AI transcript
0:00:06 You know how everyone’s obsessed with the new AI models, agents, and mind-blowing apps dropping
0:00:10 every week? Well, today we’re going deeper beneath all of that, underneath the hype,
0:00:15 the headlines, and the glossy demos, to the actual metal and concrete foundation that makes
0:00:20 it all possible. We’re talking about compute infrastructure, the biggest physical buildout
0:00:25 in modern history. So this is a really complicated topic, but I think it’s important to understand.
0:00:31 So today I brought on the smartest person I know in data centers, Evan Conrad, the CEO of SF Compute.
0:00:35 He’s building something absolutely genius. They’re creating a spot market for compute,
0:00:40 basically turning supercomputers into a tradable commodity like oil or soybeans,
0:00:43 which makes it possible to finance this infrastructure in a more rational,
0:00:48 sustainable way. Understanding this layer of AI is like understanding the foundation
0:00:53 of the internet in 1999. So let’s dive in with Evan Conrad from SF Compute.
0:01:05 Being a know-it-all used to be considered a bad thing, but in business, it’s everything. Because
0:01:13 right now, most businesses only use 20% of their data, unless you have HubSpot, where data that’s
0:01:18 buried in emails, call logs, and meeting notes become insights that help you grow your business.
0:01:23 Because when you know more, you grow more. Visit HubSpot.com to learn more.
0:01:29 Evan, thanks for coming on here today.
0:01:30 Thank you. Thanks for having me.
0:01:35 Yeah. You know, so I’ve been seeing all your posts on X about like how fast your company SF Compute is
0:01:40 growing. 3X, 5X, 7X, it sounds like SF Compute’s been taking off. And then in the meantime, I’ve been
0:01:44 seeing all this news about compute like everywhere I look, you know, like Sam Altman’s planning on
0:01:50 creating factories. They’re going to build one gigawatt of compute per week, which for people who don’t
0:01:55 know, that is a gigantic amount of compute. Elon Musk is saying he’s racing to 100 gigawatts
0:02:00 of compute in the next few years, which is just a crazy number. And you know, I’m just wondering,
0:02:04 like, are we in a compute bubble? Or where are things at?
0:02:09 I don’t think we’re in a bubble quite yet. But if there is a bubble, I can kind of explain where
0:02:14 it probably will be. Sure, that’d be great. So what I want to do is I want to walk through how you got
0:02:19 from, say, ChatGPT or an application that you might use, all the way down to the physical layer of the
0:02:23 cluster, and how the financing of that entire structure works. Because if you’re trying to find
0:02:28 the economic bubble, you want to find like who’s paying who what money and where and where the
0:02:33 miscalculation is coming from. So in most cases, if you are building up a cluster, you’re looking for an
0:02:37 offtake agreement. That means a long term contract for your cluster. So that way, you could get it
0:02:41 financed in the first place. So you’re going to buy a bunch of GPUs, you’re going to put them in a
0:02:46 data center somewhere. And then you need someone hopefully before you’ve actually bought the GPUs,
0:02:50 who’s agreed to purchase them for some period of time, like a rental agreement. And you take that
0:02:55 contract, and you go back to a lender. And now the lender says, okay, because they’ve already agreed
0:02:59 to pay for it, really, what I’m evaluating is not the broad or market or, you know, whether or not
0:03:03 people want GPUs. I’m really just evaluating, does this customer want
0:03:06 So it’s kind of like almost like commercial real estate or like a mall or something like that,
0:03:09 where you’re getting like the anchor tenants, and they’re lined up and then it helps get the
0:03:14 financing. Correct. If you have the anchor tenant, that person is the person who’s doing offtake.
0:03:19 That’s the person who you’re giving the cluster to give the customer already. And so the lender is
0:03:24 financing against the credit risk of the person who has the offtake agreement. You’ve seen this in a
0:03:28 number of different deals before. So for example, like the Microsoft core read deal, the nice part
0:03:32 about that is no one thinks Microsoft is going to default on their debts. That would be kind of like
0:03:37 a crazy thing to think. And so because of that, you can finance it very cheaply, your cost capital is
0:03:42 really low. And so that’s great. That’s super good. The problem though, is a Microsoft or somebody else
0:03:48 who’s getting the offtake agreement, then sells it to somebody else who doesn’t have good credit risk.
0:03:53 is that what they’re doing? Um, not always. Okay. So when if Microsoft is selling it to OpenAI,
0:03:58 you’re betting that OpenAI is going to continue to grow and OpenAI has very meaningful revenue. I think
0:04:04 it’s a quite reasonable case that OpenAI will continue to grow, that also they can back up what
0:04:09 they’re doing. There’s some amount of bet. There’s like a significant bet, especially in Stargate,
0:04:12 when you’re training new models, and you’re betting the new model is going to be better than the old
0:04:16 model, and you’re going to have some economic return from it. But historically, the previous models
0:04:22 worked. They trained GPT-4 and then GPT-4 made a whole bunch of money. And so GPT-5 is making a bunch
0:04:28 of money. And so presumably you will continue to have this grow, but there is the whole rest of the
0:04:33 market as well. So not everything is OpenAI, not everything is Anthropic. There is also all the
0:04:39 various application layer companies that are popping up. Basically every YC startup and every company that is
0:04:45 currently getting funded, all of those people are signing much shorter term contracts because they can’t
0:04:49 afford a longer way. And why do most of those YC companies need compute? Would they not be able to
0:04:54 just use OpenAI as API or Clot? Or are they having to do more like custom stuff on top or something? I
0:04:59 don’t really fully understand that. Sure. So in some cases, they’re using like inference layer services.
0:05:03 So this is like an open source model that they’re getting from a Fireworks or Together. There’s like
0:05:07 a hundred different services here. But eventually you’re using an open source model, or you might just be
0:05:10 using OpenAI and Anthropic, and then you’re signing short-term contracts with them.
0:05:16 The problem is that if you are either OpenAI or Anthropic, or you’re one of the other inference
0:05:20 platforms that are serving open source models, or you’re doing something like an image model,
0:05:23 or you’re doing something like a video model. So things that are kind of have a lot of other
0:05:28 competitors that aren’t OpenAI and Anthropic. If you’re signing short-term contracts with those
0:05:33 folks, a short-term contract might mean literally just a subscription. You pay a usage-based billing,
0:05:36 and then you pay at the end of the month. And then if you use less the next month, you use less the next
0:05:41 month. And so you’re just paying on a monthly basis. You’re not locked into anything. Now, of course,
0:05:45 these folks will want you to sign a long-term contract, but a lot of you aren’t on that long-term
0:05:51 agreement with their underlying inference provider. So when you do that though, now you’re betting on the
0:05:58 credit risk of a bunch of new startups who currently have a reasonable credit history because they have
0:06:04 just raised at very large valuations. And they raised at large valuations because they initially
0:06:09 made a lot of revenue very quickly because customers are actually willing to pay lots of
0:06:13 money for the AI applications, for all the different services that are happening. Their margins aren’t
0:06:20 great, but the volume is really big. And so to some extent that causes problems because the venture
0:06:26 capitalists look at the low margin, but high volume application layer products. So this is like your
0:06:32 code editor tools. Yeah. These sorts of things. And to some extent, some of the built-in valuation
0:06:38 is betting that the margins will eventually improve. And in some cases that might be true,
0:06:44 but in a lot of cases it might just literally not be true. And what that means is that if your ability
0:06:50 to sign shorter contracts or longer contracts, your credit risk basically is based on your ability to
0:06:56 raise venture capital, but that venture capital wasn’t correct about the ability for these companies
0:07:02 to eventually become higher margin like SAMS, then there is your bubble. Basically your problem then
0:07:09 becomes if a macro event were to occur, for example, if there’s some sort of larger economic concern that
0:07:15 pulls back venture capital, then what would happen is these companies might struggle to raise in the
0:07:19 current funding environment, which would harm their ability to get credit risk, which would make them
0:07:24 need shorter term contracts or their shorter term contracts would fall off because their customers
0:07:29 would fall off. Which means for the inference layers who now are basically signing long-term contracts with
0:07:35 their underlying compute clouds, but selling short-term contracts to their customers. Now they don’t have
0:07:40 the demand coming in to pay their underlying bills that they still have to pay, which then goes back to the debt
0:07:46 providers of the clusters who thought they were betting on something with reasonable credit risk, but the credit
0:07:50 risk wasn’t actually as reasonable as they thought. And so some of these folks could get wiped out.
0:07:58 And because of the total amount of money in that system, because compute is so big and AI is so big,
0:08:02 it can have like broader market effects, which probably then encourages LPs to pull their money out
0:08:08 of the venture capital firms, causing a sort of collapse. That is the fear case. And it’s one of the
0:08:12 reasons why our business exists. So you’re betting on the bubble or just like if the bubble happens,
0:08:17 it’s better for you, but if not, you’re fine? We’re trying to prevent the bubbles. We’re trying to get
0:08:22 to a stable outcome for how AI should work. Our basic premise is that a lot of people are making bad
0:08:28 decisions on the way they’re doing credit risk. We’re trying to make a much more stable state of AI.
0:08:35 So the key problem is basically that the base layer, those GPUs that you’re setting up, those GPUs, many of
0:08:39 them were built on the expectation that they would get a lot higher margins than they actually are.
0:08:44 Because in CPUs, when you’re selling CPUs, you get really high margins. So for example, if you’re
0:08:50 AWS or your GCP and you’re selling web servers, your markup on your CPUs is really high. And that’s
0:08:56 because your customer that you’re selling to is SaaS. Like we’re using a cool product right now,
0:09:01 it’s called Riverside FM. And Riverside FM probably has high margins. And they probably use somebody like
0:09:06 AWS. And because Riverside has high margins, they were able to give AWS high margins, which means
0:09:12 AWS didn’t have to sign as long of a contract in order to de-risk themselves. Because if they’re
0:09:18 making a 60 or 70% markup and maybe their capacity is down 30% one month, that’s not a big deal because
0:09:21 they made so much money on the other months and their markup is so high, there’s a lot of buffer for
0:09:27 them. But in GPUs, you’re not selling to Riverside FM, you’re selling to AI Riverside FM. And every time
0:09:32 you go to AI Riverside FM, you end up spending money on GPUs, which means that you’re super price
0:09:32 sensitive.
0:09:36 And also, I assume that CPUs don’t change as much, right? Like it seems like they just like
0:09:36 slightly get better.
0:09:41 Correct. Yeah. No one has ever really given a s*** like what CPU their web app is running on. Like you
0:09:45 just don’t even think about it. It’s not even like a concern at all. And if somebody comes along and
0:09:47 tries to sell you a new CPU and charge you more money for it.
0:09:51 Well, people used to, I mean, I was born in the 80s, you know, so like when I was a kid, I had friends
0:09:57 who like ran web servers and stuff. And it was all the pitch was like, here’s the CPUs that we have.
0:10:01 Yeah, here’s the price, you know, and that was, that’s how they pitched it was based on the CPU back
0:10:01 in the day.
0:10:03 Yeah, now no one cares.
0:10:03 No one cares.
0:10:09 But in GPUs, people totally care. And in GPUs, your margins are much thinner. And because of that,
0:10:14 they need to sign longer term contracts with their customers, because otherwise, they end up with
0:10:20 this entire like credit risk bubble that filters its way down. Or they just end up on a more localized
0:10:26 scale with basically just a concern of like, oh, I’ve signed for a three year loan. And if I sell
0:10:30 somebody a month at a time, and they drop off, then I still have to pay my three year loan,
0:10:38 and then I go bankrupt. Not great. So the problem, though, is that in order for that GPU cloud to sell
0:10:43 to customers and not go bankrupt, not like have some random collapse, they need to sell you a long term
0:10:47 contract, which is what most people do. Everyone who didn’t do that went bankrupt, basically,
0:10:50 there were a bunch of people who didn’t do that. And they either all pivoted away from doing that,
0:10:56 or they blew up a bit. And so the problem, though, is now you’re just shifting that to the
0:11:01 customer. So whoever’s buying your GPUs now has the same problem, but like one layer above. And so
0:11:06 they obviously don’t want that either. Because now they’re in this position of maybe they’re going to
0:11:12 blow up. And so really, what like the buyer of GPUs wants is like a short term contract on a lot of
0:11:17 GPUs. What the vendor wants to sell you is a long term contract on those GPUs, right? And you need
0:11:22 some way to like combine those two problems together. And the way that you would do that
0:11:29 is with a market, you allow someone to buy a long contract that they can then sell back. And because
0:11:33 they can sell back, they can exit at any time, and they may not exit at the price they got it at.
0:11:39 But it’s so much objectively better to be able to exit, that you end up saving like 40 to 50%
0:11:45 on most like load curves. And if you don’t have that, what happens is, if you’re serving,
0:11:52 say inference on your GPUs, and your margins are like 20%, and suddenly you have a 20% drop off or
0:11:56 something, you may just be underwater if you’re not able to sell back. That was a lot of stuff.
0:12:01 I’m just gonna like pause there because that I did tell you I can literally go on forever about this.
0:12:05 Yeah, I was like, okay, I’m getting like 90% of this. And at the same time, while you’re talking,
0:12:08 I’m thinking about how to summarize it in a way that’s actually very simple. And I’m not sure
0:12:10 I’m able to do that today.
0:12:15 It’s so much stuff. Yeah, it used to be the case that software was just software. And you didn’t
0:12:21 have to think that hard about stuff like this. Yeah. And now software is not software. Now software is
0:12:22 finance.
0:12:25 Yeah, it feels like there’s so many different layers to a finance and different people involved
0:12:29 where something could go wrong. And you guys by creating like a spot market, you’re trying to
0:12:34 solve that where if I had the GPUs or have a, you know, a data center, I could still get a long term
0:12:39 contract with someone like with like SF compute. But the buyers, they’re able to get a shorter term
0:12:43 deal because that’s what they want. They just want to get in and get out and test things. And yeah,
0:12:48 once they prove things out, maybe do more. It’s just we sell long term contracts, you can get out
0:12:52 of. Yeah. And if you can get out of the contract by selling back, that’s just objectively better than
0:12:54 not. Yeah. That’s basically our business model. Got it.
0:13:03 Okay, let’s be honest, your AI prompts aren’t giving you the results you deserve. But with a
0:13:08 little coaching, you can transform from basic prompts to engineering conversations that get you
0:13:15 exactly what you want from ChatGPT. That’s what this playbook delivers, not just random prompts,
0:13:21 but a step by step system with the exact techniques top AI professionals use every day. You’ll have
0:13:27 your own personalized prompt engineering system that delivers consistent results. Get it right now.
0:13:32 Scan the QR code or click the link in the description below. Now let’s get back to the show.
0:13:40 It’s pretty simple from a face value, but like how you do it and why it happens and why it’s the case is
0:13:45 this like really complex piece. So how long have you been doing this right now? It’s been like a year or two?
0:13:48 or it’s been a little bit longer than that? We’re about two years old. Yeah. When we started the
0:13:53 company, we were an audio model company. So we were training big audio models. If you’ve heard of
0:13:57 companies like Suno or Udeo, we were kind of a thing like that, but before Suno or Udeo.
0:14:01 Interesting. And it was called something else or SF Compute? It was called June Lark. It’s called
0:14:06 June Lark because it was June when we started it and we wanted to go on a Lark. And we didn’t think
0:14:10 the thing would last for more than a summer because we weren’t entirely sure what we were doing at the time.
0:14:15 But the plan was that we would fly down to LA, try to find someone who would license us the music data
0:14:21 and audio data. And then we would buy a GPU cluster for some short period of time, train our model,
0:14:25 and then like make a consumer app or something like that. And it didn’t work out very well.
0:14:31 I got stuck on the name of June Lark for some reason. I thought that was pretty funny how you
0:14:35 just started it because it was in June. It reminded me of Theo Vaughn talking about his podcast this last
0:14:39 weekend where somebody asked him like, you know, why did you start call it this last weekend? He’s like,
0:14:42 well, I started on a Monday or something like that. Yeah.
0:14:49 You know, one thing I’ve been thinking about a lot is, you know, I studied Mandarin in Taiwan. So maybe I’m
0:14:53 biased and had some connections with people in the Chinese government in the past and things like that.
0:14:57 And I’ve definitely become a person who’s like, okay, America needs to beat China, especially in AI.
0:15:02 I wonder what are your thoughts on like where things are at with like the US versus China in terms of,
0:15:08 you know, obviously the US has NVIDIA, but it looks like now China’s, you know, Huawei’s creating chips.
0:15:13 Like, are they close to being competitive? And since China has more power, are they actually ahead?
0:15:17 Or like, how do you see that? My guess is you’re going to run out of power before you’re going to
0:15:23 run out of chips. Basically, the total capacity that people have been throwing around is like more than
0:15:30 the current load of the US at any given rate. So the chips got to go somewhere else, or you got to make
0:15:34 more power in the US. And would that be like to serve like one country? Or would that be like
0:15:35 enough to serve the whole world?
0:15:48 What’s the current current load capacity of the United States? It’s like 1.3 terawatts in the US. And so if you’re
0:15:55 talking about adding another 100 gigawatts, that is a really meaningful amount that you’re trying to bring
0:15:57 on in like a year or two.
0:15:58 And that’s just for XAI.
0:16:05 Yeah. And that’s just for like XAI. You’re talking about like, a single industry increasing load growth, if you were to put all the
0:16:12 chips in the US, you’re talking about like a single industry increasing load growth by like 10 to 20%. And the rate of growth of
0:16:18 that, if it were to continue, would like double the load growth in the US. Or maybe to put it a different way, I think there’s
0:16:22 about 100 nuclear power plants in the US, you’re like doubling that.
0:16:25 Every year, like every few years.
0:16:32 Or like over a reasonable period of time. But in order to bring on these chips, you either need to be able to spin out
0:16:39 power plants faster, especially really high power density power plants, things like nuclear power plants, or you need to put
0:16:46 them outside the US. And I think for the US in particular, you probably don’t want the chips to go outside the US.
0:16:51 Like you probably want as much of it as possible within the US, partially just for economic reasons.
0:17:00 Also, because if you start to use these for higher security reasons, and I think there’s reasonable, like, case to make that you would want to do that.
0:17:12 You really don’t want somebody to walk into a data center with a gun, basically, like, right, the physical security of the data center probably matters quite a lot. And you’re just way more likely to maintain the physical security of a data center.
0:17:21 If it’s on US soil, then if it is not US soil, right. And as these things get bigger, and more critical to the world and the economy,
0:17:30 you don’t want to like leak the weights. And like, a really simple way to leak the weights is to walk into the data center with a gun, or for like a government that
0:17:36 doesn’t, like, have a lot of norms to just sort of say, hey, data center, you need to let us in now.
0:17:47 Or like an intelligence agency could do some kind of operation or whatever, you know, exactly. And, you know, even if it’s a government that today seems like a US ally, maybe tomorrow they aren’t.
0:17:59 And so I would rather have them in the United States than not in the United States. But then also you’ve got the US and China competitiveness concerns. And to solve for that, you need to bring on a lot of power.
0:18:09 Yeah, that’s where it feels like China has a huge advantage, right? It just feels like with their government structure, they will just move away quicker. They’ll just like, yes, we’re doing all this. Yes, yes, yes.
0:18:29 And they don’t have to talk to anyone else. It’s like literally, okay, getting started now. Like, when are we getting started? You know, next week. Yeah, also, the, the work ethic, I think of the culture, at least right now in China, I think people work quite a bit harder, and they’re expected to do a good job. And there’s difference culturally there right now, as well, I think is an advantage.
0:18:34 I will say that San Francisco is probably the beacon for hard work in the country.
0:18:55 Yeah, especially now, like there was a bit where it had slowed down quite a bit, at least when I was there, like when I first got there, it felt like everyone was working really hard. Then it felt like there was like a seven year period where there was, you know, everyone was talking about like work life balance. And I think it was because things had just matured. And like, it was like, okay, Facebook’s won the social media war and this, you know, and Twitter, and people had won their categories.
0:19:22 And it was just not as much competition for some huge thing that they were trying to win, right. But with AI now, it feels like everyone obviously realizes how big the stakes are like, okay, maybe Facebook doesn’t even make sense in the future because of AI, maybe these companies who thought they had won, like Google, maybe they no longer exist if they don’t win this new race. So yeah, I also think that it is a big enough shift for something important enough, that has caused a lot of people to just want to try more.
0:19:33 If you think about, you know, maybe four to five years ago, the last cycle was like a crypto cycle, maybe didn’t have as much like interest for a lot of folks, as in, it’s just distinctly different.
0:20:02 It’s very different. Like I was involved in crypto early on. I bought Bitcoin, you know, pretty early. When I saw Bitcoin, I was like, okay, at first, I wasn’t even sure. Like I came from like the game industry. So I was like, maybe this is just some stupid thing that’s gonna just like fizz out, fizzle out after a few months. But then I started realizing, okay, there’s like real uses in terms of remittances, like sending money to countries where maybe their governments control them using the currency. So that’s the one use case that always made sense to me. But then after that, people made like another 100 use cases that a lot of them, honestly, never made any sense to me.
0:20:20 Yeah, like to steel man the crypto case, like, I think just like being money and being a better version of money is like totally realistic case. And that’s like a really big and important deal. But being money, or like making a really great money, just doesn’t feel as exciting or interesting to a lot of folks as versus like, we’re gonna make robots.
0:20:36 Yeah, and we’re gonna make it possible for you to build a skyscraper for pennies, because we’ve automated the entire system, like we’re building the whole thing, all the sourcing of the materials, we like did all the legal work automated as well. And we’re gonna build that skyscraper over there for like the cost of the electricity. And by the way, we’re also going to bring down the cost of electricity.
0:20:42 By the way, the stuff you’re saying right now, I think like 99% of people have no idea that that kind of stuff is most likely coming.
0:21:06 Yeah, like that’s the sort of pitch of like, what if we just took all the costs of things and brought them to zero, or just the cost of electricity. And I think that’s a much more sort of compelling thing. And then also just the things that people end up working on, they just feel like science fiction, in a way that a lot of other previous trends felt like we’re going to do remote work, or we’re going to make things on mobile phones, or we’re going to make a bunch of characters on a social media app.
0:21:11 And it wasn’t like, we’re gonna cure cancer. Yeah, or we’re gonna make a robot.
0:21:29 Yeah, totally. When I was a kid, I’d read like science magazines and sci fi magazines, everything else. And there was no talk about money. We didn’t think about like, oh, it could be a new money. For most people like who cares, like whatever, but we definitely dreamed about robots. That was one of the biggest ones robots and space travel and yeah, building new kinds of buildings, beautiful cities, using robots, which hopefully does happen.
0:21:48 Yeah, technology used to be technology. It finally is once again, technology is technology. Like technology used to be the spacecraft, and it used to be electricity, which was like actual technology. And I think it once again, actually feels like technology. And so if you are in San Francisco, a lot of people come here because it’s once again, the actual mecca of technology.
0:22:03 Okay, let me tell you about another podcast I know you’re gonna love. It’s called Billion Dollar Moves, hosted by Sarah Chin Spellings. It’s brought to you by the HubSpot Podcast Network, the audio destination for business professionals.
0:22:13 Join venture capitalist and strategist Sarah Chin Spellings as she asks the hard questions and learns through the triumphs, failures, and hard lessons of the creme de la creme.
0:22:28 So you too can make billion dollar moves in venture, in business, and in life. She just did a great episode called The Purpose Driven Power Player Behind Silicon Valley’s Quiet Money with Mike Anders. Listen to Billion Dollar Moves wherever you get your podcasts.
0:22:56 So to get back on how the US beats China. So like, it’s cool that San Francisco is working hard, but that still doesn’t change it. We need a lot more power. You know, talking about robots, I kind of wondered if that’s a possible scenario where America could be China, because we start using robots to produce these nuclear factories and everything else. And maybe if we get ahead of China, but then I see all these videos, at least on X, where people are showing all these videos of robots in China.
0:23:15 I’m like, ha, I thought Elon Musk was probably ahead. Now I’m not so sure. So if China ends up ahead in robots, as well, we’re in trouble. And then me being here in Japan, I keep wondering, like, should I be trying to help in some way to forge some kind of alliance between the US and Japan? Because I think Japan would be a natural ally to help build those robots to be China.
0:23:35 Yeah, I think the US just needs to get rid of its autocracy thing. And this is like mostly a state or city level thing frequently. Like in California, we have CQA, which is basically just like a law that says when you want to build something, you need to prep this document that’s like every possible environmental thing that could ever go wrong, even if it’s not actually a real thing.
0:23:47 Because if you don’t prep a long enough document, you can get sued for not prepping a long enough document. So like you build a bridge or something, and your bridge needs to identify every single possible bird species that could ever have like flown into the bridge or whatever.
0:23:52 Yeah, a lot of that stuff, I’m sure it had good intentions when it started. But now over time, it’s gotten more and more, more ridiculous.
0:24:06 Yeah, it had good intentions. A lot of California and in general, a lot of the US’s problems, I think are good intentions gone wrong. And I think oftentimes people are quick to blame a person when actually just like, a lot of times you just make mistakes.
0:24:25 When you’re saying all that, I’m sitting there thinking of like Hitchhiker’s Guide to the Galaxy, the what is it, the Gorgons or something like that, I think, maybe I’m saying it wrong. You need like a stamp of approval for like almost every single possible thing. It’s like, they can do anything. It’s like, well, here, you need approval for this. Do you have this approval? Do you have this one for this? And it just can be the most basic thing possible, but like you need approval for everything.
0:24:32 Yeah, it feels like that’s what America is drowning in right now. And I’m not sure how you get out of it, because you kind of need approvals to get out of all of it.
0:24:53 Yeah, I think a lot of progress studies folks seem like they’re pushing a lot. That seems very effective. There’s like new ways to organize groups of people. There’s a thing called like sway.app or something. It’s like a way to basically organize a bunch of people to vote on like niche issues, which seems like a useful thing. But yeah, I think you just, you just need to try. You gotta get a lot of people trying.
0:24:56 Cool. I had some selfish questions.
0:24:57 Go for it.
0:25:25 So I think I told you, I’m planning on announcing that like the new direction of lore.com is I’m going to be trying to acquire data centers in collaboration with Jesse Tinsley. Jesse’s this guy who has built up MainStreet.com and Employer.com. He’s basically done a roll up of like 10 amazing like HR recruiting accounting companies. And he’s done a really great job of that. And it’s kind of saturated that market. And we’re looking to kind of take some of the financial engineering he’s learned and apply that to acquiring data centers and operating them.
0:25:46 But we’re still like at the idea phase and figuring out like, okay, what is our actual plan? And we have different ideas. If you were me, what would you be looking at? Would you be looking at like, I don’t know, maybe acquiring crypto miners and repurposing them or like finding places where maybe like the hyperscalers, they’re trying to get so big that they have smaller facilities they want to offload or I don’t know, how would you approach it?
0:25:49 So to be super clear, are you talking about the physical data center?
0:25:49 Yeah, physical.
0:25:54 Also the cluster inside of it. So it’s like the building and then there’s also the chips inside of it.
0:25:59 Probably both. I’m open minded. Do you think it’s, you know, which play do you think is more interesting?
0:26:08 Let’s say you’re doing both the data center and the chips inside. We know more about the chips inside part. The things I would try to do is do everything in your power to reduce your cost of capital.
0:26:26 And if you can own the data center and own as much as your supply chain as you can, because you can compress your cogs by basically removing the margin from the vendor you just paid. So that’s number one. Number two is you need offtake agreements for as long as possible. So you need long-term contracts with somebody who’s going to pay you before you buy the chips.
0:26:39 SF compute comes in or that is true. Like SF computes business. So I joke about this on the internet that nobody knows what SF compute does. What SF compute does is we’re an offtake machine. We generate offtake for people who are setting up clusters. That’s what we do.
0:26:48 And it’s like not at all obvious. And part of the reason it’s not at all obvious is because the way that we generate offtake, the way that we’re able to get people long-term contracts is by running a thing that looks like a cloud provider.
0:26:59 And like no customer who has ever come to us looking at us like a cloud provider ever knows the word offtake or like they don’t care. And so we try to very hammer focus on exactly what those customers care about.
0:27:09 And then on a BD sense, like us just going and talking to people, we pitch, hi, we’re your offtake provider. Like we, we’re the offtake generator. That’s the goal of the company. We’re like a machine that creates offtake.
0:27:15 So are you partnering with a lot of like private equity groups and stuff like that? Like, what are you seeing, you know, who are the main kind of customers there?
0:27:26 So we work with folks who own clusters and there are lots of different reasons why someone might own a cluster, but yes, there are many folks in traditional finance who see compute as an asset class and they would like to invest money in that asset class.
0:27:36 Kind of like what you’re considering as well. Our goal is to partner with folks like that and basically get them offtake, basically saying, hi, if you set up a cluster, we will make sure that you get a contract on it at reasonable prices.
0:27:41 So that way you can finance it correctly and you make money on it. Our goal is to try and help people make money on the clusters.
0:27:46 So another caveat here is that we typically run the clusters. So to some extent, we look like a property manager for GPU clusters.
0:27:52 So that’s the other complexity is we make all the software stacks that a Core Reave or a Lambda might’ve made for themselves.
0:27:56 And then we license it to people or we use it to operate their clusters and sell them. That’s how we work with folks.
0:28:01 So someone set up a cluster, we like go take over the cluster for them and sell it for them and give them offtake.
0:28:07 So they can focus more on treating it like a real estate asset versus having to figure out how to manage it and find all the customers and everything else.
0:28:07 Yeah.
0:28:13 Correct. And then you get a bunch of money and you take that money and then you go and set up another cluster and you just print cash.
0:28:17 That’s the goal. And that’s why we say compute works more like real estate. It doesn’t work like startups.
0:28:21 A lot of people try to do this thing where they set up clusters and then they try to build a great product on top.
0:28:29 And then they realize the customer is so price sensitive that it doesn’t matter what product they build on top of the cluster, the customer isn’t going to give you any markup on your GPUs.
0:28:37 It doesn’t work like CPUs where CPUs, you can buy some CPUs and make a great product and get like a 60 to 80% margin if you want on your CPUs by just building a really great product.
0:28:41 But no one cares about the product. They only care about the cost of the hardware in GPUs.
0:28:45 Yeah, definitely. It sounds like, yeah, it’s more real estate numbers play.
0:28:48 Do the numbers work or not than how to, you know, optimize things as much as possible.
0:28:54 I was talking to some top partners at A16Z. I’m not going to say their names, but they told me the same kind of stuff.
0:28:59 Like they’ve actually worked on with one or two of their companies that are like setting up data centers.
0:29:04 Like the deals are incredibly complicated, you know, equity and different kinds of debt.
0:29:09 Maybe not even one kind of debt, like multiple different kinds with different parties and maybe governments are involved as well.
0:29:13 Yeah, it’s crazy. So I think I get what SF Compute does now.
0:29:18 I’m not entirely sure I could summarize it in like a sentence yet to think about a little bit more.
0:29:18 We sell GPUs.
0:29:22 Yeah. Okay, that makes sense. I think I can remember that.
0:29:26 Where do you see things going like in the next five years with SF Compute?
0:29:29 Like if things are working well and you’re doing great, like where do you see the company?
0:29:32 So we think of ourselves as kind of two companies in one.
0:29:37 One is a systems engineering company that like builds the really low level stack of operating a cluster.
0:29:39 And then the other part is a fintech company.
0:29:46 That’s goal is basically to make financial products that make it easier to financially engineer a cluster and get the cost of capital down quite a lot.
0:29:47 So that’s our general goal.
0:29:49 There’s lots of stuff that we’ve got planned for that.
0:29:52 At the moment, the order book that we built today, that’s our key product right now.
0:29:55 But growing that order book, there’s lots of things that you can do with that.
0:30:06 So you think that data centers could become a huge asset class, just like premium real estate or like commercial real estate buildings, skyscrapers, malls, but they’re a bit more complicated because all the technology.
0:30:08 So you guys kind of simplify that.
0:30:12 So the finance people can treat it more just like a traditional real estate play.
0:30:15 Yeah, I think everybody wants to treat it like soybeans and instead it’s supercomputers.
0:30:16 Right.
0:30:20 So let them treat it like soybeans and you’ll take care of the supercomputer.
0:30:20 Correct.
0:30:24 Our goal is to wrap the supercomputer into something that looks like soybeans.
0:30:32 So that way you can very easily finance it and you can simplify it down enough so that the gears of Wall Street don’t have to do as much complex financial engineering.
0:30:45 And hopefully what this means is that you end up pricing the risk of owning a GPU cluster correctly and you end up with fewer bust and boom cycles and a broadly more stable environment in which lots of capital can flow in and make AI cheaper.
0:30:46 Which is good.
0:30:48 That increases the chance that America will win, which is what I care about.
0:30:50 Yeah.
0:30:51 Correct.
0:30:52 Cool.
0:30:52 Yes.
0:30:55 America’s greatest asset is its depth of capital markets.
0:31:00 And if we put those depth of capital markets to work, we can outspend anyone.
0:31:01 So let’s do it.
0:31:03 You just got to do it very safely.
0:31:04 Okay.
0:31:06 A few fun questions before we go.
0:31:08 And this has been fun and definitely different kind of episode than we normally do.
0:31:10 But just leave with some fun stuff.
0:31:14 So my son here in Japan, you know, he grew up in San Francisco, but he’s in Japan now.
0:31:15 He’s 12 years old.
0:31:20 And I’m constantly thinking about, like, what should I be teaching him so he’ll be successful in the future?
0:31:22 Because things are dramatically changing.
0:31:30 You know, when he was little, maybe he was like two or three, we had some gatherings at Lafayette Park, birthday parties, and some of my friends in tech were asking, like, what do you think he’ll do in the future?
0:31:34 And I talked about, like, anti-drone systems and stuff like this.
0:31:38 And, you know, what would you be teaching if you had a son or a daughter who was like 12?
0:31:41 What would you be teaching them to help them be as successful as possible in the future?
0:31:48 One thing is that I would probably want to make sure that they’re using the various chatbots, the AI bots, to teach themselves or learn.
0:31:52 They’re super effective teachers, just, like, incredibly good at teaching you things.
0:31:52 Yeah.
0:31:54 I would also want to question them a lot.
0:32:01 Like, the ability for them to just say kind of whatever you want them to say is entirely dependent on the person who’s making it.
0:32:05 And that seems like a thing we haven’t really reckoned with super well yet.
0:32:07 And it’s often, like, completely accidental.
0:32:09 Are you talking about the fact that you can lead the LLMs?
0:32:12 Like, if you say it in a certain way that they respond back based on how you said it?
0:32:13 Is that kind of what you’re…
0:32:19 Yeah, like, you’re fighting with your partner and you say, I said this and they said this.
0:32:28 And the way you frame the question encourages the bot to probably not give you the most wise advice, but instead to give you advice that, like, is the one that you would want to hear.
0:32:28 What do you want to hear?
0:32:38 But if you were using the same bot just a moment ago to ask it questions about, like, hey, I, you know, have this pain in my side.
0:32:43 And it says, oh, that’s, you know, this disease and you should go check it out and then you check it out and it’s real.
0:32:44 Like, well, you know what I mean?
0:32:50 Like, if you’ve used it previously to get very accurate information, things that seem really quite correct, because it’s really good at that in particular occasions.
0:32:54 But then you asked it on something else that, you know, maybe a personal life question.
0:33:02 You assume that it will give you a lot of accurate information, but instead it gives you information that it thinks you want to hear because, you know, maybe it’s subjective.
0:33:04 And I think that can lead people astray.
0:33:08 Yeah, I think a lot of people see it like it just answers things, but like, no, it helps you think it through.
0:33:16 And like, so if you need to push back on what it says, too, and like, and it should be a way of, like, probably almost like, you know, extension of your thought process versus just giving you an answer.
0:33:18 And you tell it to do that, because it will totally do that.
0:33:20 Like, if you tell it, hey, be critical of the things I’m saying.
0:33:23 And, you know, don’t totally take my advice either.
0:33:24 It totally can do that.
0:33:24 It’s great.
0:33:26 I think that’s in my custom instructions, I think.
0:33:34 Yeah, so I can imagine that for, like, a 12-year-old, just being more aware of that is probably useful when you’re, like, more impressionable.
0:33:35 Cool.
0:33:35 Makes sense.
0:33:36 That’s kind of what I’ve been trying to do.
0:33:40 I need to do it more, but, yeah, definitely been having him use all the different LLMs.
0:33:44 And even I’ve had him try things like Replit and things like that to create apps, which has been cool.
0:33:47 And also trying to teach him, like, it’s not just, like, a new Google.
0:33:52 Like, you do have to, like, actually, like, push back and talk with it and not just accept exactly what it says is, like, you know, gospel.
0:33:54 One last fun one.
0:33:56 So you get in a time machine.
0:33:57 You step out in San Francisco.
0:33:59 The year 2050.
0:34:00 What do you see?
0:34:01 Is everything the same?
0:34:02 Or is it all different?
0:34:03 Is there robots everywhere?
0:34:05 You know, what’s going on?
0:34:06 Okay, I’ll make a bunch of predictions.
0:34:06 Okay, cool.
0:34:08 First, I think we’re probably going to build a lot.
0:34:13 San Francisco is maybe the first place to stop building and hopefully the first place to start building again.
0:34:16 So I think you’ll probably see more taller buildings and more denser buildings.
0:34:19 I think self-driving cars will probably get rid of parking.
0:34:24 So I think one of the things you might literally notice walking out the street in 2050 is that there isn’t street parking.
0:34:29 I think that will genuinely be, like, a thing that it’s a visual thing that you will notice quite immediately.
0:34:33 I think you probably can walk on the road really safely because of that.
0:34:37 Like, just because if everything is self-driving, one of the things that happens with Waymos is you kind of just, like, step in front of the Waymo.
0:34:40 You don’t exactly, but, like, it’s fine.
0:34:41 It’s fine.
0:34:43 It’s crazy how good they are.
0:34:44 I saw that.
0:34:52 There was a guy, because, like, in Japan, one thing I noticed is in Japan, you know, I don’t know why, but, like, you know, people, you kind of can predict what they’re going to do more.
0:34:54 But, like, in San Francisco, it’s like, man, this is chaotic.
0:34:59 Some guy walked out in the street and stopped and checked his phone and then he turned around, like, right in front of our Waymo.
0:35:02 Yeah, because, like, the Waymo’s not going to hit you.
0:35:04 The Waymo saw you from, like, two miles away.
0:35:07 Maybe in the future, the other Waymo saw that Waymo and it’s, like, sharing data with each other.
0:35:09 Like, they already know where everybody is.
0:35:09 Yeah.
0:35:15 So, my most hopeful prediction of what could literally change, and this might change everywhere, is, do you know what pipe dream is?
0:35:17 Geez, I’ve heard of it.
0:35:18 I’m not sure.
0:35:32 I think it is the most interesting, like, piece of technology that is obvious, possible, and appears to be happening, and I really hope it happens, which is basically underground, like, pipes, and then they have a little robot in it for deliveries.
0:35:32 Oh.
0:35:41 And so what happens is you end up in your house with, like, a dishwasher-shaped thing, like, a literal thing, like a door that you pull out, and then you can get your deliveries there.
0:35:45 And that changes a bunch of stuff about the way that you go about the regular world.
0:35:51 So, for example, you’re ordering, like, Uber Eats or whatever, you don’t, like, send a car to your house for your burrito or whatever.
0:35:56 It just, like, appears in, like, this weird dishwasher thing, and, like, that’s a little thing, and then, like, you just have, like, a burrito.
0:36:00 And you want to order something from Amazon or something, it just, like, appears in your house.
0:36:04 You don’t have to worry about, like, the package getting stolen, it’s just, like, already there.
0:36:05 That’s cool.
0:36:08 So, I know Taizo-san, Masayoshi-san’s brother.
0:36:12 He was an investor in my previous company and we became friends, and he was doing this thing out in Singapore.
0:36:20 He was getting land in Singapore to build, like, a tech, you know, futuristic city, and this was one of the main pilots he was pitching was exactly what you’re talking about now.
0:36:20 Yeah.
0:36:24 It’s, like, the delivery through, you know, pipes and things like this using drones or whatever.
0:36:27 I’m thinking about, like, the things that would physically change in your life.
0:36:32 The other thing is, like, I am super hopeful that all the AI for bio stuff works.
0:36:37 I’m hoping we get just incredible new medications that just cure a whole bunch of diseases and make everything a little bit better.
0:36:41 I have no idea because I have no conception of biotech.
0:36:42 Like, I don’t know anything.
0:36:48 But we have worked with a lot of folks who seem to need GPUs, and I’m so incredibly excited for them to work.
0:36:49 You’re like, oh, a bio customer.
0:36:53 You know, like, you give them a discount, like, you know.
0:36:55 Actually, yeah, we’ve done stuff like that before, yeah.
0:36:57 We used to work with a lot of university labs that were doing stuff like this.
0:37:01 And we were one of the few places where you could get a lot of GPUs for a short period of time because that’s what our product does.
0:37:06 And that is a lot more affordable for a university than having to buy for a year or something like that.
0:37:06 Cool.
0:37:09 Evan, this has been an awesome conversation.
0:37:10 I’ve enjoyed it a lot.
0:37:12 And next time you come out to Japan, we’ll hang out?
0:37:13 Next time.
0:37:13 When I’m in San Francisco.
0:37:15 But where should people find you?
0:37:19 Maybe tell them where to find you on X or your website, what they should check out.
0:37:21 So our website is sfcompute.com.
0:37:25 And you can find me on the internet at EvanJConrad on Twitter or X or whatever it’s called now.
0:37:27 Okay, cool.
0:37:28 Evan, it’s been great.
0:37:29 Thanks for coming on.
0:37:29 Likewise.
0:00:10 every week? Well, today we’re going deeper beneath all of that, underneath the hype,
0:00:15 the headlines, and the glossy demos, to the actual metal and concrete foundation that makes
0:00:20 it all possible. We’re talking about compute infrastructure, the biggest physical buildout
0:00:25 in modern history. So this is a really complicated topic, but I think it’s important to understand.
0:00:31 So today I brought on the smartest person I know in data centers, Evan Conrad, the CEO of SF Compute.
0:00:35 He’s building something absolutely genius. They’re creating a spot market for compute,
0:00:40 basically turning supercomputers into a tradable commodity like oil or soybeans,
0:00:43 which makes it possible to finance this infrastructure in a more rational,
0:00:48 sustainable way. Understanding this layer of AI is like understanding the foundation
0:00:53 of the internet in 1999. So let’s dive in with Evan Conrad from SF Compute.
0:01:05 Being a know-it-all used to be considered a bad thing, but in business, it’s everything. Because
0:01:13 right now, most businesses only use 20% of their data, unless you have HubSpot, where data that’s
0:01:18 buried in emails, call logs, and meeting notes become insights that help you grow your business.
0:01:23 Because when you know more, you grow more. Visit HubSpot.com to learn more.
0:01:29 Evan, thanks for coming on here today.
0:01:30 Thank you. Thanks for having me.
0:01:35 Yeah. You know, so I’ve been seeing all your posts on X about like how fast your company SF Compute is
0:01:40 growing. 3X, 5X, 7X, it sounds like SF Compute’s been taking off. And then in the meantime, I’ve been
0:01:44 seeing all this news about compute like everywhere I look, you know, like Sam Altman’s planning on
0:01:50 creating factories. They’re going to build one gigawatt of compute per week, which for people who don’t
0:01:55 know, that is a gigantic amount of compute. Elon Musk is saying he’s racing to 100 gigawatts
0:02:00 of compute in the next few years, which is just a crazy number. And you know, I’m just wondering,
0:02:04 like, are we in a compute bubble? Or where are things at?
0:02:09 I don’t think we’re in a bubble quite yet. But if there is a bubble, I can kind of explain where
0:02:14 it probably will be. Sure, that’d be great. So what I want to do is I want to walk through how you got
0:02:19 from, say, ChatGPT or an application that you might use, all the way down to the physical layer of the
0:02:23 cluster, and how the financing of that entire structure works. Because if you’re trying to find
0:02:28 the economic bubble, you want to find like who’s paying who what money and where and where the
0:02:33 miscalculation is coming from. So in most cases, if you are building up a cluster, you’re looking for an
0:02:37 offtake agreement. That means a long term contract for your cluster. So that way, you could get it
0:02:41 financed in the first place. So you’re going to buy a bunch of GPUs, you’re going to put them in a
0:02:46 data center somewhere. And then you need someone hopefully before you’ve actually bought the GPUs,
0:02:50 who’s agreed to purchase them for some period of time, like a rental agreement. And you take that
0:02:55 contract, and you go back to a lender. And now the lender says, okay, because they’ve already agreed
0:02:59 to pay for it, really, what I’m evaluating is not the broad or market or, you know, whether or not
0:03:03 people want GPUs. I’m really just evaluating, does this customer want
0:03:06 So it’s kind of like almost like commercial real estate or like a mall or something like that,
0:03:09 where you’re getting like the anchor tenants, and they’re lined up and then it helps get the
0:03:14 financing. Correct. If you have the anchor tenant, that person is the person who’s doing offtake.
0:03:19 That’s the person who you’re giving the cluster to give the customer already. And so the lender is
0:03:24 financing against the credit risk of the person who has the offtake agreement. You’ve seen this in a
0:03:28 number of different deals before. So for example, like the Microsoft core read deal, the nice part
0:03:32 about that is no one thinks Microsoft is going to default on their debts. That would be kind of like
0:03:37 a crazy thing to think. And so because of that, you can finance it very cheaply, your cost capital is
0:03:42 really low. And so that’s great. That’s super good. The problem though, is a Microsoft or somebody else
0:03:48 who’s getting the offtake agreement, then sells it to somebody else who doesn’t have good credit risk.
0:03:53 is that what they’re doing? Um, not always. Okay. So when if Microsoft is selling it to OpenAI,
0:03:58 you’re betting that OpenAI is going to continue to grow and OpenAI has very meaningful revenue. I think
0:04:04 it’s a quite reasonable case that OpenAI will continue to grow, that also they can back up what
0:04:09 they’re doing. There’s some amount of bet. There’s like a significant bet, especially in Stargate,
0:04:12 when you’re training new models, and you’re betting the new model is going to be better than the old
0:04:16 model, and you’re going to have some economic return from it. But historically, the previous models
0:04:22 worked. They trained GPT-4 and then GPT-4 made a whole bunch of money. And so GPT-5 is making a bunch
0:04:28 of money. And so presumably you will continue to have this grow, but there is the whole rest of the
0:04:33 market as well. So not everything is OpenAI, not everything is Anthropic. There is also all the
0:04:39 various application layer companies that are popping up. Basically every YC startup and every company that is
0:04:45 currently getting funded, all of those people are signing much shorter term contracts because they can’t
0:04:49 afford a longer way. And why do most of those YC companies need compute? Would they not be able to
0:04:54 just use OpenAI as API or Clot? Or are they having to do more like custom stuff on top or something? I
0:04:59 don’t really fully understand that. Sure. So in some cases, they’re using like inference layer services.
0:05:03 So this is like an open source model that they’re getting from a Fireworks or Together. There’s like
0:05:07 a hundred different services here. But eventually you’re using an open source model, or you might just be
0:05:10 using OpenAI and Anthropic, and then you’re signing short-term contracts with them.
0:05:16 The problem is that if you are either OpenAI or Anthropic, or you’re one of the other inference
0:05:20 platforms that are serving open source models, or you’re doing something like an image model,
0:05:23 or you’re doing something like a video model. So things that are kind of have a lot of other
0:05:28 competitors that aren’t OpenAI and Anthropic. If you’re signing short-term contracts with those
0:05:33 folks, a short-term contract might mean literally just a subscription. You pay a usage-based billing,
0:05:36 and then you pay at the end of the month. And then if you use less the next month, you use less the next
0:05:41 month. And so you’re just paying on a monthly basis. You’re not locked into anything. Now, of course,
0:05:45 these folks will want you to sign a long-term contract, but a lot of you aren’t on that long-term
0:05:51 agreement with their underlying inference provider. So when you do that though, now you’re betting on the
0:05:58 credit risk of a bunch of new startups who currently have a reasonable credit history because they have
0:06:04 just raised at very large valuations. And they raised at large valuations because they initially
0:06:09 made a lot of revenue very quickly because customers are actually willing to pay lots of
0:06:13 money for the AI applications, for all the different services that are happening. Their margins aren’t
0:06:20 great, but the volume is really big. And so to some extent that causes problems because the venture
0:06:26 capitalists look at the low margin, but high volume application layer products. So this is like your
0:06:32 code editor tools. Yeah. These sorts of things. And to some extent, some of the built-in valuation
0:06:38 is betting that the margins will eventually improve. And in some cases that might be true,
0:06:44 but in a lot of cases it might just literally not be true. And what that means is that if your ability
0:06:50 to sign shorter contracts or longer contracts, your credit risk basically is based on your ability to
0:06:56 raise venture capital, but that venture capital wasn’t correct about the ability for these companies
0:07:02 to eventually become higher margin like SAMS, then there is your bubble. Basically your problem then
0:07:09 becomes if a macro event were to occur, for example, if there’s some sort of larger economic concern that
0:07:15 pulls back venture capital, then what would happen is these companies might struggle to raise in the
0:07:19 current funding environment, which would harm their ability to get credit risk, which would make them
0:07:24 need shorter term contracts or their shorter term contracts would fall off because their customers
0:07:29 would fall off. Which means for the inference layers who now are basically signing long-term contracts with
0:07:35 their underlying compute clouds, but selling short-term contracts to their customers. Now they don’t have
0:07:40 the demand coming in to pay their underlying bills that they still have to pay, which then goes back to the debt
0:07:46 providers of the clusters who thought they were betting on something with reasonable credit risk, but the credit
0:07:50 risk wasn’t actually as reasonable as they thought. And so some of these folks could get wiped out.
0:07:58 And because of the total amount of money in that system, because compute is so big and AI is so big,
0:08:02 it can have like broader market effects, which probably then encourages LPs to pull their money out
0:08:08 of the venture capital firms, causing a sort of collapse. That is the fear case. And it’s one of the
0:08:12 reasons why our business exists. So you’re betting on the bubble or just like if the bubble happens,
0:08:17 it’s better for you, but if not, you’re fine? We’re trying to prevent the bubbles. We’re trying to get
0:08:22 to a stable outcome for how AI should work. Our basic premise is that a lot of people are making bad
0:08:28 decisions on the way they’re doing credit risk. We’re trying to make a much more stable state of AI.
0:08:35 So the key problem is basically that the base layer, those GPUs that you’re setting up, those GPUs, many of
0:08:39 them were built on the expectation that they would get a lot higher margins than they actually are.
0:08:44 Because in CPUs, when you’re selling CPUs, you get really high margins. So for example, if you’re
0:08:50 AWS or your GCP and you’re selling web servers, your markup on your CPUs is really high. And that’s
0:08:56 because your customer that you’re selling to is SaaS. Like we’re using a cool product right now,
0:09:01 it’s called Riverside FM. And Riverside FM probably has high margins. And they probably use somebody like
0:09:06 AWS. And because Riverside has high margins, they were able to give AWS high margins, which means
0:09:12 AWS didn’t have to sign as long of a contract in order to de-risk themselves. Because if they’re
0:09:18 making a 60 or 70% markup and maybe their capacity is down 30% one month, that’s not a big deal because
0:09:21 they made so much money on the other months and their markup is so high, there’s a lot of buffer for
0:09:27 them. But in GPUs, you’re not selling to Riverside FM, you’re selling to AI Riverside FM. And every time
0:09:32 you go to AI Riverside FM, you end up spending money on GPUs, which means that you’re super price
0:09:32 sensitive.
0:09:36 And also, I assume that CPUs don’t change as much, right? Like it seems like they just like
0:09:36 slightly get better.
0:09:41 Correct. Yeah. No one has ever really given a s*** like what CPU their web app is running on. Like you
0:09:45 just don’t even think about it. It’s not even like a concern at all. And if somebody comes along and
0:09:47 tries to sell you a new CPU and charge you more money for it.
0:09:51 Well, people used to, I mean, I was born in the 80s, you know, so like when I was a kid, I had friends
0:09:57 who like ran web servers and stuff. And it was all the pitch was like, here’s the CPUs that we have.
0:10:01 Yeah, here’s the price, you know, and that was, that’s how they pitched it was based on the CPU back
0:10:01 in the day.
0:10:03 Yeah, now no one cares.
0:10:03 No one cares.
0:10:09 But in GPUs, people totally care. And in GPUs, your margins are much thinner. And because of that,
0:10:14 they need to sign longer term contracts with their customers, because otherwise, they end up with
0:10:20 this entire like credit risk bubble that filters its way down. Or they just end up on a more localized
0:10:26 scale with basically just a concern of like, oh, I’ve signed for a three year loan. And if I sell
0:10:30 somebody a month at a time, and they drop off, then I still have to pay my three year loan,
0:10:38 and then I go bankrupt. Not great. So the problem, though, is that in order for that GPU cloud to sell
0:10:43 to customers and not go bankrupt, not like have some random collapse, they need to sell you a long term
0:10:47 contract, which is what most people do. Everyone who didn’t do that went bankrupt, basically,
0:10:50 there were a bunch of people who didn’t do that. And they either all pivoted away from doing that,
0:10:56 or they blew up a bit. And so the problem, though, is now you’re just shifting that to the
0:11:01 customer. So whoever’s buying your GPUs now has the same problem, but like one layer above. And so
0:11:06 they obviously don’t want that either. Because now they’re in this position of maybe they’re going to
0:11:12 blow up. And so really, what like the buyer of GPUs wants is like a short term contract on a lot of
0:11:17 GPUs. What the vendor wants to sell you is a long term contract on those GPUs, right? And you need
0:11:22 some way to like combine those two problems together. And the way that you would do that
0:11:29 is with a market, you allow someone to buy a long contract that they can then sell back. And because
0:11:33 they can sell back, they can exit at any time, and they may not exit at the price they got it at.
0:11:39 But it’s so much objectively better to be able to exit, that you end up saving like 40 to 50%
0:11:45 on most like load curves. And if you don’t have that, what happens is, if you’re serving,
0:11:52 say inference on your GPUs, and your margins are like 20%, and suddenly you have a 20% drop off or
0:11:56 something, you may just be underwater if you’re not able to sell back. That was a lot of stuff.
0:12:01 I’m just gonna like pause there because that I did tell you I can literally go on forever about this.
0:12:05 Yeah, I was like, okay, I’m getting like 90% of this. And at the same time, while you’re talking,
0:12:08 I’m thinking about how to summarize it in a way that’s actually very simple. And I’m not sure
0:12:10 I’m able to do that today.
0:12:15 It’s so much stuff. Yeah, it used to be the case that software was just software. And you didn’t
0:12:21 have to think that hard about stuff like this. Yeah. And now software is not software. Now software is
0:12:22 finance.
0:12:25 Yeah, it feels like there’s so many different layers to a finance and different people involved
0:12:29 where something could go wrong. And you guys by creating like a spot market, you’re trying to
0:12:34 solve that where if I had the GPUs or have a, you know, a data center, I could still get a long term
0:12:39 contract with someone like with like SF compute. But the buyers, they’re able to get a shorter term
0:12:43 deal because that’s what they want. They just want to get in and get out and test things. And yeah,
0:12:48 once they prove things out, maybe do more. It’s just we sell long term contracts, you can get out
0:12:52 of. Yeah. And if you can get out of the contract by selling back, that’s just objectively better than
0:12:54 not. Yeah. That’s basically our business model. Got it.
0:13:03 Okay, let’s be honest, your AI prompts aren’t giving you the results you deserve. But with a
0:13:08 little coaching, you can transform from basic prompts to engineering conversations that get you
0:13:15 exactly what you want from ChatGPT. That’s what this playbook delivers, not just random prompts,
0:13:21 but a step by step system with the exact techniques top AI professionals use every day. You’ll have
0:13:27 your own personalized prompt engineering system that delivers consistent results. Get it right now.
0:13:32 Scan the QR code or click the link in the description below. Now let’s get back to the show.
0:13:40 It’s pretty simple from a face value, but like how you do it and why it happens and why it’s the case is
0:13:45 this like really complex piece. So how long have you been doing this right now? It’s been like a year or two?
0:13:48 or it’s been a little bit longer than that? We’re about two years old. Yeah. When we started the
0:13:53 company, we were an audio model company. So we were training big audio models. If you’ve heard of
0:13:57 companies like Suno or Udeo, we were kind of a thing like that, but before Suno or Udeo.
0:14:01 Interesting. And it was called something else or SF Compute? It was called June Lark. It’s called
0:14:06 June Lark because it was June when we started it and we wanted to go on a Lark. And we didn’t think
0:14:10 the thing would last for more than a summer because we weren’t entirely sure what we were doing at the time.
0:14:15 But the plan was that we would fly down to LA, try to find someone who would license us the music data
0:14:21 and audio data. And then we would buy a GPU cluster for some short period of time, train our model,
0:14:25 and then like make a consumer app or something like that. And it didn’t work out very well.
0:14:31 I got stuck on the name of June Lark for some reason. I thought that was pretty funny how you
0:14:35 just started it because it was in June. It reminded me of Theo Vaughn talking about his podcast this last
0:14:39 weekend where somebody asked him like, you know, why did you start call it this last weekend? He’s like,
0:14:42 well, I started on a Monday or something like that. Yeah.
0:14:49 You know, one thing I’ve been thinking about a lot is, you know, I studied Mandarin in Taiwan. So maybe I’m
0:14:53 biased and had some connections with people in the Chinese government in the past and things like that.
0:14:57 And I’ve definitely become a person who’s like, okay, America needs to beat China, especially in AI.
0:15:02 I wonder what are your thoughts on like where things are at with like the US versus China in terms of,
0:15:08 you know, obviously the US has NVIDIA, but it looks like now China’s, you know, Huawei’s creating chips.
0:15:13 Like, are they close to being competitive? And since China has more power, are they actually ahead?
0:15:17 Or like, how do you see that? My guess is you’re going to run out of power before you’re going to
0:15:23 run out of chips. Basically, the total capacity that people have been throwing around is like more than
0:15:30 the current load of the US at any given rate. So the chips got to go somewhere else, or you got to make
0:15:34 more power in the US. And would that be like to serve like one country? Or would that be like
0:15:35 enough to serve the whole world?
0:15:48 What’s the current current load capacity of the United States? It’s like 1.3 terawatts in the US. And so if you’re
0:15:55 talking about adding another 100 gigawatts, that is a really meaningful amount that you’re trying to bring
0:15:57 on in like a year or two.
0:15:58 And that’s just for XAI.
0:16:05 Yeah. And that’s just for like XAI. You’re talking about like, a single industry increasing load growth, if you were to put all the
0:16:12 chips in the US, you’re talking about like a single industry increasing load growth by like 10 to 20%. And the rate of growth of
0:16:18 that, if it were to continue, would like double the load growth in the US. Or maybe to put it a different way, I think there’s
0:16:22 about 100 nuclear power plants in the US, you’re like doubling that.
0:16:25 Every year, like every few years.
0:16:32 Or like over a reasonable period of time. But in order to bring on these chips, you either need to be able to spin out
0:16:39 power plants faster, especially really high power density power plants, things like nuclear power plants, or you need to put
0:16:46 them outside the US. And I think for the US in particular, you probably don’t want the chips to go outside the US.
0:16:51 Like you probably want as much of it as possible within the US, partially just for economic reasons.
0:17:00 Also, because if you start to use these for higher security reasons, and I think there’s reasonable, like, case to make that you would want to do that.
0:17:12 You really don’t want somebody to walk into a data center with a gun, basically, like, right, the physical security of the data center probably matters quite a lot. And you’re just way more likely to maintain the physical security of a data center.
0:17:21 If it’s on US soil, then if it is not US soil, right. And as these things get bigger, and more critical to the world and the economy,
0:17:30 you don’t want to like leak the weights. And like, a really simple way to leak the weights is to walk into the data center with a gun, or for like a government that
0:17:36 doesn’t, like, have a lot of norms to just sort of say, hey, data center, you need to let us in now.
0:17:47 Or like an intelligence agency could do some kind of operation or whatever, you know, exactly. And, you know, even if it’s a government that today seems like a US ally, maybe tomorrow they aren’t.
0:17:59 And so I would rather have them in the United States than not in the United States. But then also you’ve got the US and China competitiveness concerns. And to solve for that, you need to bring on a lot of power.
0:18:09 Yeah, that’s where it feels like China has a huge advantage, right? It just feels like with their government structure, they will just move away quicker. They’ll just like, yes, we’re doing all this. Yes, yes, yes.
0:18:29 And they don’t have to talk to anyone else. It’s like literally, okay, getting started now. Like, when are we getting started? You know, next week. Yeah, also, the, the work ethic, I think of the culture, at least right now in China, I think people work quite a bit harder, and they’re expected to do a good job. And there’s difference culturally there right now, as well, I think is an advantage.
0:18:34 I will say that San Francisco is probably the beacon for hard work in the country.
0:18:55 Yeah, especially now, like there was a bit where it had slowed down quite a bit, at least when I was there, like when I first got there, it felt like everyone was working really hard. Then it felt like there was like a seven year period where there was, you know, everyone was talking about like work life balance. And I think it was because things had just matured. And like, it was like, okay, Facebook’s won the social media war and this, you know, and Twitter, and people had won their categories.
0:19:22 And it was just not as much competition for some huge thing that they were trying to win, right. But with AI now, it feels like everyone obviously realizes how big the stakes are like, okay, maybe Facebook doesn’t even make sense in the future because of AI, maybe these companies who thought they had won, like Google, maybe they no longer exist if they don’t win this new race. So yeah, I also think that it is a big enough shift for something important enough, that has caused a lot of people to just want to try more.
0:19:33 If you think about, you know, maybe four to five years ago, the last cycle was like a crypto cycle, maybe didn’t have as much like interest for a lot of folks, as in, it’s just distinctly different.
0:20:02 It’s very different. Like I was involved in crypto early on. I bought Bitcoin, you know, pretty early. When I saw Bitcoin, I was like, okay, at first, I wasn’t even sure. Like I came from like the game industry. So I was like, maybe this is just some stupid thing that’s gonna just like fizz out, fizzle out after a few months. But then I started realizing, okay, there’s like real uses in terms of remittances, like sending money to countries where maybe their governments control them using the currency. So that’s the one use case that always made sense to me. But then after that, people made like another 100 use cases that a lot of them, honestly, never made any sense to me.
0:20:20 Yeah, like to steel man the crypto case, like, I think just like being money and being a better version of money is like totally realistic case. And that’s like a really big and important deal. But being money, or like making a really great money, just doesn’t feel as exciting or interesting to a lot of folks as versus like, we’re gonna make robots.
0:20:36 Yeah, and we’re gonna make it possible for you to build a skyscraper for pennies, because we’ve automated the entire system, like we’re building the whole thing, all the sourcing of the materials, we like did all the legal work automated as well. And we’re gonna build that skyscraper over there for like the cost of the electricity. And by the way, we’re also going to bring down the cost of electricity.
0:20:42 By the way, the stuff you’re saying right now, I think like 99% of people have no idea that that kind of stuff is most likely coming.
0:21:06 Yeah, like that’s the sort of pitch of like, what if we just took all the costs of things and brought them to zero, or just the cost of electricity. And I think that’s a much more sort of compelling thing. And then also just the things that people end up working on, they just feel like science fiction, in a way that a lot of other previous trends felt like we’re going to do remote work, or we’re going to make things on mobile phones, or we’re going to make a bunch of characters on a social media app.
0:21:11 And it wasn’t like, we’re gonna cure cancer. Yeah, or we’re gonna make a robot.
0:21:29 Yeah, totally. When I was a kid, I’d read like science magazines and sci fi magazines, everything else. And there was no talk about money. We didn’t think about like, oh, it could be a new money. For most people like who cares, like whatever, but we definitely dreamed about robots. That was one of the biggest ones robots and space travel and yeah, building new kinds of buildings, beautiful cities, using robots, which hopefully does happen.
0:21:48 Yeah, technology used to be technology. It finally is once again, technology is technology. Like technology used to be the spacecraft, and it used to be electricity, which was like actual technology. And I think it once again, actually feels like technology. And so if you are in San Francisco, a lot of people come here because it’s once again, the actual mecca of technology.
0:22:03 Okay, let me tell you about another podcast I know you’re gonna love. It’s called Billion Dollar Moves, hosted by Sarah Chin Spellings. It’s brought to you by the HubSpot Podcast Network, the audio destination for business professionals.
0:22:13 Join venture capitalist and strategist Sarah Chin Spellings as she asks the hard questions and learns through the triumphs, failures, and hard lessons of the creme de la creme.
0:22:28 So you too can make billion dollar moves in venture, in business, and in life. She just did a great episode called The Purpose Driven Power Player Behind Silicon Valley’s Quiet Money with Mike Anders. Listen to Billion Dollar Moves wherever you get your podcasts.
0:22:56 So to get back on how the US beats China. So like, it’s cool that San Francisco is working hard, but that still doesn’t change it. We need a lot more power. You know, talking about robots, I kind of wondered if that’s a possible scenario where America could be China, because we start using robots to produce these nuclear factories and everything else. And maybe if we get ahead of China, but then I see all these videos, at least on X, where people are showing all these videos of robots in China.
0:23:15 I’m like, ha, I thought Elon Musk was probably ahead. Now I’m not so sure. So if China ends up ahead in robots, as well, we’re in trouble. And then me being here in Japan, I keep wondering, like, should I be trying to help in some way to forge some kind of alliance between the US and Japan? Because I think Japan would be a natural ally to help build those robots to be China.
0:23:35 Yeah, I think the US just needs to get rid of its autocracy thing. And this is like mostly a state or city level thing frequently. Like in California, we have CQA, which is basically just like a law that says when you want to build something, you need to prep this document that’s like every possible environmental thing that could ever go wrong, even if it’s not actually a real thing.
0:23:47 Because if you don’t prep a long enough document, you can get sued for not prepping a long enough document. So like you build a bridge or something, and your bridge needs to identify every single possible bird species that could ever have like flown into the bridge or whatever.
0:23:52 Yeah, a lot of that stuff, I’m sure it had good intentions when it started. But now over time, it’s gotten more and more, more ridiculous.
0:24:06 Yeah, it had good intentions. A lot of California and in general, a lot of the US’s problems, I think are good intentions gone wrong. And I think oftentimes people are quick to blame a person when actually just like, a lot of times you just make mistakes.
0:24:25 When you’re saying all that, I’m sitting there thinking of like Hitchhiker’s Guide to the Galaxy, the what is it, the Gorgons or something like that, I think, maybe I’m saying it wrong. You need like a stamp of approval for like almost every single possible thing. It’s like, they can do anything. It’s like, well, here, you need approval for this. Do you have this approval? Do you have this one for this? And it just can be the most basic thing possible, but like you need approval for everything.
0:24:32 Yeah, it feels like that’s what America is drowning in right now. And I’m not sure how you get out of it, because you kind of need approvals to get out of all of it.
0:24:53 Yeah, I think a lot of progress studies folks seem like they’re pushing a lot. That seems very effective. There’s like new ways to organize groups of people. There’s a thing called like sway.app or something. It’s like a way to basically organize a bunch of people to vote on like niche issues, which seems like a useful thing. But yeah, I think you just, you just need to try. You gotta get a lot of people trying.
0:24:56 Cool. I had some selfish questions.
0:24:57 Go for it.
0:25:25 So I think I told you, I’m planning on announcing that like the new direction of lore.com is I’m going to be trying to acquire data centers in collaboration with Jesse Tinsley. Jesse’s this guy who has built up MainStreet.com and Employer.com. He’s basically done a roll up of like 10 amazing like HR recruiting accounting companies. And he’s done a really great job of that. And it’s kind of saturated that market. And we’re looking to kind of take some of the financial engineering he’s learned and apply that to acquiring data centers and operating them.
0:25:46 But we’re still like at the idea phase and figuring out like, okay, what is our actual plan? And we have different ideas. If you were me, what would you be looking at? Would you be looking at like, I don’t know, maybe acquiring crypto miners and repurposing them or like finding places where maybe like the hyperscalers, they’re trying to get so big that they have smaller facilities they want to offload or I don’t know, how would you approach it?
0:25:49 So to be super clear, are you talking about the physical data center?
0:25:49 Yeah, physical.
0:25:54 Also the cluster inside of it. So it’s like the building and then there’s also the chips inside of it.
0:25:59 Probably both. I’m open minded. Do you think it’s, you know, which play do you think is more interesting?
0:26:08 Let’s say you’re doing both the data center and the chips inside. We know more about the chips inside part. The things I would try to do is do everything in your power to reduce your cost of capital.
0:26:26 And if you can own the data center and own as much as your supply chain as you can, because you can compress your cogs by basically removing the margin from the vendor you just paid. So that’s number one. Number two is you need offtake agreements for as long as possible. So you need long-term contracts with somebody who’s going to pay you before you buy the chips.
0:26:39 SF compute comes in or that is true. Like SF computes business. So I joke about this on the internet that nobody knows what SF compute does. What SF compute does is we’re an offtake machine. We generate offtake for people who are setting up clusters. That’s what we do.
0:26:48 And it’s like not at all obvious. And part of the reason it’s not at all obvious is because the way that we generate offtake, the way that we’re able to get people long-term contracts is by running a thing that looks like a cloud provider.
0:26:59 And like no customer who has ever come to us looking at us like a cloud provider ever knows the word offtake or like they don’t care. And so we try to very hammer focus on exactly what those customers care about.
0:27:09 And then on a BD sense, like us just going and talking to people, we pitch, hi, we’re your offtake provider. Like we, we’re the offtake generator. That’s the goal of the company. We’re like a machine that creates offtake.
0:27:15 So are you partnering with a lot of like private equity groups and stuff like that? Like, what are you seeing, you know, who are the main kind of customers there?
0:27:26 So we work with folks who own clusters and there are lots of different reasons why someone might own a cluster, but yes, there are many folks in traditional finance who see compute as an asset class and they would like to invest money in that asset class.
0:27:36 Kind of like what you’re considering as well. Our goal is to partner with folks like that and basically get them offtake, basically saying, hi, if you set up a cluster, we will make sure that you get a contract on it at reasonable prices.
0:27:41 So that way you can finance it correctly and you make money on it. Our goal is to try and help people make money on the clusters.
0:27:46 So another caveat here is that we typically run the clusters. So to some extent, we look like a property manager for GPU clusters.
0:27:52 So that’s the other complexity is we make all the software stacks that a Core Reave or a Lambda might’ve made for themselves.
0:27:56 And then we license it to people or we use it to operate their clusters and sell them. That’s how we work with folks.
0:28:01 So someone set up a cluster, we like go take over the cluster for them and sell it for them and give them offtake.
0:28:07 So they can focus more on treating it like a real estate asset versus having to figure out how to manage it and find all the customers and everything else.
0:28:07 Yeah.
0:28:13 Correct. And then you get a bunch of money and you take that money and then you go and set up another cluster and you just print cash.
0:28:17 That’s the goal. And that’s why we say compute works more like real estate. It doesn’t work like startups.
0:28:21 A lot of people try to do this thing where they set up clusters and then they try to build a great product on top.
0:28:29 And then they realize the customer is so price sensitive that it doesn’t matter what product they build on top of the cluster, the customer isn’t going to give you any markup on your GPUs.
0:28:37 It doesn’t work like CPUs where CPUs, you can buy some CPUs and make a great product and get like a 60 to 80% margin if you want on your CPUs by just building a really great product.
0:28:41 But no one cares about the product. They only care about the cost of the hardware in GPUs.
0:28:45 Yeah, definitely. It sounds like, yeah, it’s more real estate numbers play.
0:28:48 Do the numbers work or not than how to, you know, optimize things as much as possible.
0:28:54 I was talking to some top partners at A16Z. I’m not going to say their names, but they told me the same kind of stuff.
0:28:59 Like they’ve actually worked on with one or two of their companies that are like setting up data centers.
0:29:04 Like the deals are incredibly complicated, you know, equity and different kinds of debt.
0:29:09 Maybe not even one kind of debt, like multiple different kinds with different parties and maybe governments are involved as well.
0:29:13 Yeah, it’s crazy. So I think I get what SF Compute does now.
0:29:18 I’m not entirely sure I could summarize it in like a sentence yet to think about a little bit more.
0:29:18 We sell GPUs.
0:29:22 Yeah. Okay, that makes sense. I think I can remember that.
0:29:26 Where do you see things going like in the next five years with SF Compute?
0:29:29 Like if things are working well and you’re doing great, like where do you see the company?
0:29:32 So we think of ourselves as kind of two companies in one.
0:29:37 One is a systems engineering company that like builds the really low level stack of operating a cluster.
0:29:39 And then the other part is a fintech company.
0:29:46 That’s goal is basically to make financial products that make it easier to financially engineer a cluster and get the cost of capital down quite a lot.
0:29:47 So that’s our general goal.
0:29:49 There’s lots of stuff that we’ve got planned for that.
0:29:52 At the moment, the order book that we built today, that’s our key product right now.
0:29:55 But growing that order book, there’s lots of things that you can do with that.
0:30:06 So you think that data centers could become a huge asset class, just like premium real estate or like commercial real estate buildings, skyscrapers, malls, but they’re a bit more complicated because all the technology.
0:30:08 So you guys kind of simplify that.
0:30:12 So the finance people can treat it more just like a traditional real estate play.
0:30:15 Yeah, I think everybody wants to treat it like soybeans and instead it’s supercomputers.
0:30:16 Right.
0:30:20 So let them treat it like soybeans and you’ll take care of the supercomputer.
0:30:20 Correct.
0:30:24 Our goal is to wrap the supercomputer into something that looks like soybeans.
0:30:32 So that way you can very easily finance it and you can simplify it down enough so that the gears of Wall Street don’t have to do as much complex financial engineering.
0:30:45 And hopefully what this means is that you end up pricing the risk of owning a GPU cluster correctly and you end up with fewer bust and boom cycles and a broadly more stable environment in which lots of capital can flow in and make AI cheaper.
0:30:46 Which is good.
0:30:48 That increases the chance that America will win, which is what I care about.
0:30:50 Yeah.
0:30:51 Correct.
0:30:52 Cool.
0:30:52 Yes.
0:30:55 America’s greatest asset is its depth of capital markets.
0:31:00 And if we put those depth of capital markets to work, we can outspend anyone.
0:31:01 So let’s do it.
0:31:03 You just got to do it very safely.
0:31:04 Okay.
0:31:06 A few fun questions before we go.
0:31:08 And this has been fun and definitely different kind of episode than we normally do.
0:31:10 But just leave with some fun stuff.
0:31:14 So my son here in Japan, you know, he grew up in San Francisco, but he’s in Japan now.
0:31:15 He’s 12 years old.
0:31:20 And I’m constantly thinking about, like, what should I be teaching him so he’ll be successful in the future?
0:31:22 Because things are dramatically changing.
0:31:30 You know, when he was little, maybe he was like two or three, we had some gatherings at Lafayette Park, birthday parties, and some of my friends in tech were asking, like, what do you think he’ll do in the future?
0:31:34 And I talked about, like, anti-drone systems and stuff like this.
0:31:38 And, you know, what would you be teaching if you had a son or a daughter who was like 12?
0:31:41 What would you be teaching them to help them be as successful as possible in the future?
0:31:48 One thing is that I would probably want to make sure that they’re using the various chatbots, the AI bots, to teach themselves or learn.
0:31:52 They’re super effective teachers, just, like, incredibly good at teaching you things.
0:31:52 Yeah.
0:31:54 I would also want to question them a lot.
0:32:01 Like, the ability for them to just say kind of whatever you want them to say is entirely dependent on the person who’s making it.
0:32:05 And that seems like a thing we haven’t really reckoned with super well yet.
0:32:07 And it’s often, like, completely accidental.
0:32:09 Are you talking about the fact that you can lead the LLMs?
0:32:12 Like, if you say it in a certain way that they respond back based on how you said it?
0:32:13 Is that kind of what you’re…
0:32:19 Yeah, like, you’re fighting with your partner and you say, I said this and they said this.
0:32:28 And the way you frame the question encourages the bot to probably not give you the most wise advice, but instead to give you advice that, like, is the one that you would want to hear.
0:32:28 What do you want to hear?
0:32:38 But if you were using the same bot just a moment ago to ask it questions about, like, hey, I, you know, have this pain in my side.
0:32:43 And it says, oh, that’s, you know, this disease and you should go check it out and then you check it out and it’s real.
0:32:44 Like, well, you know what I mean?
0:32:50 Like, if you’ve used it previously to get very accurate information, things that seem really quite correct, because it’s really good at that in particular occasions.
0:32:54 But then you asked it on something else that, you know, maybe a personal life question.
0:33:02 You assume that it will give you a lot of accurate information, but instead it gives you information that it thinks you want to hear because, you know, maybe it’s subjective.
0:33:04 And I think that can lead people astray.
0:33:08 Yeah, I think a lot of people see it like it just answers things, but like, no, it helps you think it through.
0:33:16 And like, so if you need to push back on what it says, too, and like, and it should be a way of, like, probably almost like, you know, extension of your thought process versus just giving you an answer.
0:33:18 And you tell it to do that, because it will totally do that.
0:33:20 Like, if you tell it, hey, be critical of the things I’m saying.
0:33:23 And, you know, don’t totally take my advice either.
0:33:24 It totally can do that.
0:33:24 It’s great.
0:33:26 I think that’s in my custom instructions, I think.
0:33:34 Yeah, so I can imagine that for, like, a 12-year-old, just being more aware of that is probably useful when you’re, like, more impressionable.
0:33:35 Cool.
0:33:35 Makes sense.
0:33:36 That’s kind of what I’ve been trying to do.
0:33:40 I need to do it more, but, yeah, definitely been having him use all the different LLMs.
0:33:44 And even I’ve had him try things like Replit and things like that to create apps, which has been cool.
0:33:47 And also trying to teach him, like, it’s not just, like, a new Google.
0:33:52 Like, you do have to, like, actually, like, push back and talk with it and not just accept exactly what it says is, like, you know, gospel.
0:33:54 One last fun one.
0:33:56 So you get in a time machine.
0:33:57 You step out in San Francisco.
0:33:59 The year 2050.
0:34:00 What do you see?
0:34:01 Is everything the same?
0:34:02 Or is it all different?
0:34:03 Is there robots everywhere?
0:34:05 You know, what’s going on?
0:34:06 Okay, I’ll make a bunch of predictions.
0:34:06 Okay, cool.
0:34:08 First, I think we’re probably going to build a lot.
0:34:13 San Francisco is maybe the first place to stop building and hopefully the first place to start building again.
0:34:16 So I think you’ll probably see more taller buildings and more denser buildings.
0:34:19 I think self-driving cars will probably get rid of parking.
0:34:24 So I think one of the things you might literally notice walking out the street in 2050 is that there isn’t street parking.
0:34:29 I think that will genuinely be, like, a thing that it’s a visual thing that you will notice quite immediately.
0:34:33 I think you probably can walk on the road really safely because of that.
0:34:37 Like, just because if everything is self-driving, one of the things that happens with Waymos is you kind of just, like, step in front of the Waymo.
0:34:40 You don’t exactly, but, like, it’s fine.
0:34:41 It’s fine.
0:34:43 It’s crazy how good they are.
0:34:44 I saw that.
0:34:52 There was a guy, because, like, in Japan, one thing I noticed is in Japan, you know, I don’t know why, but, like, you know, people, you kind of can predict what they’re going to do more.
0:34:54 But, like, in San Francisco, it’s like, man, this is chaotic.
0:34:59 Some guy walked out in the street and stopped and checked his phone and then he turned around, like, right in front of our Waymo.
0:35:02 Yeah, because, like, the Waymo’s not going to hit you.
0:35:04 The Waymo saw you from, like, two miles away.
0:35:07 Maybe in the future, the other Waymo saw that Waymo and it’s, like, sharing data with each other.
0:35:09 Like, they already know where everybody is.
0:35:09 Yeah.
0:35:15 So, my most hopeful prediction of what could literally change, and this might change everywhere, is, do you know what pipe dream is?
0:35:17 Geez, I’ve heard of it.
0:35:18 I’m not sure.
0:35:32 I think it is the most interesting, like, piece of technology that is obvious, possible, and appears to be happening, and I really hope it happens, which is basically underground, like, pipes, and then they have a little robot in it for deliveries.
0:35:32 Oh.
0:35:41 And so what happens is you end up in your house with, like, a dishwasher-shaped thing, like, a literal thing, like a door that you pull out, and then you can get your deliveries there.
0:35:45 And that changes a bunch of stuff about the way that you go about the regular world.
0:35:51 So, for example, you’re ordering, like, Uber Eats or whatever, you don’t, like, send a car to your house for your burrito or whatever.
0:35:56 It just, like, appears in, like, this weird dishwasher thing, and, like, that’s a little thing, and then, like, you just have, like, a burrito.
0:36:00 And you want to order something from Amazon or something, it just, like, appears in your house.
0:36:04 You don’t have to worry about, like, the package getting stolen, it’s just, like, already there.
0:36:05 That’s cool.
0:36:08 So, I know Taizo-san, Masayoshi-san’s brother.
0:36:12 He was an investor in my previous company and we became friends, and he was doing this thing out in Singapore.
0:36:20 He was getting land in Singapore to build, like, a tech, you know, futuristic city, and this was one of the main pilots he was pitching was exactly what you’re talking about now.
0:36:20 Yeah.
0:36:24 It’s, like, the delivery through, you know, pipes and things like this using drones or whatever.
0:36:27 I’m thinking about, like, the things that would physically change in your life.
0:36:32 The other thing is, like, I am super hopeful that all the AI for bio stuff works.
0:36:37 I’m hoping we get just incredible new medications that just cure a whole bunch of diseases and make everything a little bit better.
0:36:41 I have no idea because I have no conception of biotech.
0:36:42 Like, I don’t know anything.
0:36:48 But we have worked with a lot of folks who seem to need GPUs, and I’m so incredibly excited for them to work.
0:36:49 You’re like, oh, a bio customer.
0:36:53 You know, like, you give them a discount, like, you know.
0:36:55 Actually, yeah, we’ve done stuff like that before, yeah.
0:36:57 We used to work with a lot of university labs that were doing stuff like this.
0:37:01 And we were one of the few places where you could get a lot of GPUs for a short period of time because that’s what our product does.
0:37:06 And that is a lot more affordable for a university than having to buy for a year or something like that.
0:37:06 Cool.
0:37:09 Evan, this has been an awesome conversation.
0:37:10 I’ve enjoyed it a lot.
0:37:12 And next time you come out to Japan, we’ll hang out?
0:37:13 Next time.
0:37:13 When I’m in San Francisco.
0:37:15 But where should people find you?
0:37:19 Maybe tell them where to find you on X or your website, what they should check out.
0:37:21 So our website is sfcompute.com.
0:37:25 And you can find me on the internet at EvanJConrad on Twitter or X or whatever it’s called now.
0:37:27 Okay, cool.
0:37:28 Evan, it’s been great.
0:37:29 Thanks for coming on.
0:37:29 Likewise.
The infrastructure behind AI is massive – learn to leverage it with our free guide to advanced prompt engineering: https://clickhubspot.com/fnw
Episode 84: What’s the real infrastructure powering every mind-blowing AI app and model you see today—and are we heading for a “compute bubble”? Nathan Lands (https://x.com/NathanLands) is joined by Evan Conrad (https://x.com/NathanLands), CEO of SF Compute and a leading expert in data centers and AI infrastructure.
Evan, previously an AI audio model founder, now leads SF Compute—building a groundbreaking spot market for AI compute that transforms supercomputers into a tradable, financeable commodity. He’s at the heart of the rapidly growing AI data center industry, shaping how AI is built, scaled, and funded.
This episode pulls back the curtain on the biggest physical build-out in tech history—compute infrastructure. Nathan and Evan break down how AI companies actually get the raw power they need, the economics behind GPU clusters, credit risk bubbles, power constraints in the US vs. China, and why making compute tradable could make or break the future of AI. Whether you’re an investor, founder, or just love tech, this is your crash course in the “invisible” industry driving the AI revolution.
Check out The Next Wave YouTube Channel if you want to see Matt and Nathan on screen: https://lnk.to/thenextwavepd
—
Show Notes:
-
(00:00) AI Cluster Economics Explained
-
(05:59) Venture Capital Dependency Risks
-
(09:34) GPU Cloud Requires Long-Term Contracts
-
(10:23) GPU Contracts and Market Solution
-
(13:57) US vs China: AI Competition
-
(19:33) Crypto vs. Automation Ambitions
-
(22:03) US Plutocracy and Environmental Barriers
-
(26:03) GPU Cluster Investment Management
-
(26:55) Compute as Real Estate Model
-
(29:52) Future Skills for a 12-Year-Old
-
(34:20) Automated In-Home Delivery System
—
Mentions:
-
Evan Conrad: https://www.linkedin.com/in/evan-conrad
-
SF Compute: https://sfcompute.com/
-
Pipedream: https://pipedream.com/
-
Udio: https://www.udio.com/
-
Suno: https://suno.com/
-
Anthropic: https://www.anthropic.com/
-
Waymo: https://waymo.com/
Get the guide to build your own Custom GPT: https://clickhubspot.com/tnw
—
Check Out Matt’s Stuff:
• Future Tools – https://futuretools.beehiiv.com/
• Blog – https://www.mattwolfe.com/
• YouTube- https://www.youtube.com/@mreflow
—
Check Out Nathan’s Stuff:
-
Newsletter: https://news.lore.com/
-
Blog – https://lore.com/
The Next Wave is a HubSpot Original Podcast // Brought to you by Hubspot Media // Production by Darren Clarke // Editing by Ezra Bakker Trupiano
Leave a Reply
You must be logged in to post a comment.