You can follow me on X. Obviously, Meyer and my colleague Jesse Jenkins. Essie Jenkins. You can follow both of us on Bluesky under our names. Just type them in.
- [78:50]
A
You'll find us.
- [78:51]
C
You'll find us. Or LinkedIn. There are so many different social networks you could use. Shift Key is a production of heatmap News. Our editors are Gillian Goodman and Nico Loricella. Multimedia editing and audio engineering is by Jacob Lambert and by Nick Woodbury. Our music is by Adam Komalow. Thank you so much for listening. We'll see you next week.