How the Hedge Fund Magnetar Is Financing the AI Boom

0:00Your best bottling plant employs 3 ,300 people. How do you get 3 ,300 people working at peak efficiency? Your best store has reduced waste, water, and energy usage. How do you make every store like your best store? Your best property has every guest raving. How do you make every property like your best property? The answer is Ecolab. Better performance, better outcomes, better impact. Ecolab. Now every location is your best location. How many vendors does it take to meet all your organization's food needs? Just one. EasyCater, the workplace food platform that lets teams order from a huge variety of restaurants, over 100 ,000 nationwide, all through a single vendor.

0:45In addition to all that variety, EasyCater also gives you full visibility of your organization's food spend with invoicing, centralized reporting, and seamless integration with expense management systems, all on one platform. EasyCater, your business tool for food. To learn more, visit EasyCater.com slash podcast.

1:08Bloomberg Audio Studios. Podcasts, radio, news.

1:25Hello and welcome to another episode of the Odd Lots podcast. I'm Tracy Alloway. And I'm Joe Weisenthal. Joe, AI is so hot right now. In the immortal words of Mugatu, AI is so hot. It is. Yes, it is really hot. But, you know, you hear something, oh, there's a little bit of slowing down and some of the progress on the models. But the recent NVIDIA results speak for themselves. There is nothing that I've seen yet that would suggest that this macro trend, at least as an investment trend, and I'm not talking about stocks per se, is anywhere close to, quote, slowing down. Yeah. And the interesting thing is we seem to be having more and more players, some new types of players that are getting into the space.

2:10So, you know, we have AI funds kind of launching left and right. And one of the newest players is a hedge fund called Magnetar. And I know them like primarily for credit stuff. I think they were big in red cap trades for a while. Yeah. And now they're launching an AI fund, a VC fund, which is kind of unusual for this type of hedge fund to do. Totally. I mean, I've heard of Magnetar for a long time, obviously going back to the early 2010s at least. And look, I'm not surprised that various investors are looking for what is their distinct way into this space. And of course, look, we've done interviews with VCs of various nature and positions in the past.

2:53And so I guess, you know, there's sort of two questions to my mind anytime we're going to be talking to someone investing in early stage or any stage of AI, which is obviously what is the thesis? What's going to win out? Where will value accrue? But then from an investor perspective, given so many entrants into this space, specifically whether on the public equity side, whether on the private side, whether on the VC side, early stage, late stage, what do they as a fund or an investor bring to the table or will be able to see that the other billions of dollars competing for AI profits do not see?

3:29I have a slightly different question, which is for these types of investors, like how much is it about how good the technology is that they're investing in versus how much is it about getting in the right position in the capital stack? So that's a great question. I think it's going to be really interesting to talk to someone who's coming from this perspective. And without further ado, we have the perfect guest. We're going to be speaking with Jim Prusco. He is a partner and senior portfolio manager on Magnetar's alternative credit and fixed income team. Jim, welcome to the show. Thank you. Great to be here.

4:05So how does someone on a hedge fund's fixed income team get into AI? Well, we have a long history of investments in private companies really dating back to an increased focus after the financial crisis when spreads and yields got tighter and the private markets seem more interesting. And we've often partnered with platforms where we thought we could grow the platform and generate an interesting asset, either a pool of cash flowing assets or help grow the company and participate in that growth and support them through financing and other things like we can support them through helping them with hiring or accounting or other systems they need and just to help them grow generally.

4:48And so I've been doing that a long time, and we've been in a number of areas like auto lending in Ireland. And then we've moved into various fintech companies. We were one of the first institutional investors in Open Door before they went public. We were supporting and investing in a very interesting fintech that is financing restaurants right now. And so we felt we had experience in that space. And then that sort of overlapped with our relationship and our investment in CoreWeave, where we were the first institutional investor in CoreWeave in 2021. So we were very early in the trend of putting capital into the AI infrastructure space.

5:32And that's just sort of grown as this whole market has grown to encompass literally everything now. You know, we continue to look for smart ways to invest. And, you know, one of those ways we felt was what can we provide that's of value? And one of the things we can provide besides the general help we can give a growth stage company is compute because that is the scarce resource right now. And that's where all the capital is going to the various parts of the value chain to deliver compute. And so there's a competition to get compute. And if you're a smaller company with limited capital or limited access to capital, it can be difficult to get that.

6:14And so that was sort of the value proposition we thought we could bring to bear. Joe, I have this vision in my head of VCs, like going into startups, burying baskets full of chips. Yeah, exactly. Instead of just saying that, like, our pitch is the relationship and the coaching aspect. We have access to the chips or the energy plus chips. Just for point of clarification, listeners should know, we've talked to CoreWeave at least twice on the show. And it feels like in the AI space specifically, this is one of those names that's a very big deal, but not many people don't know it the way they know, say, an NVIDIA at the very back end or a chat GPT at the very front end.

6:55But they build a lot of the AI data centers that are filled with NVIDIA chips. I want to get more into the business model there because I have a lot of questions in the business of selling compute, et cetera. But talk a little bit more about – you said your experience in the private side is like this expertise with platforms per se. And when I think of platforms, I think of companies that can acquire lots of other companies or a lot can be built onto them. Talk to us about how the platform-specific expertise informs your thinking with a core weave or any other AI investment that you're making now.

7:32So we've tried to put capital into companies that are trying to build their business in a particular space. And oftentimes that could be a space where they generate a cash flowing asset, like in the auto loan example. In the open door example, they were acquiring real estate, which was a hard asset. In that restaurant fintech example, they're acquiring restaurant credit. And so we've tried to support businesses that had some kind of asset or cash flow and work with them on a number of ways that we can add value. I think first and foremost is all these growth stage companies need financing. And I think we have great expertise from debt to equity, private to public, and we can be innovative in trying to bring the best, most appropriate, lowest cost capital to these growth stage companies.

8:26And like I said, as well as – So just to be clear, just to understand in this context, what makes AI distinct, say, from other waves of tech or what makes it distinct for, say, a magnetar is in part this distinct capital demands that was not perhaps as big of a deal during the SaaS wave of the 2010s. Yes. And not only a general capital demand, but in many cases for many of these companies, a very specific demand to have capital to deploy with compute. And because they need this very specific scarce resource, helping to deliver that resource, and in particular, helping to deliver that resource in a high-quality way where you have a partner like CoreWeave that has, I think there's a lot of evidence that they have the highest performing AI training cluster.

9:18And so that is really valuable to these companies that might otherwise struggle to get enough compute to further their business model. Speaking of core weave, I'm really curious how that conversation actually started because this was a new and novel thing. I don't think we had chip based loans before, to my knowledge. And I keep hearing that asset based financing is going to be like this next big thing in private credit or it's the last real frontier in private credit. How did you come up with this idea, this deal? Well, asset-based financing is really a classic private credit tool. And there's a number of examples.

9:59Just if you think about my example with the Irish auto lender, if you buy a loan for a car, so the Irish auto lenders generating car loans and those go and you buy them in a vehicle, you have primarily the security of the people paying on those loans. And so you get paid back by the cash flow of the borrowers paying their car loans back. But there's credit risk to that. They could potentially stop paying. And in the case where they stop paying, then you have the car's collateral. And really, that metaphor applies almost directly to GPUs, where if you're a company delivering high-performance compute like CoreWeave has, you're contractually selling that compute to some counterparty that's going to use it.

10:49And in their case, that's often a very large, very creditworthy hyperscaler, but not always. There could be smaller startups that have riskier business models. And in that case, primarily by funding the GPU, you're getting paid back with those contractual cash flows on the use of the GPU. But in the case that company fails, then as backup, you have the GPU itself. Now, the GPU isn't really like the car where you'll probably go out and sell it, but you get that time back on the GPU, which you can then resell to somebody else. And being a scarce asset, you can think about what value that would have in a future time.

11:30One difference that I could imagine with the GPU versus other forms of assets, say whether it's a car or say whether it's a house, is a certain here in 2024 still unpredictability about many things in the future? Will NVIDIA always be the gold standard, so to speak, in AI chips? Maybe it looks like a yes, but it doesn't seem guaranteed. How fast will the current generation of chips that are deployed degrade in value? I imagine there are fairly predictable sort of depreciation curves for cars that perhaps are more uncertain for chips. And then also the uncertainty of actual deployment given permitting and challenges with energy and the other operational things that have to do with a new company building a data center.

12:20Talk to us about modeling or at least thinking through some of the uncertainties with chips specifically? Well, depending what stage you get involved, you have the breadth of all those different risks potentially. So if you're investing in high-performance compute, but it's a greenfield data center, then you have to think about all those things. You have to think about the delivering of the power. You have to think about the timing on all the components to get to the data center. If you're making what we've been talking about, which is sort of a GPU-based loan, then usually that loan is based upon a running GPU in an existing high-performance compute data center.

13:02So you don't really have to think about some of the earlier stage issues. You more have to think about how long is my contract? How good is my contract? What do I think the value of renting that chip out will be at the end of that contract? How much rent on that chip could I get if I had to re-rent that in the middle of the contract? So it's more near-term things on actually having a functioning GPU in the data center. But all those other things have to be financed too. And there's going to be innovative and large amounts of capital dedicated to financing those things. Setting aside the financing for a second, how hard has it been just to find physical space in data centers?

13:46Well, it's been extremely scarce. And a lot of that is driven by the search for power. the data centers required for the new AI chips are much different than the old data centers so it isn't really cost efficient in most cases to go and take an old data center and try to retrofit it because the amount of power just alone that has to go there is you know transcending an order of magnitude more per rack of gpus now and so that's just you just can't really retrofit that efficiently. It's better to build your own building. And so it's really come down to things like permitting, availability of power, and time to get all your components.

14:31And all these things have their own lead time. So it had an interesting back and forth to Brian on curing transformers, all these little nuances come into play when you have to build a data center. And so So because power is really the limiting factor, most of all, you're seeing a lot of moves towards where the power is. And there was recently an article on Bloomberg, I think, about a company in Texas that owns a bunch of land that's now worth$40 billion, right? And that's because they're near all this renewable power. But that isn't the only thing. It's incredibly complex to operate this high-performance compute.

15:10So then you have to think about, if I try to build my data center out there where the power is, can I get everything out there, including operational expertise, right? Can I staff my data center with the kind of experts I need to run this kind of highly technical, high-performance compute? And each generation is just getting more complicated. We're going to have liquid cooling on the next generation of NVIDIA chips, probably immersion cooling right after that. It's very complicated, very expensive. and very difficult to scale, much harder to do in a large size than it is to do in a small size.

15:46Maybe Magnetar can finance a small modular nuclear reactor. No, seriously, because if you're financing the compute and securing that on behalf of companies that you want to invest in, you could go one layer down, finance the energy. And we're certainly interested in that. And we have a history investing in energy. We have investments right now in a developer of utility-scale solar power in the U.S. who has leased some of that solar power to various hyperscalers. So that is certainly a space where we're interested in. I was just in Miami meeting with a company that has a novel heat sink battery technology that they want to deploy to data centers that they're talking to a bunch of data center type companies about launching that product there.

16:32So there's a ton of interesting things. And just like every other part of this ecosystem, it's going to require an immense amount of capital. I guess just since we're sidetracked on the energy component for now, while we're here, novel battery technologies, there's a lot of them out there. There's a lot of startups that have something novel in energy. And often one of the things that they talk about is this chicken and egg problem where they need capital. They need sort of financing of some sort or another to build the stuff, but the lenders don't really want to give it until there's demand and no one is going to promise to buy it until it's shown that it can work.

17:12Can you talk a little bit? I mean, again, I know there's a little bit off track from GPUs themselves, but since you were talking about - Well, it's similar with AI, right? Yeah, since you're talking about batteries, can you talk a little bit about that dynamic as it affects solving the energy side of the equation? Yeah, for sure. And it has some overlap with the way you look at an AI company too. Yeah. If you think about the core things that we really want to look at, it's technology, team, and traction. So does their technology really work? That's first and foremost. What is this product? Does it have some kind of advantage?

17:47And then traction, like time to market. That's super important. I was just talking to IsoCon at Poolside and to him, those are the two most important things. Speed to product, speed to market, because it's a race. And even if you have the greatest technology, if you take too long, someone's going to be using something else. And that's certainly true in the energy space where energy is of critical importance. So I think that for these startups on the traction side, they really need some strategic partnerships. because their cost of capital is very high. The strategic partnership is kind of like an existing company that has a demand.

18:27It also has a lot of cash and could theoretically be a buyer of their solution. Yes, and really on the other side, too. So, for example, because their cost of capital is so high, there's certain things that it's hard for them to do. And one of the things that it's really hard for all these startups to do, and this was true in the recycling industry and other industries, is build a plant. Like very expensive, time-consuming to build a plant. You don't really want to raise BC capital to build a plant. And so it's important to have a partnership on the manufacturing side too. And that was really like the first thing this battery startup that I just visited talked about is like getting that because you've got to be able to deliver your product and you have to deliver it on scale.

19:12And ideally, you don't want to be wasting time building your own plant on that. And then, like you said, on the other end, you want to have a partnership with the users of the energy, which is all the people that either have data centers or use data centers or are customers of data centers. And you want them to ideally put together an attractive financing relationship where, you know, in some form or fashion, they're front loading their payments to you. so that you can use that capital to actually build the product that they need.

20:01Your best restaurant location gets five-star reviews. How do you make every location like your best location? Your best paper mill has been operating at peak productivity. How do you make every mill like your best mill? Your best data center has optimized every drop of water. How do you make every data center like your best data center? The answer is Ecolab. Better performance, better outcomes, better impact. Ecolab. Now every location is your best location. How many vendors does it take to meet all your organization's food needs? Just one. Easy Cater, the workplace food platform that lets teams order from a huge variety of restaurants, over 100 ,000 nationwide, all through a single vendor.

20:47In addition to all that variety, Easy Cater also gives you full visibility of your organization's food spend with invoicing, centralized reporting, and seamless integration with expense management systems, all on one platform. Easy Cater, your business tool for food. To learn more, visit easycater.com slash podcast. So Joe and I went to San Francisco a little while ago and we saw some cool things. I had my first ride in a Waymo and we saw some cool battery related technology. We also saw a lot of VCs. Everyone very excited about AI, obviously. They were also talking about the difficulty of chasing deals right now.

21:26How do you compete with those traditional VCs or are you just not competing with them directly because you're taking the slightly different GPU backed approach? You know, I think it's both. I think you're competing with them and to an extent partnering with them. And that's the thing we had to ask ourselves before launching the fund is, you know, what are we bringing to bear that's value added? And in this case, we're bringing to bear the compute. And so often these startups, even if they're backed by a strong VC, can have a bit of a chicken and egg problem, which is they need compute to develop their product and they need capital to buy that compute.

22:06But if they don't have the compute lined up and the price locked in, then the capital might be hesitant to go in because they'd be like, we could put our capital into you and then it could take you an extra six months to get your compute. And by that time, some competitor has passed you by or the technology has changed. And on the other hand, because they're a startup, they don't really have the creditworthiness to just contract the compute. They most likely have to pay up front. And so we bridge that gap. And so if we go into a fundraising round where there's a bunch of VCs putting cash in, if they know that we're putting compute in alongside them and that the second the round closes, that compute will be available to the company, that makes it easier to raise the cash part of it.

22:52So we are competing and we need that value added to be part of the equation. But also, I think it helps them to raise from traditional VCs because we take that one risk off the table. How big is the market of companies that need compute? Because there are plenty of AI companies that just build on top of an existing model like GPT or Anthropics model, et cetera. How many companies are actually out there and who, like not who are they specifically, but what are the types of companies for whom actual access to compute is an important part of their business? Yes. Well, you know, it starts, of course, with the LLM companies are using massive, huge, huge amounts of compute.

23:39But then if you look at the rest of sort of the AI stack, there's a couple areas where you're going to need compute. And one is all the small model, custom model companies. And small can mean a lot of different things. So you can have some very small companies that are using a very targeted model, like say in a vertical stack, you might have a robotics company that is specifically training a model to run a robot in a particular situation. And that could be anything from a warehouse to doing surgery, right? And they need compute to train that model. Or another one, which is huge and dominated by an existing big player is autonomous driving.

24:25But there are other autonomous driving companies that are trying to be deployed at other automakers that need compute to train those models. Or weather models. There's some really good companies that we've talked to doing weather models. They need compute to train their model. And so that whole model layer, and then even on the app layer, there might be custom elements of small models that they have that sit on top of the big LLMs that they need some amount of compute for. So there's quite a range. It's not everyone. It's more in that model application layer and less in the infrastructure layer that need compute.

25:06So this is one thing I always wonder about AI investment, which is you have a lot of companies that are building on top of existing models, as Joe mentioned. And to some extent, that makes sense because they can save a lot of money by doing it. And realistically, are you going to compete with Google or Microsoft? Probably not. But on the other hand, I always wonder if you're building on top of an existing model, how do you ring fence that business? Because my assumption is if AI gets better, maybe at some point the AI can replicate any AI model, basically. So this is the first thing we always worry about is does some giant company already have this product in a closet with like 20 PhDs working on this?

25:53And somebody, I was just at this conference and somebody coined the phrase incumbent maximalist. Oh, nice. And that's when you think the incumbents are going to do everything and no one else will ever succeed. And I think there's a few use cases. There's things where it's a very specific task that is hard to do well with a giant general model and probably isn't worth doing well. If you're focused on growing tens to hundreds of billions of dollars of revenue, you can't be distracted by trying to do every little thing. And we've seen this in previous tech revolutions as well. And so it can be something that's very focused on a space.

26:33We've seen legal, accounting, sales. There's some great companies that have virtual employees that they're doing, things that are very task specific. There's some companies doing text-to-language and language-to-text and other things for very specific applications. So, you know, that's one way. The other way is data. The greatest ring fence to any AI company or business is data because you've seen as the performance of some of the LLMs has supposedly flattened out, a lot of that is because they've just used all the data. Like they've trained on the whole internet. There's nothing left. And so now you have to have other ways to train or novel sources of data.

27:18So proprietary data is super valuable. And then there just could be areas where they're conflicted. They don't want to compete with their customers right now, although competing with your customers is a great tradition in the tech space. But there could be situations where it's not worth it to them yet to compete with their customers. And so I think there's those different use cases where you're going to see a small number of companies succeed. I have a very stupid question. And actually, I shouldn't even be asking you. I should have asked it the last time we talked to CoreWeave. But since you're here, I'm going to take them all again on the question I didn't ask them.