The circular financing criticism doesn't apply to real demand—there are no 'dark GPUs', enterprise AI TAM was $37B last year and growing, and the unit economics of deploying intelligence are positive.

(00:00:06) Hi, listeners. (00:00:06) Welcome back to No Priors. (00:00:08) Today I'm here with Neil Tawari of Magnetar Capital. (00:00:11) This is a $22 billion alternative asset manager at the center of the AI compute buildout. (00:00:16) We talk about the financial innovation, depreciation of GPUs, and what's next in AI compute. (00:00:23) Welcome. (00:00:24) Thanks so much for doing this, Neil. (00:00:25) Absolutely. (00:00:25) You know, really happy to be here. (00:00:27) So you are leading AI infrastructure at (00:00:30) at Magnetar, you're at the center of the build out, enabling it, financing it. (00:00:35) For any of our listeners who haven't heard, can you just explain a little bit what Magnetar is? (00:00:39) Sure. (00:00:40) So Magnetar has been around for actually, this is our 20th year. (00:00:44) We're an alternative asset manager, and that can mean a lot of different things. (00:00:48) But we have three primary strategies. (00:00:50) The first one is private credit. (00:00:52) The second one is a venture strategy. (00:00:54) And the third is more of a systematic or quantitative focused public strategy as well. (00:01:00) And so I think, when people look at us and why are we here in this moment, especially on building out AI infrastructure, I think a lot of it has to do with kind of our unique lens on helping to build capital-intensive businesses and using creative financing, whether it's venture or other structures with unique elements. (00:01:21) And I think we're going to talk a lot about that, but to build out and optimize the balance sheets for these capital-intensive businesses. (00:01:30) I was hearing about you guys originally. (00:01:32) So you're the first investor I think we've ever had on the podcast. (00:01:34) That's exciting. (00:01:35) Thank you. (00:01:36) I remember hearing about you and Magnetar initially around, I was like, who's this big owner of Coreweave? (00:01:43) And also, you know, helping Open AI with some of their early build-outs. (00:01:47) When did you guys first start looking at the problem and thinking about how to solve it? (00:01:52) Yeah, so we actually, you know, stumbled across the compute problem before it was compute. (00:01:58) We met Core Weave back in 2021, and that was when they were actually transitioning from mining Ethereum into high-performance compute. (00:02:08) And at that time, it was using the GPU as an instrument to mine cryptocurrencies. (00:02:16) And interestingly, that same instrument could be used for high-performance computing applications. (00:02:22) And the first one was visual effects. (00:02:25) So think of things like movies, Marvel movies and things like that. (00:02:28) And so they were transitioning at that point between crypto mining into the first kind of high performance compute use case. (00:02:36) And this was all before AI. (00:02:38) And so we made our first investment before the AI trade started, but we added a lot of optionality where, you know, we could envision a world where the GPU could be used for a lot of different high performance kind of computing applications. (00:02:52) I think (00:02:53) AI was on the radar, machine learning was on the radar for us. (00:02:57) But I wouldn't say that we could foresee everything that happened. (00:03:00) We just happened to be at the right place at the right time. (00:03:03) And we continued to double down as the company progressed and started shifting into more workloads that were machine learning and kind of AI training based. (00:03:12) Did you have like an existing significant data center? (00:03:17) No, I mean, I think, interestingly at Magnetar, there, we have invested across asset classes. (00:03:23) So we've done a lot of property investing, real estate investing as an example, investing in energy. (00:03:28) We had an energy business historically. (00:03:30) And so a lot of the elements for, you know, what constitutes a data center, power, energy, land, real estate, you know, we had a lot of the background in those spaces. (00:03:40) I think we were new to compute, right? (00:03:42) Like that was a new sector for us. (00:03:44) And so kind of those two worlds merging, we obviously, came up on the curve on the compute side, but we had a lot of, background on the elements that constitute what it means to build a cloud. (00:03:58) So you guys just really, you were in this company, you saw the demand and you said like, it's going to grow and we're going to make this a big part of our business. (00:04:05) Exactly. (00:04:05) I think (00:04:06) what was interesting was we made our first investment in 2021, and then about a year later, we continued to see expansion of use cases for, at that time, it was called high-performance compute. (00:04:17) And then it was kind of towards the end of 22, the whole AI discussion started. (00:04:22) And as we entered 2023, (00:04:25) Coreweave started to train models for OpenAI. (00:04:29) And that's when things really started growing because the sheer amount of compute that was needed to train an LLM, this was like the first time it had ever been done. (00:04:37) And what was interesting was what kind of allowed them to take advantage of that opportunity was the historical kind of backgrounds of a lot of the founders were in energy, asset management, (00:04:49) And when you fast forward to today and you look and what constitutes your ability to build a GPU cloud, it's your ability to manage these highly complex assets. (00:05:00) And it fundamentally comes down to access to power and energy. (00:05:04) And so they had these elements with them. (00:05:06) They obviously brought on a lot of talent on the cloud side. (00:05:09) And to put all these together, and at that moment, it allowed them to build very large-scale, reliable clusters for OpenAI and obviously many other customers since then. (00:05:21) And I think the last comment I'll make is, (00:05:23) What really allowed them to kind of win this market early on was focus on two things. (00:05:27) It was scale and reliability. (00:05:30) And I think those were the two things that are really difficult for a lot of the new entrants since then, because scale has to do with your access to capital, your access to energy, power, data center, and then reliability really had to do with their ability to manage a giant fleet of GPUs, which is actually quite complicated, you know, whether it's reliability from, you know, (00:05:53) a few failures or software challenges, building a fleet that can healthily be online all the time at 99.9% reliability is incredibly difficult. (00:06:02) And that's something that they had started back in 2017, 2018 timeframe, and they were at the right moment at the right place with the right technology stack to really build the optimal cloud for that moment. (00:06:15) I've definitely experienced that with our portfolio of companies that are building large training clusters. (00:06:23) Core Weave has a reputation for reliability that not everyone has reached. (00:06:27) Can you just help characterize, if you fast forward, like 2 1/2, three years now, like what is the scale of the problem today? (00:06:34) Yeah, so if you look at kind of CapEx, right, let's starting with that. (00:06:39) So CapEx for AI compute and infrastructure in 2026, you know, at least from the hyperscalers is projected to be between 660 and $690 billion. (00:06:49) And over the next several years, (00:06:52) that scales to trillions of dollars, right? (00:06:55) And so the scale of the problem is how do you build that size of CapEx efficiently? (00:07:03) And I think a lot of that has to do with not only your ability to have access to those core elements, energy, power, and your ability to have data center space, et cetera, (00:07:16) But I think one of the things that's not talked about as much is capital and access to capital and how is capital structured. (00:07:24) And what I mean by that is (00:07:27) This is, billions to trillions of dollars of CapEx. (00:07:31) And just using equity dollars alone is not an efficient way to scale this. (00:07:35) That's obviously massive dilution. (00:07:37) You know, there's, it's not an easy problem to solve. (00:07:39) When we first met, I had like slowly come to this realization. (00:07:42) I was like, I don't think we should take the dilution for the cluster. (00:07:45) Yeah, right, exactly. (00:07:47) And so that's where I think, you know, when you and I have talked about like structuring and I can give a couple examples. (00:07:52) if that's helpful. (00:07:53) I think the first one was DDTL structures or SPV DET structures that had a, think of it as like an SPV. (00:08:03) Inside of the SPV are the is the CapEx, the collateral, which is the GPUs. (00:08:09) and the contracts themselves. (00:08:11) And so in this example, the actual asset or collateral was not really just the GPUs themselves. (00:08:19) It was really the contracted cash flows from, in this case, investment grade counterparties. (00:08:25) And so I think the reason. (00:08:26) This is the consumer of the. (00:08:28) consumer of the computer, exactly. (00:08:29) You know, your Microsofts, your Metas, et cetera, of the world. (00:08:32) And I think (00:08:33) The reason that was done is really twofold. (00:08:37) When you look at the scale of the problem, those particular contracts needed billions of dollars of debt to finance the CapEx. (00:08:47) Obviously, for a nascent and new and growing company, that's really hard to raise. (00:08:51) So part of structuring it this way is ensuring that you have kind of guaranteed off-take on the back end to minimize the risk for debt holders. (00:09:03) And I think that's a lot of what the market got wrong, especially when there was a lot of press about this early on, where it was, there's billions of debt on these highly depreciating assets, and it's extremely speculative. (00:09:17) And what was oftentimes characterized in the media was these debt structures had GPUs as collateral, and that's like putting a used car as collateral, which is obviously just going to depreciate incredibly fast. (00:09:29) You know, that's a very risky kind of structure. (00:09:32) And I think what got missed was the GPUs themselves were actually like the second, second or tertiary level of collateral in those instruments. (00:09:40) The primary collateral was the contract of cash flows from investment grade counterparties. (00:09:46) It's Microsoft or Nvidia or somebody like that saying, I'm committed to pay you. (00:09:51) Exactly. (00:09:51) I know you can pay me. (00:09:52) Take or pay contracts and they're like five years in length. (00:09:55) So I think that was like one feature that's unique to talk about. (00:09:59) And then the second one really has to do with (00:10:02) the debt itself and how it amortizes. (00:10:05) And so in simple terms, when you have debt, you have principal and interest and you have to pay it off over time. (00:10:11) And in these structures, typically the payback period on the CapEx was roughly 2 to three years. (00:10:19) And the structures themselves, the debt was over 5 years, four to five years in length. (00:10:25) where the entire debt amortized during the outstanding period that the debt was out. (00:10:32) And so at the end, you ended up with 0 balance for the debt, and there was no balloon payment or anything that was really due on the back end. (00:10:42) And so the question that often comes up is, isn't that a very risky type of structure, because these things are depreciating incredibly quickly. (00:10:51) So I think there's two comments here. (00:10:54) First is, (00:10:55) On that depreciation question, in these kind of debt structures, it doesn't really matter because the debt's fully paid off by the end of the debt term against committed contractual contracts from investment-grade counterparties. (00:11:09) And then at the very end, the actual upside or residual value, and I know there's a lot of questions on residual value, is held by, you know, the cloud player in this example, right, Court Week, right, or, you know, any others. (00:11:25) And that's a really interesting prospect because you can see a world where all of this CapEx is paid off incredibly quickly, and there's an opportunity to redeploy it, where you can redeploy it without having to pay for any additional debt, obviously, against that redeployment. (00:11:42) How have the instruments changed? (00:11:44) They've changed in several ways, where the first is, and when you look at these SPVs, I think you're starting to see ways to change the portfolio construction of who can go inside of one of these debt structures, and so, early on in the... (00:12:00) early days, these were all only investment grade counterparties. (00:12:04) Because there was, the space was so nascent, the operators had no experience. (00:12:08) And I think now what you're starting to see is a blend of investment grade and non-investment grade. (00:12:13) So like, what does that actually mean? (00:12:15) What that means is, you know, you're seeing these structures with investment grade counterparties, like your hyperscalers and your other corporates that are IG. (00:12:22) mixed alongside some of the AI native companies. (00:12:26) And so think of the AI model companies, the labs, software companies that are building AI startups. (00:12:31) You're seeing those companies get mixed in alongside the IG companies to build a portfolio. (00:12:37) Because now you have the history that you can do this. (00:12:40) And now you have structures where you can kind of balance the risk with (00:12:44) IG and non-IG. (00:12:45) And we're continuing to see that kind of move to be able to help finance, really the model companies and a lot of these startups. (00:12:52) Obviously, that was difficult to do, three or four years ago. (00:12:55) That's starting to become easier as these companies have more runtime and ability to, you know, make the compute fungible. (00:13:02) All our portfolio companies that buy compute tell me it's a supply constraint in the market today. (00:13:10) One, is that true? (00:13:11) And 2, when you think about like (00:13:14) continuing to grow your business or grow this ecosystem, like what's going to stop it? (00:13:20) Like what could slow down a build out? (00:13:23) Yeah, I mean, I think what's interesting is if you look at like 2023, 2024, we were very supply constrained and the supply constraint was chips. (00:13:33) No one could get access to chips. (00:13:35) Yes, we bought chips. (00:13:36) We bought chips, right? (00:13:37) Yeah. (00:13:38) And, there was this thought that, okay, there's going to be an overbuild of chips and then the supply constraints will go away. (00:13:44) Well, fast forward to 2026 and what we see is, there is obviously more availability of chips, but to build and operate these, data centers requires (00:13:54) people, power, infrastructure, a lot of these things that have a lot of bottlenecks. (00:14:01) And so actually taking these chips and then making them into useful revenue generating assets is really the bottleneck now. (00:14:08) It's also not clear that there is supply of chips at the latest generation at scale soon, which is how everybody wants them. (00:14:17) Exactly. (00:14:18) I think, you know, you see (00:14:20) Not only, you're starting to see interesting, and not only just the high-end players want access to the latest chips, you're seeing the latest, you know, obviously startups want access to those. (00:14:28) And I think it has to do with efficiency. (00:14:30) You know, one of our friends or one of your friends as well, Dylan Patel over at Semi Analysis, posted this interesting article last week on inference and inference spend and inference kind of performance. (00:14:42) And there's a lot of, jokes made about Jensen math. (00:14:46) And it was interesting because the. (00:14:47) Seems pretty good at math. (00:14:49) He's actually great at math. (00:14:51) And so for the Hoppers, the H100 or H200 series of GPUs into the Blackwells, there was a claim made that it could be 30 times more efficient. (00:15:01) And I think the data from, you know, some analysis showed that it was 90 to 100 times more efficient in terms of inference performance. (00:15:09) And so I think part of the need to go to these new chips is not, it's yes, more computing power, but it's actually the, it can be cheaper to operate. (00:15:17) It's price performance. (00:15:19) Price performance, exactly. (00:15:21) Yes, my favorite Jensenism is the more you buy, the more you save. (00:15:24) Exactly. (00:15:24) It's actually true. (00:15:26) Yeah. (00:15:27) Crazy. (00:15:29) Help me address like this criticism around circular financing. (00:15:33) Yeah, I know. (00:15:34) It's obviously a topic du jour. (00:15:36) And (00:15:37) I think, the way we see it and frame it really has to do with the demand signals and who are the eventual buyers and how is this being used. (00:15:47) And so, at least from what our perspective, we continue to see insatiable demand. (00:15:55) And if you go back to, the previous kind of big tech build out back in the early 2000s, (00:16:01) There's obviously a lot of fiber that was being built and you had dark fiber, and an overbuild happening. (00:16:07) And I think what you see here is I've, you know, you don't see any dark GPUs, any GPUs used. (00:16:15) And then #2, you're starting to see (00:16:18) actual economic value. (00:16:20) So I think last year, enterprise AI had about 37 billion of total TAM, and it's continued to grow like crazy. (00:16:26) And at least personally, and I'm sure you see this too, but I use these tools all the time and I find it incredibly valuable, right? (00:16:34) The actual tokenomics of positive ROI is actually here now, I think from our perspective. (00:16:41) And so the circularity, you know, comment, I think applies when you're building (00:16:48) speculative computing capacity, or if you're, purely doing vendor financing and it's, you're trying to do some type of, unique, some type of, rev rec type item related to that, and that's not what we see. (00:17:01) Like, what we see is financing to support to build out the demand against use cases that are very positive in their ROI. (00:17:10) And so, like... (00:17:11) Our perspective is that that's not a real concern that we have, and it really has to do with who are the ultimate buyers here. (00:17:19) The ultimate buyers have been at scale, the hyperscalers, they're deploying this. (00:17:25) at scale and the economics are positive when you look at a unit economic basis in terms of deploying intelligence. (00:17:32) And I think we're at a moment in time where you're really starting to see that. (00:17:35) In my own experience, I have been a heavy AI user for several years, but reasoning advances the ability to scale up inference, especially around code, means I'm up against my max limit all the time in a way that was not true. (00:17:51) initially. (00:17:52) How does the inference workloads actually growing? (00:17:55) I mean, it's a good demand signal that there's value, but how does that change your business? (00:18:00) Yeah, so I think one thing that's interesting that we're seeing is obviously there's been the shift from training to inference, you know, over the last few years. (00:18:08) That split continues to grow on the inference side as usable and ROI positive applications get developed. (00:18:15) I think the two things I see on the inference side now is (00:18:21) Inference is a lot more complex than I think initially thought. (00:18:25) And what I mean by that is it's not as simple as you train a model and then it's easy to inference it. (00:18:32) In certain cases, you can do that on similar infrastructure, but there are issues around latency, fungibility of that, and really optimizing the cost of your compute on the inference side. (00:18:45) How do you manage (00:18:46) peaks of inference demand, and obviously it's not linear like training, your GPUs are on all the time, 100% of the time. (00:18:54) And so with inference, you have a lot more variability. (00:18:57) And so there's a lot more nuances in optimizing inference. (00:19:02) I think the second thing that's observed that I've seen is inference is definitely a memory problem, a memory throughput problem. (00:19:10) on the inference side, you have these kind of phases called pre-fill and decode, right? (00:19:16) And how you optimize that across a fleet of GPUs is actually a unique technical problem. (00:19:21) And then the third is what I would say is distribution. (00:19:25) You know, a lot of times training infrastructure is quite centralized. (00:19:29) What you're seeing with inference is in many use cases, as this becomes more ubiquitous, you're going to have more and more decentralized (00:19:38) inference clusters. (00:19:39) And actually one of my favorite companies is one of your companies, Base10, which is really optimizing distributed inference at scale. (00:19:47) And I think one thing that's interesting when you look at companies like that and other inference clouds is how do you optimize the compute and build out these clusters (00:19:58) that could actually look very different than a training cluster, where training cluster might be 50, 100, 150 megawatts in one kind of four walls. (00:20:07) I think you're starting to see distributed inference, which could be, you know, four or five megawatts and five separate data centers and stitching them together in different areas, right? (00:20:17) And that looks very different from a kind of power perspective, how you, know, the software matters a lot more when you're doing like distributed inference. (00:20:26) And then in terms of your question, how it impacts us, I think one of the things that we've been focused on is, where we started this conversation with you on financing compute, that was really obviously, it started with mostly training. (00:20:42) A lot of those hyperscalers are now doing a lot of inference on that same infrastructure, but these are investment grade counterparties. (00:20:49) You know, it's easy to, it's easier to lend money to build out these clusters to those customers. (00:20:55) I think now that you have this new crop of inference clouds and application layer companies that are needing tons of inference, I think the key question that we're really focused on is how can we finance the next build, which is distributed inference. (00:21:11) And maybe the last one or two takeaways would be (00:21:14) One thing I'm seeing is, for every application layer company out there, the highest line item from COGS is compute. (00:21:22) And then the inference companies and inference clouds out there, most of them are purchasing up compute from either other clouds or unused capacity. (00:21:34) And when you look at margins for that, you've got like layered margins. (00:21:38) And so there's a push to kind of own your own infrastructure to really drive and increase profit margins, but also it's the ability to kind of have control of your own destiny. (00:21:49) And I think a lot of folks are starting to, the application layer companies and inference clouds are grappling with how can we build and own and operate our own infrastructure. (00:21:57) And that's something I'm really looking into. (00:22:00) I am too. (00:22:00) And I think one of the things that (00:22:02) is going to make a big difference in this ecosystem is like, can the inference clouds like base 10, can they deliver reliability that you would expect from a cloud, like a traditional cloud? (00:22:17) Because the like, (00:22:20) distributed data center operations that they consume today do not offer that reliability. (00:22:26) Right. (00:22:26) And the other thing that's interesting is, this is additional reporting from last week. (00:22:30) If you're familiar with Silicon Data, they put together a lot of data on spot pricing and price per token performance. (00:22:38) This is Carmen Lee's company. (00:22:40) And one thing that I think was really interesting in an article she published last week had to do with (00:22:47) how two pieces of compute that look identical on paper have wildly different performances, everything from reliability to cost to speed. (00:22:56) And I think as you distribute, you know, have distributed inference, how do you mash together very different types of compute and try to optimize for reliability, I think is super interesting. (00:23:09) And that gets to kind of one thing I find really interesting that NVIDIA is doing is this concept of AI factories and building AI factories, behind corporates and AI companies. (00:23:22) And maybe the way I unpack that is you've got kind of more large monolithic cloud players, the hyperscalers and the neo clouds that are building large scale, you know, cloud environments and a lot of (00:23:35) where I think Nvidia and others see this going is, yes, those are going to be important components and those are going to be huge markets, but corporates, Fortune 500 AI companies that use a ton of compute will want dedicated AI factories associated with workloads that they run and that they have control over. (00:23:54) And so I think you're starting to see the early indications of how do you find