Free with a sign-in.
The best 13 min of 39
The Accidental YC Founder Solving AI’s GPU Crisis | SF Compute’s Evan Conrad
Evan Conrad argues that credit, not GPU supply alone, constrains AI compute—and that a physically settled market can make clusters easier to finance.
Upstarts · Alex Konrad
- Evan Conrad is CEO and co-founder of SF Compute.
The best minutes
- 1:11GPU markets can make compute easier to finance4 min
- 7:15GPU clusters can fail in ways buyers may not expectPro pick 4 min
- 10:53Compute deposits can force startups to raise at huge valuationsPro pick 6 min
Also worth watching
- 28:16GPU futures need to account for performance differences4 min
- 36:34B300 rental rates and workload value may converge soon2 min
- 22:59A $245 million compute deal is small by industry standards3 min
- 34:11Compute financing could become a US geopolitical lever3 min
- 20:25Conrad tried around 40 products while searching for a business3 min
Skip the intro and sponsor reads, the earlier YC and first-product stories, the investor-pitch discussion, and the detailed market roadmap.
Summary · 4 min read
Evan Conrad is CEO and co-founder of SF Compute. He says the company grew out of his first AI lab, which had about $500,000 in the bank and a GPU bill of about $400,000 a month after buying too much capacity.
Conrad argues that AI compute’s key bottleneck is credit: lenders hesitate to finance clusters for young labs, so buyers face large deposits and startups raise at high valuations to cover them. He says SF Compute addresses this by operating clusters owned by partners and selling actual GPU capacity in an independent, physically settled market. He sees that structure as a way to let smaller labs access capacity, while noting that compute is technically nonfungible and hard to commoditize. He also predicts continued scarcity and rising GPU prices, and says data-center restrictions could shift financing and compute capacity across geopolitical lines.
A Pro write-up, open to you here. Try Pro free for 7 days
A Pro write-up. The rest of the summary and the minute by minute are in Real Alpha Pro; the episode, its moments and the transcript stay free.
Read it with Pro, 7 days freeMinute by minute
- 1:11Credit constrains compute access
Conrad says GPU buyers face large deposits because lenders are reluctant to underwrite AI labs’ credit risk. He says SF Compute created an independent, physically settled market so buyers can secure GPU capacity in the future and lenders can assess the asset’s value.
- 4:45SF Compute operates clusters for owners
Conrad describes the company as a property manager for supercomputers: it helps owners design, build and operate clusters, then sells their capacity. Owners include investors and nonprofits; short-term sales can leave capacity available for a portfolio company or grantee that needs it later.
- 7:15Reliable delivery requires operators
Conrad says deployment problems pushed SF Compute from subleasing into operating clusters, including voltage issues and broken InfiniBand networks. He says the company has encountered data centers in chicken coops and flooding during a monsoon, and argues that compute needs an operator to make it a reliable market.
- 10:44Credit drives up startup valuations
Conrad says GPU clouds can make more money by selling long-term capacity to frontier labs, while SF Compute chose a different model. He says deposits can force new AI labs to raise at very high valuations: he recalls initial valuations of $100 million a year or two earlier and about $1 billion at the time, and projects $5 billion, $15 billion and $30 billion.
- 13:29Liquid capacity can arm smaller labs
Conrad says SF Compute wants startups like its former lab, June Lark, to get compute without first raising at a billion-dollar valuation. He contrasts owned GPUs, which allowed xAI to sell capacity, with leases that some labs could not sublease, and says the market should serve more than the largest labs.
- 17:44A therapy exercise led to YC
Conrad says he went to a Y Combinator interview because it frightened him and he was working through cognitive behavioral therapy. His first YC company was Quirk, a mental health app; he later tried a distributed-systems product and, over about four years, around 40 product ideas. He recalls Dalton Caldwell’s advice to keep going and not die.
- 21:15Survival kept the company going
Conrad says SF Compute often had a month of runway and that he did not want to fire the team or let people down. He says he had not initially understood the market’s scale and once thought $5 million was a lot.
- 23:10Conrad calls $245 million modest
Conrad says the company nearly did not announce its $245 million contract because it seemed small by compute-market standards, though it mattered to SF Compute. He says the company is still in survival mode and does not want to become a neocloud that owns and resells most of its hardware. In his view, an independent operator is needed to build a trusted forward market.
- 25:56The market resists a simple pitch
Conrad says the company’s first description was a power grid for compute, but that phrase led people to ask why compute needed a grid. He says investors sometimes mistook SF Compute for a competitor or a simple GPU-matching marketplace, even though operating a physical market involves finance and supercomputing.
- 28:16Compute is not fully fungible
Conrad says compute futures cannot work like wheat because GPU capacity varies by operator and hardware version. He cites a 40% difference in GPU kernel-launch performance between some VM implementations and says early H100s performed worse than later ones. SF Compute’s market delivers GPU time, rather than settling a bet against a price index in cash.
- 31:43Lower deposits depend on a forward market
Conrad says he wants buyers to find available capacity, pay for shorter lead times and face smaller or no deposits. He argues that lenders could finance clusters based on expected future compute prices rather than each buyer’s credit. A liquid market could also let operators replace failed nodes sooner and reduce delays.
- 34:21Compute could become geopolitical leverage
Conrad predicts GPU prices will rise and says supply will remain behind demand for at least four to five years. He argues that credit and revenue streams, not land and power alone, constrain new clusters. He says data-center moratoriums could push compute outside the United States and make capital flows a geopolitical asset.
- 37:24GPU economics favor owners with advantages
Conrad says a B300 could generate $20 to $50 an hour depending on its use, while its price was $5.50 an hour, or perhaps $6. He expects those figures to move closer together. He says buyers with strong credit, power or land—or a portfolio reason to reserve GPUs—should consider owning them; others should probably lease.
Published Oct 29, 2026.
Transcript · 34 min read
Upstarts, hosted by Alex Konrad, with Evan Conrad.
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
A mid-sized deal in Compute is bigger than many venture funds' entire AUM. Finding GPUs is difficult, buying them is expensive, and operating them is risky. San Francisco Compute calls itself the power grid for Compute, and CEO and co-founder Evan Conrad says he has invented a new kind of market for supercomputing, which is even more complicated than it sounds. For like the last 30 years, companies were really simple ideas, and you could easily explain them in like a single sentence. Welcome to fucking Compute. It is not a simple sentence. Today on the podcast, we'll talk about the services that people who own supercomputing hardware really need, how to survive the GPU race one month at a time, and why SF Compute thinks it's important to arm the rebels.
Plus, how to get into Y Combinator without really trying. I went and did a YC interview because it seemed very terrifying. It was like a therapy exercise. This episode is brought to you by Jira by Atlassian, where teams and agents get the context, coordination, and control to move work forward. Try it free at Jira.com. That's J-I-R-A dot com. Evan, welcome to the podcast. Thanks for having me. You are also a Conrad, but spelled the more common way, I guess. It's not a bad way. First double Conrad we've had on the show. I feel like you've got to get other Conrads. I've had other Alex's, but I haven't had other Conrads.
“The reason why you had to place that deposit in the first place is because the lender doesn't want to underwrite your credit risk as an AI lab.”
Read the whole transcript
models, doing a lot of stuff in AI today, and companies big and small want these, are you guys... also sort of building this capacity and operating it? Are you more matching capacity to the companies? How should we be thinking about this stack and where SF Compute fits? Many people in the world have land... our capital, credit, or co-location space, or maybe they went out and bought a big GPU cluster and they would like to become a cloud or they would like to own their own hardware, that's what we do.
We're like a property manager for supercomputers. So we go to folks who want to become a cloud and we help them do that. So for context, our CTO is the co-founder of Voltage Park, a multi-billion dollar GPU cloud. So we've done this before at quite large scale. Well, we help people who want to have GPU clusters, design them, build them, operate them, and then sell them to people. We can work in some capacity like Stripe. So we operate GPU clusters for folks, and then you can put your own website in front of us. And the customer never needs to know that San Francisco Compute exists. We just sell the capacity, and then you can sort of treat us like an API, but we're the ones operating the supercomputer.
Or you can directly go to sfcompute.com and buy from us. The reason why people work with us typically is either, one, they need the offtake. They have some asset. They would like to turn it into a GPU cluster, and they want somebody to buy it. Two, they want the GPUs because they want it for their portfolio. An example is a venture firm or a nonprofit. They're worried about a compute shortage in the future, and so they have us build a cluster for them that they own. And then in the meantime, we're just selling the capacity for them. And when their portfolio, either some non-profit grantee or, you know, a venture firm's portfolio company, starts to fly off and really need the GPUs, we can stop selling the GPUs to whoever we were selling it to because we're only selling them on short-term contracts and then hand them back to the portfolio company or the portfolio project that needed that capacity.
Also, there's just a lot of people who would like a pure financial return. They fully understand that compute is a very big asset class. There's a really high IRR if you do it right. But they don't know how to operate a supercomputer. And so they look for a partner who can operate a supercomputer for them, and we're that partner. I mean, that's a lot of stuff. I guess the initial use case when you're this small startup, you have other GPU compute that you've basically overbought and you need to find takers for it. is that how then the business looked and then you've kind of added more of the full stack or has it been full stack kind of throughout? So what happened when we were subleasing our cluster
“We've seen data centers that were literally chicken coops. We've gotten hit by a monsoon at one point and got flooded during a site visit.”
There are companies trying to build, I mean, there are very highly venture backed, like multi-billion dollar companies and also public companies who are trying to build that extra
“a billion dollar valuation right off the gate in order to pay for a big deposit, which then makes them, makes the whole thing financeable.”
And they're going to sell more of it to the folks who have really great credit. How do you move fast but also stay in control? You have to ship quickly, but you want to ship with confidence. And a lack of context is breaking that balance. Your AI is only as good as the context it has. And for most teams, that context is scattered across tools, threads, and DMs. But you don't have to ship with Blindspots when you use Jira by Atlassian. What's the goal behind a project? What got decided in Slack last week? Atlassian's teamwork graph pulls together across Jira, Confluence, GitHub, Slack, and more.
so nothing falls through the cracks. Teams get better answer quality while also using fewer tokens. And you can share that context with the agents you already use like Claude, Cursor, or GitHub Copilot or connect your own tools through Atlassian MCP. Keep the speed, stay in control, and ship with confidence with Jira by Atlassian. At Upstarts, we document startup journeys that upend the status quo while building great businesses. And iterating on business models that best reflect your product innovations is a big part of doing that. Chargebee is the billing, metering, and monetization infrastructure that enables just that for some of the fastest-scaling AI-native startups.
Charge for active agent minutes, multimodal usage credits, different seat types, or any other custom combination by heading to chargebee.com slash upstarts. That's C-H-A-R-G-E-B-E-E dot com slash upstarts. June Lark was, I believe, audio-focused AI Lab? Yeah. Was this something that you had been interested in for a while, or how did you end up kind of a founder in the first place? I had very severe panic disorder for a while. Like, I had lots of panic attacks. I did a thing called cognitive behavioral therapy, which got over the panic attacks. But one of the things that you do when that happens is you do things that make you afraid.
I wandered in to a coffee shop, and I talked to these folks who were talking about their YC interview, and it sounded really terrifying. And so I went and did a YC interview. No way. Because it seemed very terrifying, and it was like a therapy exercise. I, like, looped in my brother to do it, and I was working on this, like, CBT app at the time. And so I went in and pitched that. We got into YC. I then went back to my— Did you tell them, by the way, I was only doing this as a therapy staff? I think I did later. I don't actually remember if I did. Like, I liked making products. 99% of applicants or whatever to YC never get a callback or never get in.
You do it almost accidentally, it sounds like. I mean, like, I had a thing. But do you think maybe you were more confident in the pitch because you were like, I'm not—my hopes and dreams are not tied to the result here. I would not describe myself either now or before as confident. Well, obviously, you showed something to get into YC, though. I think I just liked making things. Since I was a young kid, I was just making shit all the time. Okay. Yeah. Like what? I would constantly make different types of mobile apps. And then in college, my roommates and I had this consulting company in which we would make stuff for other people.
We made a point-of-sale system for a golf cart where there would be a teenager or something that would come around in a golf course. and take orders for people from the kitchen back at the main spot. Oh, got it. And it would coordinate with... You were DoorDash for golf courses. We were DoorDashes for golf courses, basically. We made DoorDashes for golf courses. That's a really good, extremely good description. Well, I'm just thinking of the YC pitch you didn't have to make for that one. Exactly, yes. So did you go to YC with this idea, or what happened next? Yeah, so we went to YC with Quirk, which was our cognitive behavioral therapy app, and pretty much immediately ran into the problem that every mental health app has, which is like if you do well at all like if your thing works your customers churn And so there's really negative incentives in that.
Kind of like, I guess, dating app challenges as well, right? Similar to dating app challenge, yeah. So we kind of went back to people and said, what should we do? And they said, well, I guess you could try other shit. And so we tried a thing called a room service. A room service was like a distributed systems company. We built the component piece of Figma that makes things multiplayer. We went with this, and then after nine months of this, realized that the people who were most likely to buy it were the people who were most likely to
“And tried like 40 products of different things. Gave up on some of them way too early because somebody told me that they were stupid.”
Could you tell it was going to be a huge market if you guys could survive for, you know, as Because you're going month to month in those early, early moments. I don't think I actually thought that much about the scale in the early times.
“We were almost not going to announce that because I thought it was like, it just isn't that big of a deal in compute.”
Or where do you think San Francisco Compute has been most an upstart in kind of the philosophy or the approach that you're taking? For like the last 30 years, companies were really simple ideas. And you could easily explain them in like a single sentence. Welcome to fucking compute. It is not a simple sentence. Like, I'm sorry, there isn't an elevator pitch. You're going to have to understand finance. You're going to have to understand supercomputing.
It's the fucking supercomputing company. It's not going to be a simple explanation. It's really easy to explain the problem. It's really easy to say, look, everyone sold out, you know, big, long lead times. You have massive deposits. Everyone can see the problem. When we've explained the company to people, most people didn't want to engage because it was a really fucking complicated problem and a really fucking complicated company. And was this investors? Was this employees, potential customers? Who were the people who were shying away? Our early rounds were quite difficult because of this. We would go and try and talk to folks, and if you couldn't explain it in a simple sentence, most people either assumed that we were competitive with something we were not competitive with.
They assumed that they just didn't understand the company. Good portion of that's my fault. There's both the complexity of the problem, and then also you do actually have to be able to explain this to some extent. But some part of the company is just irreducible complexity. If you had to describe the company as much as possible in a sentence today, have you practiced it? Can you give us a shot? With a power grid for compute. I mean, that seems pretty simple to me. Super simple. But it took you a while to get there. No, it was the immediate first sentence we would say. Really? And that just didn't resonate?
The immediate next question is, why would you need a power grid for compute? And if you keep pulling on that thread, you end up realizing that the entire company is extremely complicated. I think most people would like to look at the company and see a traditional software company that has traditional software problems. Or just a very simple marketplace where it's like, okay, there are GPUs over here, everybody wants them, these guys matchmake, they take a spread. simple, right? Yes. Not so simple. It's not so simple. Nothing about the way that you would think you would make a GPU market, nothing that intuitively makes sense works.
And we experienced that really directly because we started with step one and then we just kept going and going and going. So many of the things that we do don't make a lot of sense to people if you just think we're like Airbnb for computer or something.
“you can have a 40% performance difference on GPU kernel launches from one operator to another.”
Your vision of the future, when you guys are at the scale you want to be, what does this market look like and what is SF Compute fully empowering there? We want it to be the case that if you go buy GPUs, there's a price you can buy it at. Right now, that is just straight up not the case.
On anywhere other than SF Compute, almost everybody sold out. On SF Compute, you can buy capacity. Like, that's cool. Step one. Step one, guys. We did it. Then, if you want capacity, and you want a lot of capacity, you can pay more for a lead time that is shorter. On SF Compute, there's a maximum amount of capacity that you can get. You can't go beyond a certain amount because we either don't have that capacity for sale, or maybe we just don't have that capacity total. We want it to be the case that you can pay for a shorter lead time. There's a bunch of different ways that we can make this happen, but that seems like an advantageous thing that should exist.
Then, we want it to be the case that when you go to buy capacity, you don't have to pay as big of a deposit. Either you don't pay any deposit, or you pay very little deposit. The reason why that's possible is that if you know what the price of compute will be in the future, the lender doesn't have to underwrite the credit risk of the customer, they have to underwrite the price of compute. That is a super advantageous thing to exist, and we think that if you make a commercial forwards market, But you can create a future in the future that solves that problem. Once you have the capacity, we want it to come online sooner and be delayed less often.
The way that you make delays not happen is that the cloud was able to go buy the capacity before they had the customer in hand, which they were able to do because the lender was pretty confident on what the price of compute would be in the future. So they knew what it was going to be sold at. When the cloud can place the purchase well before the customer has even seen it, then by the time they're selling the capacity, they already have the hardware, potentially. Or the construction process is very de-risked. And so the delays are way less likely to happen. Right now, this is like financially infeasible for most people.
Once you have gotten the hardware, we want to make it possible for you to have fewer types of failures. And the way you can make fewer types of failures is you have a market that has very liquid. You can pay to swap in a node when another one fails. Right now, what happens is when you get a cluster, you first do a burn-in, which kills a bunch of nodes, and then those nodes that die go through an RMA process, meaning you have to go get new nodes, and sometimes you'll have spares, but not necessarily always, which means you have to wait for more capacity to come online. And you should be able to pay for that capacity to come online now,
“the leverage that the United States has to shift where those capital flows go is going to probably become a geopolitical asset.”
“Right now, if you have a B300, depending on what you're running and what you're doing, you can make somewhere between like $20 to $50 an hour on that B300.”
