Compute prepayments are pricing startups out of clusters

Transcript

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

1. 0:56

   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

2. 1:53

   that hardware is often the situation in which the person who's operating that hardware is just as 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

3. 2:49

   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 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

4. 3:47

   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

5. 4:34

   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.

6. 5:10

   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. 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.

7. 5:45

   but it's like one-to-one that's why it's happening what do you mean by the world's compute is sold