MediaTalk | Season 3 | Ep. 17 - How AI Workloads Are Shining a Light on Neoclouds

View Full Transcript

Mike Reynolds: Hi, I'm Mike Reynolds, a senior reporter covering the media industry with S&P Global Market Intelligence's tech, media and telecom news team. Welcome to "MediaTalk," a podcast hosted by S&P Global, wherein the news and research staff explore issues in the ever-evolving media landscape. Today, we're going into the cloud, so to speak. The rapid evolution of artificial intelligence has led to a new breed of cloud service providers known as AI neoclouds. The business is currently attracting considerable investment, even though questions remain about its long-term prospects. My colleague, Iuri Struta, is going to try and help us make some dollars and sense of this here. Iuri is covering tech, M&A, and capital markets for S&P Global Market Intelligence. Iuri, how are you doing today?

Iuri Struta: Hey, Mike. Nice to be here.

Reynolds: Iuri, forgive me. Neoclouds, it sounds very surreal, Salvador Dali-ian, if you will. What are we talking about here? How long have neoclouds been a thing?

Struta: Yeah, I think that's a great question to start with. Most of the neoclouds were previously crypto miners. Following the crypto bust and the emergence of AI, they pivoted to powering AI workloads. To get deeper into this, basically, to mine cryptocurrencies like Bitcoin and Ethereum, you would build these data centers and fill them up with what is most of the time NVIDIA GPUs, graphics processing units, or some other sorts of accelerated computing chips. So, you find new bitcoins and then sell them in the open market — Bitcoins, Ethereums, whatever cryptocurrencies you want to mine.

And this works as long as the price of crypto keeps going higher and higher. But I think in 2022, there was a crypto bust, and a lot of these companies started to struggle. They were looking to find new ways for their businesses, and they were lucky. Then ChatGPT emerged, GenAI took off around the end of 2022, and they basically found an even better use case—they could rent GPUs for AI workloads. This could be used either for training or inference. So most of them pivoted. And if you look at CoreWeave, for example, they recently had their IPO, and CoreWeave was born as a crypto miner. They were originally known as Atlantic Crypto. So we did some work around this, and we found about 54 of these neoclouds, most of them based in the US, but some of them in Europe as well.

Reynolds: Alright, in layman's terms, Iuri, what exactly do they do or do best?

Struta: Yeah, so I think they just provide AI cloud infrastructure for their clients. So if you are a customer that needs to train or run some AI workloads, you would go to one of them and start renting GPUs by the hour. That is usually much cheaper than going and setting up your own servers. And in many cases, for small businesses, for example, it might be even very hard to get access to these chips. Even if you have the money, you can't purchase them because you don't have the relationships with the chip makers.

Reynolds: I gotcha. So neocloud providers aside, AI workloads will also be handled by hyperscalers like Amazon.com Inc.'s Amazon Web Services or Microsoft Corp.'s Azure. But are the neoclouds better at this type of thing?

Struta: Oh, it is hard to say if they're better than the hyperscalers. This degree of specialization definitely helps. For certain workloads, a survey by 451 Research, which is a tech research firm part of S&P Global Market Intelligence, shows that companies might choose a different provider for their AI workloads than their cloud provider for general-purpose computing. So, for example, if my cloud provider for regular workloads is Google LLC or Amazon, I might be inclined to use Google for my AI workloads as well. But 451 surveys show that this is not the case all the time. It very much depends on the needs of each client, and a lot of the clients may be shopping around for better prices or better tools. And sometimes they could find this neocloud is the best solution for AI.

Reynolds: So providing that optionality, that's helping the marketplace, which is growing. Can you give us a sense of investments in neoclouds at this juncture?

Struta: Yeah, so there's a lot of money being invested into these neoclouds. You have venture capital firms, you have private equity firms, and you also have the corporates, all of them either investing in equity or in debt. In 2024, you had more than $10 billion invested into neoclouds. This is across these 54 companies that we identified and across more than 30 transactions. And this comes from almost nothing in 2020. In 2020, you didn't have basically this category. This category was crypto miners that no serious investor was looking into or no institutional investor was looking into. And this year also seems to be off to a good start, with about one billion raised by neoclouds so far. And I think if you look at CoreWeave, which is the leader, the detached leader of the neocloud cohort, their revenues grew from $230 million in '23 to $1.9 billion in '24, and other neoclouds like Crusoe Energy Systems LLC and Vultr have also seen strong headcount growth in the past two years or so. This basically indicates they are experiencing strong revenue growth as well. We don't have this private company, so we don't have their financials right, but looking at the headcount, we can see that headcount is growing. So it's definitely the size of a business that is growing very strong.

Reynolds: That underscores the health, and I think you talked about it a little bit, that CoreWeave has got a good relationship with NVIDIA, and that's helping matters tremendously.

Struta: Yeah.

Reynolds: Yeah.

Struta: That's one of the most interesting things about this neocloud. If you look at CoreWeave, its strong growth has been possible because of its, like you suggested, cozy relationship with NVIDIA. NVIDIA sells GPUs to CoreWeave. They also invested in the equity of CoreWeave, and they are a large shareholder. They also rent back compute from CoreWeave for its own AI workloads, so it's also a customer. So if you're looking at NVIDIA, NVIDIA is CoreWeave's supplier, investor, and customer. So they're participating across the entire business functions of CoreWeave. And NVIDIA, this suggests to you that NVIDIA, a chipmaker, is invested in CoreWeave’s success. And if you're stepping back and looking at the game that NVIDIA is playing here, NVIDIA needs these neoclouds to be successful. They sell most of their chips, a big chunk of their AI chips, to companies, to these hyperscalers like Microsoft, Amazon, Google, and Meta Platforms Inc. But increasingly, many of these companies started to build their own silicon, their own AI chips. So NVIDIA, obviously, encourages the creation of these neoclouds. The idea here is if neoclouds build a sustainable business, it'll be a good source of revenues for NVIDIA, and it's diversified revenues. We looked at NVIDIA's investments in the space. They invested in five neoclouds. Some of these neoclouds only provide NVIDIA GPUs. Some of them also provide Advanced Micro Devices Inc. GPUs. Speaking of AMD, they're a competitor of NVIDIA, and they also make GPUs for AI workloads. They're following the same playbook. They invested in two neoclouds. This is what we know is public information. Maybe they invested in more, and this hasn't been disclosed. So you have these two chipmakers trying to encourage and build companies that are separate from them. But they are invested very much in these companies being successful.

Reynolds: NVIDIA always making news. It looks like they're committing to working with manufacturing partners to design and build factories for producing AI supercomputers entirely in the US following the Trump administration's initial announcement that tariffs would be imposed on semiconductor imports. That's going to help CoreWeave and others in their neocloud efforts.

Struta: I think this definitely increases the availability of chips for many of these neoclouds and also potentially shields them from geopolitical factors. You don't rely on sourcing your chips from abroad or Taiwan, Taiwan Semiconductor Manufacturing Co. Ltd. in this case, or at least not all of them. You have some homegrown sources. I don't think this is going to come very soon. Probably best-case scenario, I think NVIDIA said in about one year, one and a half years, they were going to have the factory ready. But still, I think it's definitely positive for them.

Reynolds: Yeah, again, it takes time to get there. So we'll see. 18 months seems long, but maybe not so far away in the grander scheme of things. Alright, absolutely. Neoclouds making strides. But is there a long-term business potential here, Iuri? I think CoreWeave’s loss widened in 2024 from '23. More of, I think it's really reliant on a couple of key customers. And maybe if one of them backed out, it would be detrimental to its prospects.

Struta: Yeah, that's a great question, and I don't think anybody has a definite answer yet. But if you look at CoreWeave, which is among the few that have detailed financials public and maybe another company's Nebu Holding, this is a company based in Europe. At its core, the business seems very simple. You buy GPU servers mostly by selling debt and equity. And you then rent them out to customers, hoping that the total revenue you get from each GPU is going to be substantially higher over time than the cost of the GPU, its installation, and other services. In the meantime, you'll also be able to use that revenue you are getting from the GPUs to comfortably service your debt. This is a very capital-intensive business, so you need to spend a lot of money upfront to build the infrastructure. If you are looking to grow quickly like CoreWeave does, then inevitably you are going to have huge losses in the first few years or maybe even more because you need to finance that capital-intensive growth until the market becomes more established and you have the leaders of this new neocloud cohort shaping up. So you finance that through the issuance of equity and debt, but this debt-heavy business model could get the company into trouble. If demand for AI slows down, so if revenue declines and the company might have trouble servicing the debt, then the business might not be very sustainable. On the other side of the coin, of course, if demand for AI continues to strengthen, then you potentially have a big winner here. If you look at traditional cloud and its growth trajectory, there was a lot of skepticism about the business model for traditional cloud at the beginning, but in the end, it proved to be a massive winner. And that keeps on giving for companies like Amazon, Microsoft, and Google. The biggest thing that these neoclouds need to do is to avoid what we call the commoditization of the business because, in essence, the business is a commodity. You don't need a lot of knowledge to buy GPUs and then rent them out. How you decommoditize the business—a lot of companies try to add additional services on top. This is how traditional cloud has grown, Amazon, Microsoft, and Google, and there are signs that CoreWeave is trying to do the same. They recently acquired Weights and Biases Inc. This is a company that helps developers build their AI products. So they're saying to customers, look, if you're coming to us, we don't only have GPUs, we don't provide only access to GPUs, but we're also giving you the tools to build your AI. And I think this is probably where the key lies between a commodity business model and one that adds value.

Reynolds: You want to diversify your options there. You want to be able to be more things to more people as it goes. So that makes a lot of sense. Are there any other headwinds, though, at this juncture?

Struta: Yeah, I think some of our analysts that look into these companies are thinking more macro issues. The chips may become more expensive, especially now with tariffs, and a potentially bigger headwind is probably lack of power, especially as a lot of these companies are looking to expand, and yeah, this is a wider problem—not enough availability of power.

Reynolds: Yeah. You mentioned tariffs there, and again, if we have this conversation in a couple of weeks, maybe it's a little different story, but as things stand now, what's your sensibility for tariffs as a factor in the neocloud world?

Struta: Yeah, they are a factor. Maybe not direct, but in as long as it makes the silicon that they're putting into their servers more expensive. So it's not a direct impact, as most of these companies have their data centers locally. But another impact could probably be on demand, which could be more damaging. Look, if you are a tech organization and you're seeing tariff uncertainty, do you approve budget increases for some new AI workloads in the cloud, or are you waiting for more clarity? So this is, I think, where the uncertainty has the potential to hit demand and, in turn, neoclouds as well.

Reynolds: Sure. Neoclouds are not alone in looking for clarity, so we'll see how that maps out in the weeks and months ahead. Your sensibility for the market here is, are there more IPOs on the horizon here?

Struta: So we were actually expecting a pickup in the number of tech IPOs this year as broadly the tech sector. And to a certain extent, it seems this will be a better year than the last two or three, but it's worth noting that the IPO window was essentially shut for tech IPOs in '23 and '24, and the number of IPOs during those years was well below pre-pandemic levels and had trended down. Generally, the IPO market right now is in pretty bad shape. And this is just to provide some background on where we stand here. I think now the outlook for IPOs is much muddier than it was at the end of last year. Tariffs have a big impact. It creates volatility in the market, and the volatile market is not conducive to an IPO recovery. We are already seeing some of the companies delaying their IPOs. Klarna Group PLC is a FinTech buy-now-pay-later, for instance. They delayed their IPO.

Reynolds: I gotcha. On the other hand, are investors lining up some exit strategies here?

Struta: Yeah. So in the tech space generally, but also within the neoclouds, with the IPO window not strongly open, you can look forward to M&A, and M&A has been one of the key and, I would say, only exit options that have been open to investors for about three years now.

Reynolds: Your crystal ball, Iuri, what does it say as you look skyward? What kind of business will neoclouds be at the end of this year? And as we look forward into '26 and '27?

Struta: Yeah, I think it very much depends on demand for GenAI technologies. If that demand will continue to increase, then you will probably have a lot of these companies doing well. You're going to have some consolidation at a certain point. I don't think having 54 neoclouds is sustainable. So you're going to have a lot of consolidation, maybe not in '25, but maybe later. Also, looking at demand growth in technology is not usually a straight line. Usually, after the initial hype, growth can stagnate for a while. This could happen now with what we see with tariffs, and after maybe tariffs uncertainty, this demand could go up again for GenAI. So it's very important for many of these companies to be prepared for this kind of drawback, which could be short-term. I'd say if a bout of weakening demand gets you into trouble, then the business model is not very sustainable. This is where you actually can see more consolidation in the space when companies get into trouble.

Reynolds: So growth doesn't follow a straight line, even if you're going up into the stratosphere and neoclouds. We'll have to come back to this maybe later this year, early next year. That concludes this episode of "MediaTalk." I wanted to thank my colleague, Iuri Struta, for discussing this complex and growing world of neoclouds. Very interesting stuff, Iuri. Thanks.

Struta: Thanks, Mike. It has been great.

Reynolds: Alright, and this is Mike Reynolds. I also want to thank all of you for listening. We'll catch up soon on the next installment of “MediaTalk.”