# AI Startups Locked Out of Compute in 6 Months | Evan Conrad

Markdown copy of https://realalpha.app/v/I2UR4asz-RE/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

Evan Conrad says AI startups are being priced out of compute as lenders demand creditworthy tenants or huge upfront payments.

- Show: MTS
- Hosts: [Theo Jaffee](https://realalpha.app/people/theo-jaffee/) & [Sophia Dew](https://realalpha.app/people/sophia-dew/)
- Guest: [Evan Conrad](https://realalpha.app/builders/evan-conrad/) is co-founder and CEO of SF Compute.
- Published: Oct 28, 2026
- Length: 22 min

## The best minutes

- [0:49](https://realalpha.app/v/I2UR4asz-RE/?t=49) [Compute prepayments are pricing startups out of clusters](https://realalpha.app/s/compute-prepayments-are-pricing-startups-out-of-clusters/) (5 min)
- [6:39](https://realalpha.app/v/I2UR4asz-RE/?t=399) [GPU clouds face correlated customer and financing risk](https://realalpha.app/s/gpu-clouds-face-correlated-customer-and-financing-risk/) (3 min)
- [16:04](https://realalpha.app/v/I2UR4asz-RE/?t=964) [SF Compute aims to make GPU rights usable as collateral](https://realalpha.app/s/sf-compute-aims-to-make-gpu-rights-usable-as-collateral/) (4 min)

## Also worth watching

- [10:20](https://realalpha.app/v/I2UR4asz-RE/?t=620) [Compute access may require startups to raise far more capital](https://realalpha.app/s/compute-access-may-require-startups-to-raise-far-more/) (2 min)
- [13:19](https://realalpha.app/v/I2UR4asz-RE/?t=799) [A small GPU cluster may be hard to resell after its tenant fails](https://realalpha.app/s/a-small-gpu-cluster-may-be-hard-to-resell-after-its-tenant/) (3 min)

Skip the roughly two-minute sponsor roll after the interview and the brief Lovable sponsor read.

## Summary

Evan Conrad is co-founder and CEO of SF Compute. The company operates GPU clusters without owning the hardware and is developing a market for GPU-use rights that lenders could accept as collateral.

Conrad argues that AI startups already face growing barriers to buying compute, as lenders demand investment-grade credit or prepayments of 30% to 100% for long-term GPU contracts. He says uncertain resale values and reliance on cloud operators create correlated risk, which can steer capital toward the largest firms and leave smaller startups behind. Conrad advises startups to raise as much as possible or partner with a creditworthy company, while SF Compute is trying to make GPU-use rights usable as collateral. He says that could reduce upfront payments, while simply building more compute may not solve the financing problem.

### Minute by minute

- [0:49](https://realalpha.app/v/I2UR4asz-RE/?t=49) **Compute access depends on credit.** Conrad says many startups already struggle to buy compute because lenders want investment-grade or near-investment-grade borrowers. Since lenders are unsure what GPU clusters will be worth later, they may require 30% to 100% prepayment on a three-to-five-year contract, a burden startups may not be able to fund.
- [3:58](https://realalpha.app/v/I2UR4asz-RE/?t=238) **Creditworthiness could split the market.** Conrad says companies that get creditworthy enough to secure compute can use it to generate returns and become more creditworthy. He expects this cycle to favor large players and consolidate the industry unless financing changes.
- [5:43](https://realalpha.app/v/I2UR4asz-RE/?t=343) **New Blackwell clusters are pre-sold.** Conrad says some Hopper and Ampere capacity remains available, but Blackwells are sold out. Clouds often built Hopper clusters before finding buyers; for Blackwells, customers generally signed contracts first, leaving most new clusters spoken for.
- [6:39](https://realalpha.app/v/I2UR4asz-RE/?t=399) **Lenders fear correlated downside.** Conrad says lenders may worry that GPU clouds will struggle to find future customers as investment-grade labs and hyperscalers build or operate compute themselves. In a downside scenario, clouds could be left with less creditworthy AI-lab customers that fail, though Conrad says he is not predicting that clouds will suddenly blow up.
- [10:59](https://realalpha.app/v/I2UR4asz-RE/?t=659) **Startups need capital or partners.** Conrad advises a Series A startup to raise as much as possible, saying the bar could be around $1 billion, or to partner with an investment-grade company that can co-sign a lease. He says investors could also buy compute, and warns that venture portfolios may consolidate or be wiped out if the debt problem remains.
- [12:23](https://realalpha.app/v/I2UR4asz-RE/?t=743) **More compute alone will not solve it.** Conrad says building more compute could make the financing problem worse. He thinks lenders need better collateral and more confidence that clusters retain value; SF Compute aims to provide backup operations so lenders can rely on rights to GPUs rather than only on the hardware and its operator.
- [14:33](https://realalpha.app/v/I2UR4asz-RE/?t=873) **Small clusters risk being stranded.** Conrad says new clouds may build five-megawatt clusters and sell them to tenants that later fail. Large investment-grade customers such as Microsoft, Google and Anthropic may not want clusters that small, leaving the equipment difficult to resell.
- [16:04](https://realalpha.app/v/I2UR4asz-RE/?t=964) **SF Compute makes GPU rights financeable.** Conrad says SF Compute operates clusters it does not own and is building a market for time-bound rights to use GPUs, which lenders could use as collateral. Customers set the price they are willing to pay, while funds or data-center owners can own the assets and SF Compute provides cluster operations.

## Transcript

MTS, hosted by Theo Jaffee and Sophia Dew, with Evan Conrad.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[0:00](https://realalpha.app/v/I2UR4asz-RE/?t=0) Now it's not like the first month up front. It's like 100% prepayment, somewhere between 30 and 100%. We've seen 100%, meaning that as time goes on, you are asked to pay for a three to five year contract on a supercomputer with all the cash up front or a large portion of the cash up front. And that's a death nail for startups. That's one in which startups don't get to get compute because they will go back to their venture capitalists and they'll ask for more money and they won't necessarily get that more money. And we are back. We are live with Evan Conrad, who is the co-founder and CEO of SF Compute.

[0:39](https://realalpha.app/v/I2UR4asz-RE/?t=39) He just wrote this really interesting and viral article called The End of Startups, The Coming Credit Crunch, which was, you know, the sentiment behind this article was also posted about by Elon and many other people. And it starts with this very provocative claim. Within the next six months, most AI startups will not be able to buy Compute. so interesting evan welcome to the studio thanks thanks for having me absolutely so why do you believe this so this article was more of a prediction of a trend that is already happening not necessarily a prediction of oh and i probably should have said um currently most startups cannot buy compute um it is only simply getting harder and harder and harder um it's like a progressively increasing thing. The problem is that if you want to finance a GPU cluster, you need to have some way to finance that. You need to first have collateral for your GPU cluster, which is typically the hardware itself. And the problem of that collateral, though, is nobody wants to buy the hardware. They want to buy the hardware in a data center that's operated by a company that works and sells you the thing. And the situation in which you recoup that hardware is often the situation in which the person who's operating that hardware is just as

[1:59](https://realalpha.app/v/I2UR4asz-RE/?t=119) likely to blow up. And not only are they likely to blow up, so is all the people who can operate GPU clusters. There's a lot of correlated risk in that collateral. And meanwhile, the lender doesn't know what the cluster is worth in the future. And this is regardless of like what you think about AI or what I think about AI. Like I obviously think that the whole thing is going to go off to the right, but a lender may not think this. And so in order to get a cluster online, you need to have a credit-worthy counterparty who's buying the cluster. And so if you are the startup that's trying to buy a cluster, you don't have any credit. To the eyes of a lender, you look like you're going to blow up. And so many new clusters, both sort of the whole time, but now increasingly more and more so, are being asked for investment-grade credit or near-investment-grade credit, which startups don't have. So to make up for that fact, you're instead asked for a big prepayment. So initially, this might have been a small prepayment. You can imagine you're going to get a lease for a new rental unit, and they ask you to pay the first month up front or something like this the same sort of situation happens in gpu clusters except now it's not like

[3:20](https://realalpha.app/v/I2UR4asz-RE/?t=200) the first month up front it's like 100 prepayment it's or large amounts you know somewhere between 30 and 100 we've seen 100 meaning that as time goes on you are asked to pay for a three to five year contract on a supercomputer with all the cash up front or a large portion of the cash up front and that's a death nail for startups that's one in which startups don't get to get compute because they will go back to their venture capitalists and they'll ask for more money and they won't necessarily get that more money so will the will the top echelon of startups basically be fine because they can just raise more from vcs to make these up from payments uh so maybe yeah i i think there's like some cutoff where if you can get to some credit worthiness status then the returns to the credit worthiness um kind of continue to spiral around you because you got compute and so therefore you can make the returns from those compute which makes you look more credit worthy and so what i expect to see assuming we don't change anything is you get consolidation in the industry now we as sf compute are working to try and prevent that this is what sfc does um but it takes a very long time to do that um and in the

[4:34](https://realalpha.app/v/I2UR4asz-RE/?t=274) short term what looks to be the case is you've got um the sort of big players off raising multi billion dollar rounds and you've got the um smaller players kind of not knowing this is happening which how could they possibly not know this is happening i mean don't they have to deal with debt markets and whatnot all the time? So most AI labs don't know how the thing they are buying works. From their perspective, they think, oh, I'm being asked for a really big prepayment by all of these different clouds. It must be because I don't have leverage or I'm not a good negotiator. And what they don't know is that, no, the cluster you're buying does not exist. All the clusters that have been built are sold.

[5:21](https://realalpha.app/v/I2UR4asz-RE/?t=321) So when you're buying a cluster, you're buying a new cluster and that cloud has to go get debt for that new cluster or they already have debt and it's got covenants on it or so on. And so most customers, I think, most like AI labs, don't seem to necessarily connect the I have to do a big prepayment to the debt markets are freezing up. but it's like one-to-one that's why it's happening what do you mean by the world's compute is sold out like is there no way if you're a startup to get like old like hopper or like ampere NVIDIA chips if you want them you can you have to like do you have to like buy 5090s on ebay or something like um you can still get some hopper and you can still get some ampere um but if you want black wells yeah they're sold out um in part because the black wells weren't made until you ordered them And so the only one, nobody, very few people speculatively built Blackwells.

[6:18](https://realalpha.app/v/I2UR4asz-RE/?t=378) Lots of people speculatively built hoppers. Built or bought? Built. So, for example, in hoppers, what happened is the cloud builds a cluster. Then the cloud sells the cluster. That's not how it worked for Blackwells. Blackwells, the customer signed a contract. Then the cloud built the cluster. And so every Blackwell cluster, for the most part, is spoken for. So if you want a new cluster, it's a new cluster that you're going to get. Interesting. So how bad is the correlated risk here? So it seems like you're going to have a lot of investing-grade tenants that take on leverage in order to be able to buy a ton of GPUs, and then they'll get downgraded because of that.

[7:06](https://realalpha.app/v/I2UR4asz-RE/?t=426) And this seems like a chain of risk. Do you think that this could just blow up enormously, 2008 financial crisis style? Especially because much of the American stock market right now is propped up by the AI boom. So I want to be clear. I don't think this is happening. I don't think that the clouds are going to blow up suddenly or AI is doomed or whatever. But it doesn't matter what I think. It matters what a lender thinks. And as you push the debt markets to the limit, you're asking them to deploy larger and larger amounts of capital. And that means that they will be more and more freaked out about how that capital is being used.

[7:46](https://realalpha.app/v/I2UR4asz-RE/?t=466) And so what matters is like, what is the actual scenario that occurs in the downside? And that scenario does indeed look like correlated risk. I can describe one that is if I were a lender, here's what I would. worried about. You have a GPU cloud. That GPU cloud makes money right now, but may not make money in the future, because to make money in the future, they have to deploy a new cluster.

[8:12](https://realalpha.app/v/I2UR4asz-RE/?t=492) **Evan Conrad:** However, their biggest customers are ones who themselves want to be the GPU cloud. The Frontier Labs, the Hyperscalers, the people who have investment-grade credit want to do the thing you're doing, and will simply try to disintermediate you if they have the opportunity to.

[8:32](https://realalpha.app/v/I2UR4asz-RE/?t=512) So if you were investing or if you were writing debt against the cloud based on their cash flows and you were projecting that they would be able to sign another cluster in the future, that may not necessarily be true, except with tenants who are not investment grade. because, you know, the Series C startup from San Francisco that is totally printing money and would totally buy a big cluster is still not going to necessarily try and become the cloud themselves. And so you have the opportunity to sell them, but the bank may not think that's great. And so from the bank's perspective, they may be looking at this and saying, wow, this cloud has all these investment-grade customers right now, but in the future, they may not have investment-grade customers.

[9:14](https://realalpha.app/v/I2UR4asz-RE/?t=554) And so I have the ability to predict that in the future, those people might die. And now those non-investment grade customers are new AI labs that are competing with this company, Anthropic, that seems now to be doing amazing. And like, what if Anthropic just wins? And then what happens is the clouds can't sell to Anthropic because Anthropic just wants to do this. Meanwhile, they're stuck selling to these other competitors who might die. And so maybe that's my losing condition where this cloud doesn't have the ability to sell into a market because their market moved away from them. Now, is that true?

[9:47](https://realalpha.app/v/I2UR4asz-RE/?t=587) I don't think that's necessarily true, but this is like the scenario that people are worried about. And through Lovable's MCP server, you can create and deploy your project directly from the agents you already use. Turn ideas into software people love with Lovable. Lovable.dev. Now, on to the episode. Like you said, you know, you don't believe this, but it doesn't matter what you think. It matters what lenders think. So, like, do you think that a scenario like this is possible because, like, lenders will just freak out? So, right now, lenders are freaking out um i hope that the anthropic um s1 like calms a lot of people down because i think a lot of people were basing up things off of just like incorrect information um and so hopefully that changes things um but at least where we stand right now it is hard to get debt which then ripples through the whole economy and it's not good for startups so if you're a startup raising a series a right now like what would you do i would raise as much money as i can um i would raise um i would take your round and i would add 10x to it um and if i was an investor i would buy compute instead um i would yeah basically i would either try to find a big investment grade normal

[11:29](https://realalpha.app/v/I2UR4asz-RE/?t=689) company that i can either partner with or sell to um in some cases um and use their balance sheet to co-sign on the lease with me so I get investment grade credit or I would try to raise the biggest round I possibly can and so I'm above the bar. What do you think the bar is now? Like a billion dollars. Like if you're trying to raise a round

[11:52](https://realalpha.app/v/I2UR4asz-RE/?t=712) **Evan Conrad:** I think you should try and raise like a billion dollars.

[11:54](https://realalpha.app/v/I2UR4asz-RE/?t=714) It depends on your industry obviously. That sounds like, you know, kind of like... Oh, you know, why didn't I think about raising a billion dollar round? I know. That's the problem. This is what I'm saying. This is why the article is called End of Startups, is because I don't think many people can raise a billion dollar round. I think that's extremely difficult to do. And the ones that you see are the ones that are like playing with red dots. It's just survivorship bias. I think that in general, if you don't fix the debt problem, most of the portfolios of various venture firms are going to consolidate or get wiped out or they will convert into basically GPU clouds.

[12:39](https://realalpha.app/v/I2UR4asz-RE/?t=759) So how does this get solved? Just manufacture a lot more compute? That doesn't necessarily solve it. It kind of makes it worse. um so um over time um you can build up the credibility that these things have value um like the the gpu clusters have value and that these companies are making money over a long period of time so i think like in four years this may all be resolved um because um anthropic will be ripping and i don't know maybe we'll have like a cure for cancer or something and so uh we'll all feel much better about it. But until then, you're going to get, like, there's just a lot of risk in the system. And we're being asked to do more risk in the system. So the way that we think it gets solved, and like, obviously, we're biased about this, is we think that you need to make the collateral that the lender is lending against into something that's more valuable, rather than what they are currently lending against. So right now, people are lending against this GPU cluster, but that GPU cluster needs an operator. That operator is just as likely to blow up in the case in which you need them to sell that collateral. And so what we're trying to do is be the place that can make that collateral be something that's worthwhile. To do

[13:56](https://realalpha.app/v/I2UR4asz-RE/?t=836) that, we have the whole operating capacity of a GPU cloud. So we operate GPU clusters, but we don't own the hardware. And because of that, we're not exposed to the same risk. And so we end up as a backup servicer to these folks um i don't know if you know what that does you know what the term servicer means in this case um what does it mean it basically means that like we have the ability to operate the hardware um but we don't have the same exposure risk um that everyone else has right and so because of that um we get to offer um uh the like right to some gpus um as the collateral rather than the hardware itself and if you recoup that right to the gpus we can sell that right and we don't blow up right when you need us to sell it.

[14:38](https://realalpha.app/v/I2UR4asz-RE/?t=878) Right. So is this what Elon meant when he said there's going to be a lot of stranded compute? Yeah, so some of the ways that I think you could have stranded compute is there are a lot of new, what we would call neoclouds. There are a lot of new clouds. Everyone and their mother is trying to suddenly become a cloud because you just make a lot of money doing it. um however um they will get sold to companies that won't be able to grow because um sorry the new clouds are starting with smaller clusters they're starting starting with like a five megawatt deal um and they will sell that to somebody um who or they they will sell that to somebody who like looks like a less investment grade tenant or at least initially that was the plan, then you have to find an investment-grade tenant, which looks like Microsoft or Google or Anthropic, these groups of people.

[15:37](https://realalpha.app/v/I2UR4asz-RE/?t=937) Anthropic's almost investment-grade, I guess, or now is. And these people don't give a shit about your 5-megawatt cluster at all. Like, it's too small for you to care about. So what seems like it may happen is you get that 5-megawatt cluster. You sell it to someone who then blows up.

[15:56](https://realalpha.app/v/I2UR4asz-RE/?t=956) **Evan Conrad:** You don't have the ability to sell that cluster to anybody else because your investment grade customer.

[16:00](https://realalpha.app/v/I2UR4asz-RE/?t=960) are, like, don't care about that small of compute. I think that's where you're going to get started, compute. Interesting. So what is SF Compute doing about this? You said building a physically settled forwards market for compute. Yeah, so we have an enormously, like, complicated challenge ahead of us, but functionally what we're trying to do is make the debt cheaper. Whenever we've tried to explain San Francisco Compute, it's like this elephant, and we, like, show you one view of the elephant. One way of explaining San Francisco Compute is, like, we sell GPUs, man, don't worry about it. and another way of explaining SF Compute is we're like this forwards market but the sort of longer explanation of what we do is we operate GPU clusters, we don't own

[16:42](https://realalpha.app/v/I2UR4asz-RE/?t=1002) **Evan Conrad:** the assets and then we run a market on top of the GPU clusters that we operate and this allows us to create this piece of paper that says I own the rights to these GPUs between this

[16:52](https://realalpha.app/v/I2UR4asz-RE/?t=1012) time and that time and you can lend against that or that's the goal for us. We're going to make that possible to lend against And we think that is a better form of collateral that's more viable than the hardware itself. And if you can do that, then you can bring down the requirements for credit quality that the customer has assigned, which basically means that if you use that as the collateral for the loan, then the customer's experience is they don't have to pay a big prepayment. The prices go down for them, and they get to succeed at all. Why don't lenders already do that? Because this type of collateral doesn't exist.

[17:25](https://realalpha.app/v/I2UR4asz-RE/?t=1045) We're making that collateral. To do that, it took us three and a half years just to build the operating capacity of a GPU cloud. We had to do that while we didn't own the hardware. So we basically had to slow grow for three years. We had to look like the dumbest cloud in the world. Because the moment you buy the hardware, then you're exposed to the same risk. And so we had to not buy the hardware while still building the operating capacity of a GPU cloud. Then we had to build this forwards market, which you can't even build until you build the hardware. Or until you can build the operating capacity of the cloud.

[17:55](https://realalpha.app/v/I2UR4asz-RE/?t=1075) You have to get liquidity on that, and then you can do lending. It was like a multi-stage process that is almost impossible to explain to anybody who is not deep in the industry. Meanwhile, it's like the worst possible path for a company to do, and so just nobody has done it. How do you price forwards on GPUs, especially very long-dated forwards, when GPUs depreciate? You just have a customer who prices it. We don't price it. The customer does. The customer says, I would like to buy this between this time and this time, and I'm willing to pay up to this. Really? Well, who's the counterparty then?

[18:29](https://realalpha.app/v/I2UR4asz-RE/?t=1109) So we operate the clusters, and we're like a property manager for the person who owns the cluster. And so there's sort of like a mixed structure there. The term counterparty here is probably not exactly what I think you mean. You're asking who the customer is, or you're asking who the vendor is? Who the vendor is. The vendor is like a fund. So the people who are building GPU clusters today that are working with us are people who otherwise wouldn't have built a GPU cluster. They're not a cloud. Like, they're sometimes kind of a cloud and will help new clouds get started. But a lot of people are like, I've got a data center, and I could sell the data center to a cloud, or I could become the cloud.

[19:10](https://realalpha.app/v/I2UR4asz-RE/?t=1150) But then they don't have the ability to operate a GPU cluster. And so we step in and we partner with them. And we say, look, we'll do the GPU cluster part. You do the data center part. But you're going to own the assets. and so you'll make more money because you own the assets. We're going to give up the right to do that because we want to do this debt thing over here, which is why I say like SF Compute is kind of like a very stupid GPU cloud. We're sort of like doing the economically dumb thing so that we get the ability to do the smart thing, which ultimately saves startups. Totally. Well, I think that's a good place to wrap.

[19:42](https://realalpha.app/v/I2UR4asz-RE/?t=1182) Evan, thank you so much for coming on MTS. Thank you. A huge thanks to MTS sponsors. 11labs, AI that communicates at human level across every channel and modality. 11labs.io slash mts. Scale your startup on Neon. Millions of developers and startups have already chosen Neon for their back end. Start on the free plan or get up to $100,000 in credits for your startup at neon.com slash MTS. Special thanks to our sponsor Kong, the AI connectivity platform. Connect APIs, LLMs, agents, and systems with serious security and governance. KongHQ.com. Finally prove what your AI investment is worth with Lairdyn.

[20:23](https://realalpha.app/v/I2UR4asz-RE/?t=1223) Head to Lairdyn.com today to start measuring your AI's bottom line. MongoDB, trusted to run critical applications by 75% of the Fortune 100, the intelligent data platform for the era. Apply today for access to up to $15,000 in Atlas credits, dedicated technical support, and exclusive events and go-to-market opportunities. MongoDB.com. Chargebee, billing and monetization infrastructure for the AI era. Chargebee.com. Blitzy, autonomous software development for enterprise code bases ship 5x faster blitzy.com adquick make your brand a billboard out of home advertising as easy to scale as digital adquick.com arena measuring ai performance in the real world arena.ai support for the show comes from vcx the public ticker for private tech the u.s stock market started history's greatest wave of wealth creation from factory workers in detroit to farmers in omaha anyone could own a piece of the great American companies. But today, our most innovative companies are staying private longer, which means everyday Americans are missing out. Until now. Introducing VCX, a public ticker for private tech.

[21:35](https://realalpha.app/v/I2UR4asz-RE/?t=1295) Visit getvcx.com for more info. That's getvcx.com. Carefully consider the investment material before investing, including objectives, risk, charges, and expenses. This and other information can be found in the Innovations Fund prospectus at getvcx.com. This is a paid sponsorship. You