Why AI Startups Will Disappear Without Compute Markets — Evan Conrad, San Francisco Compute

So, this is a conversation about So, this is a conversation about computing markets computing markets computing markets . Hmm, but this is . Hmm, but this is . Hmm, but this is more of a conversation about more of a conversation about more of a conversation about why it's worth why it's worth why it's worth creating a market for creating a market for creating a market for computing computing computing technologies. Hmm, if technologies. Hmm, if technologies. Hmm, if you don't know me, I'm you don't know me, I'm you don't know me, I'm Evan. Hmm, we run a Evan. Hmm, we run a Evan. Hmm, we run a company called company called company called San Francisco Compute. Hmm, we San Francisco Compute. Hmm, we San Francisco Compute. Hmm, we are creating a physically are creating a physically are creating a physically settled settled settled forward order market forward order market . Hmm, so if you were . Hmm, so if you were . Hmm, so if you were here for the last here for the last here for the last talk, the last talk, the last talk, the last talk was talk was talk was basically about basically about basically about why all the why all the why all the GPU cluster operators, hmm, GPU cluster operators, hmm, GPU cluster operators, hmm, don't really don't really don't really perform very well. Hmm, perform very well. Hmm, perform very well. Hmm, when we say when we say when we say physically settled physically settled , we mean that , we mean that , we mean that we are the operator. We are we are the operator. We are we are the operator. We are both an both an both an equipment operator, just as the equipment operator, just as the equipment operator, just as the cloud is an cloud is an cloud is an equipment operator, and we equipment operator, and we equipment operator, and we are also a are also a are also a marketplace at the same time. It's kind of like a marketplace at the same time. It's kind of like a marketplace at the same time. It's kind of like a grid for grid for grid for calculations. So the calculations. So the calculations. So the general idea is general idea is general idea is that in order to that in order to that in order to run run run and operate a and operate a and operate a supercomputer, um, supercomputer, um, supercomputer, um, we think we think we think the problems start the problems start the problems start right at the design stage right at the design stage of the supercomputer, um, of the supercomputer, um, right at the design stage of the right at the design stage of the bill of materials, bill of materials, um. This is also a control um. This is also a control um. This is also a control over the unifying over the unifying over the unifying panel; it is also panel; it is also panel; it is also the control of, ahem, the control of, ahem, the control of, ahem, "bare "bare "bare metal" orchestrations; it also includes metal" orchestrations; it also includes metal" orchestrations; it also includes services on top, services on top, services on top, virtual machines, virtual machines, virtual machines, etc. Hmm, there are many etc. Hmm, there are many etc. Hmm, there are many different parts of the stack different parts of the stack different parts of the stack that need to be done that need to be done that need to be done efficiently; otherwise, um, efficiently; otherwise, um, efficiently; otherwise, um, people who have a people who have a people who have a data center, um, and data center, um, and data center, um, and would like to make more money from would like to make more money from would like to make more money from that data center that data center by becoming a GPU cloud, um, by becoming a GPU cloud, um, will ultimately will ultimately will ultimately fail for the customer.

fail for the customer. fail for the customer. This is what we do. We This is what we do. We This is what we do. We offer what's offer what's offer what's called called called supercomputing as a supercomputing as a supercomputing as a service. We service. We service. We are reaching out to people are reaching out to people are reaching out to people who want to become a who want to become a who want to become a cloud. It's like people cloud. It's like people cloud. It's like people who want to become what who want to become what who want to become what Nvidia would call an NCP, hmm Nvidia would call an NCP, hmm , or venture capital firms , or venture capital firms , or venture capital firms who want to have who want to have who want to have clusters for their clusters for their clusters for their portfolio, hmm, and want to portfolio, hmm, and want to portfolio, hmm, and want to be able to be able to be able to redeploy that redeploy that redeploy that capacity when a capacity when a capacity when a company like Instinct company like Instinct company like Instinct starts to grow, starts to grow, starts to grow, hmm, or AI companies hmm, or AI companies who would like to get the who would like to get the residual value of residual value of residual value of the compute because the compute because the compute because they've realized that it's they've realized that it's they've realized that it's really important to really important to really important to their bottom line, hmm, their bottom line, hmm, their bottom line, hmm, but don't want to but don't want to but don't want to run run run a supercomputer. a supercomputer. a supercomputer. So, those are physical So, those are physical So, those are physical calculations, and then calculations, and then calculations, and then in addition to that, in addition to that, in addition to that, we keep an we keep an we keep an order book for order book for order book for capacity that you capacity that you capacity that you can purchase and can purchase and can purchase and reserve now reserve now reserve now or in the future. Hmm, or in the future. Hmm, or in the future. Hmm, this, um, together, is not at all what this, um, together, is not at all what this, um, together, is not at all what we thought we would we thought we would we thought we would do. do. do. We were originally a company We were originally a company We were originally a company called Junlark that called Junlark that called Junlark that was trying to create was trying to create was trying to create something like Sunno or something like Sunno or something like Sunno or UDIO, but before Sunno or UDIO... UDIO, but before Sunno or UDIO... Hmm, at that time, hmm, to Hmm, at that time, hmm, to Hmm, at that time, hmm, to actually launch actually launch actually launch a lab, you had to a lab, you had to raise a big raise a big deposit, a huge one, deposit, a huge one, deposit, a huge one, which at that time which at that time which at that time was $10 million was $10 million was $10 million . Hmm, we thought . Hmm, we thought that was an insane amount of that was an insane amount of that was an insane amount of money. Hmm, and so we money. Hmm, and so we money. Hmm, and so we thought: no, this is thought: no, this is thought: no, this is stupid. We're stupid. We're stupid. We're going to raise 500 grand going to raise 500 grand going to raise 500 grand , and obviously we're , and obviously we're , and obviously we're going to going to going to rent the rent the rent the space from someone space from someone space from someone else. Hmm, obviously else. Hmm, obviously else. Hmm, obviously today, hmm, it's not today, hmm, it's not today, hmm, it's not $10 million, it's a $10 million, it's a $10 million, it's a billion dollars. And, um billion dollars. And, um , nevertheless, um, we , nevertheless, um, we , nevertheless, um, we raised our 500 thousand.

raised our 500 thousand. raised our 500 thousand. We buy a large We buy a large We buy a large cluster and then cluster and then cluster and then work work work on our cluster every month. on our cluster every month. on our cluster every month. Um, and if we hadn't Um, and if we hadn't subleased our cluster, we would have gone subleased our cluster, we would have gone bankrupt. Um, bankrupt. Um, bankrupt. Um, if you've heard of if you've heard of if you've heard of GPU brokers, we were the GPU brokers, we were the GPU brokers, we were the first GPU broker because first GPU broker because first GPU broker because we were forced to be a we were forced to be a we were forced to be a GPU broker. So for the GPU broker. So for the GPU broker. So for the first year of the company, first year of the company, first year of the company, we were just trying we were just trying we were just trying to sell our cluster, to sell our cluster, to sell our cluster, otherwise we would have gone otherwise we would have gone otherwise we would have gone bankrupt. Um, bankrupt. Um, bankrupt. Um, so I was hanging around with these so I was hanging around with these so I was hanging around with these real, um, real, um, real, um, pieces of paper in Nurups that pieces of paper in Nurups that said H100 and A100 at the same time, because that's said H100 and A100 at the same time, because that's how I made, um, how I made, um, how I made, um, I made the, I made the, I made the, um, poster so quickly. Um, and so um, poster so quickly. Um, and so um, poster so quickly. Um, and so this conversation is this conversation is this conversation is n't really necessarily about, n't really necessarily about, n't really necessarily about, um, how the computing market works um, how the computing market works . It's about how the . It's about how the computing market computing market computing market , um, emerges. , um, emerges. , um, emerges. Um, we like to Um, we like to Um, we like to think of it as if there's think of it as if there's think of it as if there's this brochure, and this brochure, and this brochure, and on the front of the on the front of the on the front of the brochure it says that brochure it says that brochure it says that building Neolab is a building Neolab is a building Neolab is a nightmare. There are all these different nightmare. There are all these different nightmare. There are all these different financial and financial and financial and economic issues, economic issues, economic issues, and on the back of the brochure it and on the back of the brochure it and on the back of the brochure it says these are says these are computing markets. It's not entirely computing markets. It's not entirely clear how one thing clear how one thing clear how one thing gets to the other gets to the other gets to the other until you open the until you open the until you open the brochure, and then, oh brochure, and then, oh brochure, and then, oh my god, there's so much in this my god, there's so much in this my god, there's so much in this brochure brochure brochure , and it all , and it all , and it all goes on and on, and goes on and on, and goes on and on, and so many folds. This is a so many folds. This is a so many folds. This is a very complex and very complex and very complex and unpredictable unpredictable unpredictable problem because problem because two surprising two surprising problems need to be solved to get it problems need to be solved to get it problems need to be solved to get it working. You have to working. You have to working. You have to create a market, which, ahem create a market, which, ahem , is a uniquely , is a uniquely , is a uniquely difficult problem difficult problem difficult problem because there's the problem of a because there's the problem of a because there's the problem of a cold start. You cold start. You cold start. You also need to also need to also need to build a cloud and build a cloud and build a cloud and then not become the cloud then not become the cloud . Hmm, that's a whole . Hmm, that's a whole . Hmm, that's a whole different challenge. But different challenge. But different challenge. But these are problems that these are problems that labs face, and these labs face, and these will be problems that will be problems that will be problems that everyone will face.

everyone will face. everyone will face. So, if you work So, if you work So, if you work at a startup, these are your at a startup, these are your at a startup, these are your problems right now. Hmm, problems right now. Hmm, problems right now. Hmm, everyone's everyone's everyone's running out of running out of running out of computing computing computing resources now. Hmm, that, um, resources now. Hmm, that, um, resources now. Hmm, that, um, probably won't probably won't probably won't change for a change for a change for a while. Hmm, that means while. Hmm, that means while. Hmm, that means you'll have to you'll have to you'll have to buy a new cluster, buy a new cluster, buy a new cluster, brand new, and there's a brand new, and there's a brand new, and there's a lead time. lead time. lead time. You do not have the ability You do not have the ability You do not have the ability to pay to launch to pay to launch to pay to launch this cluster now. this cluster now. this cluster now. You will pay for its You will pay for its You will pay for its launch later. launch later. launch later. There is nothing on the market right now There is nothing on the market right now There is nothing on the market right now that that that can actually launch this can actually launch this can actually launch this cluster online for cluster online for cluster online for you. you. In the meantime, you will pay your entire bank balance as a deposit. The reason In the meantime, you will pay your entire bank balance as a deposit. The reason you pay you pay you pay your entire your entire your entire bank balance bank balance bank balance as a deposit is as a deposit is as a deposit is because the lender because the lender because the lender has no idea how to has no idea how to has no idea how to guarantee AGI. Um, how guarantee AGI. Um, how guarantee AGI. Um, how can you can you can you do that? You have two do that? You have two do that? You have two startups in San startups in San Francisco, or Francisco, or Francisco, or maybe in New York, maybe in New York, maybe in New York, that say, "I'm that say, "I'm that say, "I'm going to build a going to build a going to build a superintelligence." superintelligence." superintelligence." No lender No lender No lender knows how to guarantee knows how to guarantee knows how to guarantee superintelligence. Um, superintelligence. Um, superintelligence. Um, so they have no so they have no so they have no idea how much the idea how much the idea how much the asset is worth. Um, in the meantime, asset is worth. Um, in the meantime, asset is worth. Um, in the meantime, your cluster your cluster your cluster will most likely be will most likely be will most likely be delayed. This is because delayed. This is because delayed. This is because the lender needed to the lender needed to the lender needed to guarantee the guarantee the guarantee the credit risk credit risk credit risk of the customer, so they of the customer, so they of the customer, so they needed to know needed to know needed to know who the customer was before who the customer was before who the customer was before you deployed you deployed you deployed the cluster. So the the cluster. So the the cluster. So the moment moment moment the client signs is the client signs is the client signs is the riskiest the riskiest the riskiest moment in the entire moment in the entire moment in the entire transaction. Um, you transaction. Um, you transaction. Um, you wanted to first wanted to first wanted to first create a cluster create a cluster create a cluster and then sell it.

and then sell it. and then sell it. This also helps This also helps This also helps reduce all sorts of reduce all sorts of reduce all sorts of different types of problems, different types of problems, different types of problems, um, when clusters um, when clusters um, when clusters break. Meanwhile, break. Meanwhile, break. Meanwhile, most of the people who most of the people who most of the people who create clusters create clusters create clusters are people who have an are people who have an are people who have an actual asset that actual asset that actual asset that makes the thing cheap. makes the thing cheap. makes the thing cheap. You already have space in the You already have space in the You already have space in the data center. data center. data center. What they don't What they don't What they don't get is get is get is expertise in expertise in expertise in supercomputers, as supercomputers, as supercomputers, as we saw in the we saw in the we saw in the previous report, previous report, previous report, if you, um, if you, um, if you, um, watched it. Meanwhile, watched it. Meanwhile, watched it. Meanwhile, you will face the you will face the you will face the same problem as Junark, same problem as Junark, same problem as Junark, namely: you will either namely: you will either namely: you will either overestimate or overestimate or overestimate or underestimate underestimate . Hmm, that's because AI doesn't . Hmm, that's because AI doesn't . Hmm, that's because AI doesn't have margins like have margins like have margins like SAS companies. Therefore, SAS companies. Therefore, SAS companies. Therefore, you don't have much you don't have much you don't have much margin for error. So margin for error. So margin for error. So the chances that you'll either the chances that you'll either the chances that you'll either outperform and someone outperform and someone outperform and someone else wins because else wins because else wins because they did it they did it they did it right, or you'll right, or you'll right, or you'll overestimate and then overestimate and then overestimate and then just become a just become a just become a stable AI, or you'll stable AI, or you'll stable AI, or you'll explode, you'll, um, explode, you'll, um, explode, you'll, um, become a character become a character become a character AI, um, are very high. AI, um, are very high. AI, um, are very high. So, it's all the same So, it's all the same So, it's all the same problem. These are problem. These are problem. These are just different types of just different types of just different types of risks. This is technical risks. This is technical risks. This is technical risk, financial risk, financial risk, financial risk, credit risk, credit risk, credit risk. Hmm, it's all one risk. Hmm, it's all one risk. Hmm, it's all one big combined big combined big combined risk problem. Uh, that's the risk problem. Uh, that's the risk problem. Uh, that's the problem that we're problem that we're problem that we're ultimately solving, and ultimately solving, and ultimately solving, and it's also the problem it's also the problem it's also the problem that we think the that we think the computing markets in general, the entire computing markets in general, the entire industry, are solving. So we're industry, are solving. So we're industry, are solving. So we're going to look at each of going to look at each of going to look at each of these points and these points and these points and see how that see how that see how that plays into the plays into the computing market, because if you computing market, because if you don't understand that, don't understand that, don't understand that, it's not entirely it's not entirely it's not entirely clear why you clear why you clear why you need a need a computing market. And it almost computing market. And it almost seems like the kind of seems like the kind of seems like the kind of problem that someone problem that someone problem that someone thinks, “Wow. "Computer markets thinks, “Wow. "Computer markets are cool, are cool, so I'm going to so I'm going to so I'm going to create one." We create one." We create one." We believe that they believe that they believe that they should stem from should stem from should stem from real problems that real problems that people face, namely: first, people face, namely: first, everyone is sold out. Why are everyone is sold out. Why are everyone is sold out. Why are they sold? Because they sold? Because they sold? Because the lender couldn't the lender couldn't the lender couldn't finance the cluster finance the cluster finance the cluster unless you had a unless you had a unless you had a long-term long-term long-term purchase and sale agreement.

purchase and sale agreement. purchase and sale agreement. This was not the This was not the This was not the case with processors, because the buyers of case with processors, because the buyers of case with processors, because the buyers of processors are processors are processors are SAS companies. SAS companies. SAS companies. SAS companies have high margins. SAS companies have high margins. SAS companies have high margins. And because they have And because they have And because they have high margins, you high margins, you high margins, you can create a can create a can create a product that product that product that rewards you for having a rewards you for having a rewards you for having a great great great processor that processor that processor that rewards you with rewards you with rewards you with high margins on high margins on high margins on your processors. your processors. Think about it this way. Think about it this way. If you're a SAS company and If you're a SAS company and If you're a SAS company and your underlying cloud your underlying cloud your underlying cloud doubles in price, you won't doubles in price, you won't doubles in price, you won't go bankrupt. You go bankrupt. You go bankrupt. You hate your hate your hate your cloud, you're angry, but you're cloud, you're angry, but you're cloud, you're angry, but you're not going bankrupt. If not going bankrupt. If not going bankrupt. If you're an you're an you're an AI company and AI company and AI company and your underlying cloud your underlying cloud your underlying cloud doubles in price, you're going doubles in price, you're going doubles in price, you're going to go bankrupt unless you to go bankrupt unless you to go bankrupt unless you move to some move to some move to some other underlying cloud. other underlying cloud. other underlying cloud. And that means you are And that means you are And that means you are actually actually actually price sensitive. And since you're price sensitive. And since you're price sensitive. And since you're actually actually actually price sensitive, um, and you're not price sensitive, um, and you're not price sensitive, um, and you're not held in any held in any held in any particular cloud. So particular cloud. So particular cloud. So you can't you can't you can't sell on demand, sell on demand, sell on demand, because your customers because your customers because your customers will immediately go to will immediately go to whoever is cheaper. And that whoever is cheaper. And that whoever is cheaper. And that means that to means that to means that to finance finance finance GPU clusters, you GPU clusters, you GPU clusters, you need to have need to have need to have long-term long-term long-term take-and- take-and- pay contracts, hopefully pay contracts, hopefully pay contracts, hopefully with someone who has an with someone who has an with someone who has an investment-grade investment-grade investment-grade credit rating. credit rating. credit rating. However, if you However, if you However, if you don't have liquidity behind don't have liquidity behind don't have liquidity behind them, if you them, if you them, if you sell sell sell reserved reserved reserved resources and your customer resources and your customer resources and your customer can't get out, it can't get out, it can't get out, it also means that also means that also means that some other customer some other customer some other customer who wants who wants who wants capacity right capacity right capacity right now because they now because they now because they instinctively do and their instinctively do and their instinctively do and their product is gaining product is gaining product is gaining popularity can't popularity can't popularity can't pay for anything.

pay for anything. pay for anything. They have to They have to They have to fund a new fund a new fund a new cluster that has a cluster that has a cluster that has a certain certain certain implementation time because the implementation time because the implementation time because the physical world just physical world just physical world just exists and it will take exists and it will take exists and it will take some time. Sometimes you some time. Sometimes you some time. Sometimes you even have to even have to even have to build that damn build that damn build that damn data center. data center. data center. Um, and then since Um, and then since Um, and then since you're funding a new you're funding a new you're funding a new cluster, you'll cluster, you'll cluster, you'll have to pay have to pay have to pay a deposit because a deposit because a deposit because almost no one in AI almost no one in AI almost no one in AI has a good has a good has a good credit history. And credit history. And credit history. And since this is a since this is a since this is a completely new completely new completely new technology, all of these technology, all of these technology, all of these companies are new. companies are new. companies are new. Meanwhile, even as you, Meanwhile, even as you, Meanwhile, even as you, um, start to um, start to um, start to scale up those scale up those scale up those ladders, you're still going to ladders, you're still going to ladders, you're still going to run into run into credit issues all along the credit issues all along the way. This is the way. This is the way. This is the number number number one bottleneck for AI at the one bottleneck for AI at the one bottleneck for AI at the moment. This is not moment. This is not moment. This is not power. You power. You power. You get credit get credit get credit long before you long before you long before you reach AI, or reach AI, or reach AI, or long before you long before you long before you reach, um, uh, reach, um, uh, reach, um, uh, power. So, um, power. So, um, power. So, um, that means that every that means that every that means that every company that company that company that does logical does logical does logical inference, um, uh, inference, um, uh, inference, um, uh, every Neol, is limited by every Neol, is limited by every Neol, is limited by the amount of cash in their the amount of cash in their the amount of cash in their bank account bank account . And so they . And so they . And so they go back to the go back to the go back to the venture venture venture capitalist who capitalist who capitalist who says, "Oh, I'm going to says, "Oh, I'm going to says, "Oh, I'm going to fund at the fund at the fund at the model level, the model level, the model level, the application level, the application level, the application level, the inference level, inference level, inference level, because I don't want to because I don't want to because I don't want to fund all this fund all this fund all this capital expenditure." capital expenditure." capital expenditure." Hmm, and then the venture Hmm, and then the venture Hmm, and then the venture capitalist funds capitalist funds capitalist funds all the capital all the capital all the capital expenditure by giving, um, the expenditure by giving, um, the expenditure by giving, um, the new Neolab or the new new Neolab or the new new Neolab or the new inference company inference company inference company , um, a check, , um, a check, , um, a check, which they then which they then which they then send to the cloud, send to the cloud, send to the cloud, um, that's a down payment for the um, that's a down payment for the um, that's a down payment for the cluster, and that's how cluster, and that's how cluster, and that's how you secure you secure you secure the customer. So, um, the customer. So, um, the customer. So, um, the clusters then the clusters then the clusters then get delayed because get delayed because , um, the cloud couldn't , um, the cloud couldn't , um, the cloud couldn't actually deploy actually deploy actually deploy the cluster until it had a the cluster until it had a the cluster until it had a customer. And that was the customer. And that was the customer. And that was the riskiest riskiest riskiest moment. So, hmm, what would have moment. So, hmm, what would have moment. So, hmm, what would have happened is that the cloud happened is that the cloud happened is that the cloud bought the cluster bought the cluster bought the cluster in advance and then in advance and then in advance and then waited for it to waited for it to waited for it to arrive at the arrive at the arrive at the data center and data center and data center and then sold it. And then sold it. And then sold it. And this means that you are

this means that you are this means that you are much less much less much less likely to experience delays likely to experience delays . Meanwhile, the cluster is . Meanwhile, the cluster is . Meanwhile, the cluster is very likely to very likely to very likely to break when you break when you break when you get the cluster as a get the cluster as a get the cluster as a client because you have client because you have client because you have no liquidity. And no liquidity. And no liquidity. And so if the so if the so if the failure rate is higher than you failure rate is higher than you failure rate is higher than you expect and you expect and you expect and you don't have the ability don't have the ability don't have the ability to migrate a node, then to migrate a node, then to migrate a node, then you don't have anywhere to you don't have anywhere to you don't have anywhere to pull nodes from, which pull nodes from, which pull nodes from, which means if you means if you means if you need new nodes need new nodes need new nodes to make up for the to make up for the to make up for the failed ones, you'll failed ones, you'll failed ones, you'll have to RMA them, which have to RMA them, which have to RMA them, which means you'll have to means you'll have to means you'll have to get a get a get a new node. And this new node. And this new node. And this brings us back to brings us back to brings us back to a few previous a few previous a few previous steps, where the runtime steps, where the runtime steps, where the runtime issue arises again issue arises again issue arises again . And . And . And that's why everything keeps that's why everything keeps that's why everything keeps breaking down for everyone. breaking down for everyone. breaking down for everyone. Once you, as a new Once you, as a new Once you, as a new AI lab AI lab AI lab , get , get , get through all of this, your through all of this, your through all of this, your problems are that problems are that problems are that you know you you know you you know you sold out no one, just like you sold out sold out no one, just like you sold out sold out no one, just like you sold out everyone. Then you everyone. Then you everyone. Then you had to wait for a had to wait for a had to wait for a certain lead time, certain lead time, certain lead time, you paid a large you paid a large you paid a large deposit, the thing deposit, the thing deposit, the thing was delayed, and then was delayed, and then was delayed, and then it broke, and it broke, and it broke, and finally you finally you finally you went through all of that. went through all of that. went through all of that. The likelihood that The likelihood that The likelihood that you will correctly you will correctly you will correctly predict your predict your predict your demand is very low. demand is very low. demand is very low. So you will either be So you will either be So you will either be over-supplied over-supplied over-supplied or under- or under- or under- supplied, and you have supplied, and you have supplied, and you have no way out. There is no no way out. There is no no way out. There is no way to sell way to sell way to sell your capacity to your capacity to your capacity to someone else, and someone else, and someone else, and there is no way to there is no way to there is no way to buy more buy more buy more capacity if capacity if capacity if you need it. you need it. you need it. So, computer So, computer So, computer markets, in order to solve markets, in order to solve markets, in order to solve these problems, have to these problems, have to these problems, have to solve a bunch of different solve a bunch of different solve a bunch of different problems at once.

problems at once. problems at once. You need to You need to You need to guarantee the value of the guarantee the value of the guarantee the value of the GPU cluster GPU cluster GPU cluster , not , not , not guarantee the value guarantee the value guarantee the value or credit risk or credit risk or credit risk of the customer. Here's how it of the customer. Here's how it of the customer. Here's how it might work. There is no might work. There is no might work. There is no guarantee that this guarantee that this guarantee that this will work. What happened in other will work. What happened in other will work. What happened in other industries is that industries is that industries is that everything just everything just everything just consolidated. You consolidated. You consolidated. You have one big have one big have one big creditworthy creditworthy creditworthy client that client that client that handles the whole handles the whole handles the whole damn industry. This is exactly where the damn industry. This is exactly where the development of development of artificial intelligence is currently heading. artificial intelligence is currently heading. artificial intelligence is currently heading. Instead of Instead of Instead of being, ahem, a liquid being, ahem, a liquid being, ahem, a liquid market where anyone can market where anyone can market where anyone can participate, it participate, it participate, it will look like TSMC. will look like TSMC. will look like TSMC. You will have one You will have one You will have one buyer, hmm, or buyer, hmm, or buyer, hmm, or several buyers who several buyers who several buyers who will become clouds. will become clouds. will become clouds. You will have several You will have several You will have several model model model labs that are labs that are artificial intelligence labs. And artificial intelligence labs. And then you won't have any then you won't have any then you won't have any other startups. Hmm, other startups. Hmm, other startups. Hmm, we think that's we think that's we think that's going to happen. And that's why going to happen. And that's why going to happen. And that's why 3 years ago, when 3 years ago, when 3 years ago, when we were creating our we were creating our we were creating our artificial artificial artificial intelligence lab, we had a intelligence lab, we had a intelligence lab, we had a chance. We could chance. We could chance. We could say, “No, we’re not say, “No, we’re not say, “No, we’re not going to do going to do going to do our new, um, we’re not our new, um, we’re not our new, um, we’re not going to, you know, going to, you know, going to, you know, endure a few years endure a few years endure a few years just trying to just trying to just trying to keep our keep our keep our company alive. "Hmm, we're company alive. "Hmm, we're company alive. "Hmm, we're just going to just going to just going to create a new create a new create a new artificial artificial artificial intelligence lab." But we intelligence lab." But we intelligence lab." But we saw this problem saw this problem saw this problem and said, "Oh, [ __ ], and said, "Oh, [ __ ], and said, "Oh, [ __ ], if we don't do anything about it if we don't do anything about it if we don't do anything about it , we're , we're , we're just going to be stuck with just going to be stuck with just going to be stuck with this in 3 years." And this in 3 years." And this in 3 years." And now, 3 years later, we are now, 3 years later, we are now, 3 years later, we are facing this facing this facing this problem, and everyone will problem, and everyone will problem, and everyone will face it. So, face it. So, face it. So, as if you were watching the as if you were watching the as if you were watching the previous conversation, hmm previous conversation, hmm , it's very difficult to solve , it's very difficult to solve , it's very difficult to solve this problem unless you this problem unless you this problem unless you manage both manage both manage both the market and the market and the market and the clusters at the same time. So, this is the clusters at the same time. So, this is the clusters at the same time. So, this is Dylan Page. Dylan is Dylan Page. Dylan is Dylan Page. Dylan is the co-chair of the the co-chair of the the co-chair of the CNCF Infrastructure Group. This is the CNCF Infrastructure Group. This is the CNCF Infrastructure Group. This is the organization that organization that organization that develops Kubernetes, I develops Kubernetes, I develops Kubernetes, I previously worked at Lambda previously worked at Lambda , working on , working on , working on infrastructure there. Hmm, infrastructure there. Hmm, infrastructure there. Hmm, and the reason, and he and the reason, and he and the reason, and he joined San Francisco

joined San Francisco joined San Francisco Compute, is because, hmm, Compute, is because, hmm, Compute, is because, hmm, GPU computing is GPU computing is GPU computing is not not not standardized and standardized and standardized and reliable enough for the reliable enough for the reliable enough for the market operator to be market operator to be market operator to be separated from the separated from the separated from the physical physical physical infrastructure. This doesn't infrastructure. This doesn't infrastructure. This doesn't work if you're not the one work if you're not the one buying the thing. buying the thing. buying the thing. Sorry, this doesn't work Sorry, this doesn't work Sorry, this doesn't work if the person you're if the person you're if the person you're buying from doesn't manage it. buying from doesn't manage it. buying from doesn't manage it. To give you an example of To give you an example of To give you an example of why this doesn't why this doesn't why this doesn't work, I'm sure work, I'm sure work, I'm sure some of you some of you some of you are flying in from are flying in from are flying in from overseas or, sorry, overseas or, sorry, overseas or, sorry, um, from another city, um, um, from another city, um, um, from another city, um, and you have a hotel and you have a hotel and you have a hotel room. Hmm, wouldn't it be room. Hmm, wouldn't it be room. Hmm, wouldn't it be great if you great if you great if you could, you know, could, you know, could, you know, sell this sell this sell this hotel room to hotel room to hotel room to someone else. I'm coming someone else. I'm coming someone else. I'm coming to you and I want to buy to you and I want to buy to you and I want to buy that hotel room. that hotel room. that hotel room. Um, and I say, “Can Um, and I say, “Can Um, and I say, “Can I order I order room service?” Can room service?” Can I have my I have my I have my room cleaned? Um, I'm having a room cleaned? Um, I'm having a room cleaned? Um, I'm having a problem with my problem with my problem with my number. Who should I number. Who should I number. Who should I call? Um, if you call? Um, if you call? Um, if you don't run a hotel, don't run a hotel, don't run a hotel, you can't solve you can't solve you can't solve any of these problems. any of these problems. any of these problems. This is why This is why This is why cluster brokerage doesn't work. cluster brokerage doesn't work. cluster brokerage doesn't work. You should manage You should manage You should manage the cluster like a hotel the cluster like a hotel . You have to run this . You have to run this . You have to run this whole damn whole damn whole damn enterprise. enterprise. enterprise. Otherwise, you Otherwise, you Otherwise, you won't really be able to create a won't really be able to create a won't really be able to create a market. This doesn't work. market. This doesn't work. market. This doesn't work. You cannot You cannot You cannot perform physical perform physical perform physical calculations. Um, so calculations. Um, so calculations. Um, so why doesn't anyone why doesn't anyone why doesn't anyone do the physical do the physical do the physical calculations? Um, calculations? Um, calculations? Um, the problem is that there is a the problem is that there is a the problem is that there is a very strong incentive to very strong incentive to very strong incentive to never do never do never do physical calculations.

physical calculations. physical calculations. Um, that's because if you're Um, that's because if you're Um, that's because if you're building physical building physical building physical compute for compute for compute for compute, that compute, that compute, that literally means literally means literally means you're building all the you're building all the you're building all the operating operating operating power of the GPU cloud. And power of the GPU cloud. And power of the GPU cloud. And if you're building all the if you're building all the GPU cloud processing power, GPU cloud processing power, you should just be a you should just be a you should just be a GPU cloud because it GPU cloud because it GPU cloud because it looks like you're going to make an looks like you're going to make an looks like you're going to make an incredible amount of incredible amount of incredible amount of money. Um, you'll money. Um, you'll money. Um, you'll be signing be signing be signing billion-dollar deal after billion-dollar deal after billion-dollar deal after billion-dollar deal. billion-dollar deal. billion-dollar deal. So, to hell with everything. So, to hell with everything. So, to hell with everything. Let's not Let's not Let's not create GPUs, create GPUs, create GPUs, or let's...not or let's...not or let's...not create a market for create a market for create a market for computing computing computing technologies. Let's technologies. Let's technologies. Let's just just just go GPU cloud instead, which is go GPU cloud instead, which is go GPU cloud instead, which is basically what everyone is doing basically what everyone is doing basically what everyone is doing , because the numbers are , because the numbers are , because the numbers are astronomical and it makes a astronomical and it makes a astronomical and it makes a lot of sense. Um, lot of sense. Um, lot of sense. Um, the reason we the reason we the reason we ended up creating a ended up creating a ended up creating a physical settlement was because physical settlement was because physical settlement was because we we we knew that if we did knew that if we did knew that if we did n't start that, then, um, we would n't start that, then, um, we would n't start that, then, um, we would n't be able to launch the n't be able to launch the n't be able to launch the AI ​​lab AI ​​lab AI ​​lab , which was what , which was what , which was what we wanted to we wanted to we wanted to do from the do from the do from the beginning. We wanted to beginning. We wanted to beginning. We wanted to create a laboratory create a laboratory . We couldn't create . We couldn't create . We couldn't create a lab if a lab if a lab if it didn't exist. So it didn't exist. So it didn't exist. So when we started, it when we started, it when we started, it was 10 million, um, was 10 million, um, was 10 million, um, as a deposit that as a deposit that as a deposit that had to be had to be had to be paid, um, to paid, um, to paid, um, to launch a new launch a new launch a new lab. This is similar lab. This is similar lab. This is similar to how to overcome to how to overcome to how to overcome the problem of credit the problem of credit the problem of credit risk. Um, then the risk. Um, then the risk. Um, then the cluster sizes cluster sizes cluster sizes became larger. became larger. became larger. The minimum contribution to the The minimum contribution to the The minimum contribution to the space has become larger. Back space has become larger. Back space has become larger. Back then it was 100 then it was 100 then it was 100 million. Now it's a million. Now it's a million. Now it's a billion. We don't billion. We don't billion. We don't think this can think this can think this can continue. We don't continue. We don't continue. We don't think there's a think there's a think there's a world where you can world where you can world where you can get a $10 billion round get a $10 billion round right away before you right away before you raise, uh, raise, uh, raise, uh, before you have before you have before you have revenue, or earnings, or your revenue, or earnings, or your revenue, or earnings, or your product, or anything else.

product, or anything else. product, or anything else. Um, so what...That means Um, so what...That means Um, so what...That means if you don't if you don't if you don't solve this problem, solve this problem, solve this problem, we're now on track to we're now on track to we're now on track to not having not having not having any any AI startups. Hmm, one of the AI startups. Hmm, one of the things you could things you could things you could imagine is that imagine is that imagine is that maybe maybe maybe new new inference companies and new inference companies and new model labs could emerge, model labs could emerge, model labs could emerge, and the way they would and the way they would and the way they would solve the credit problem is solve the credit problem is they would just they would just move the risk to the move the risk to the move the risk to the application layer. They application layer. They application layer. They say, "Listen, say, "Listen, say, "Listen, I need to I need to I need to get over my get over my get over my credit problem, hmm credit problem, hmm ." So I'm ." So I'm ." So I'm only going to sign only going to sign investment investment grade clients, grade clients, grade clients, you know, Fortune you know, Fortune you know, Fortune 500. Um, I'm 500. Um, I'm 500. Um, I'm only going to sign only going to sign only going to sign five-year five-year five-year token contracts. I'm token contracts. I'm token contracts. I'm only going to sign only going to sign only going to sign tokens that require a tokens that require a tokens that require a large deposit large deposit large deposit up front. Hmm, that's all it up front. Hmm, that's all it up front. Hmm, that's all it ends with. ends with. Hmm, if you have a Hmm, if you have a bottleneck at the bottleneck at the bottleneck at the logical inference level logical inference level logical inference level due to credit, you can't due to credit, you can't due to credit, you can't get get get tokens unless you tokens unless you tokens unless you have credit or a have credit or a have credit or a large deposit. And large deposit. And large deposit. And the problem is that the problem is that the problem is that tokens tokens tokens are used by the are used by the are used by the whole damn whole damn whole damn economy. There is no economy. There is no economy. There is no rule that says rule that says startups have to startups have to startups have to exist. Hmm, if exist. Hmm, if exist. Hmm, if every startup...

every startup... suddenly becomes a bottleneck suddenly becomes a bottleneck suddenly becomes a bottleneck because of credit because of credit because of credit to access to access to access information, what information, what information, what happens is that happens is that happens is that existing firms just existing firms just existing firms just supply everything. They supply everything. They supply everything. They produce the produce the produce the product themselves because they product themselves because they product themselves because they have credit, and then have credit, and then have credit, and then they get income they get income they get income from that product, and from that product, and from that product, and that improves their that improves their that improves their credit history, and that credit history, and that credit history, and that 's the flywheel that 's the flywheel that 's the flywheel that drives drives drives consolidation throughout the consolidation throughout the consolidation throughout the economy. If you don't economy. If you don't economy. If you don't solve this solve this solve this bottleneck, there bottleneck, there bottleneck, there will be no startups. That's why we will be no startups. That's why we will be no startups. That's why we started doing this. And started doing this. And started doing this. And because at some because at some because at some point, um, uh, or what point, um, uh, or what point, um, uh, or what we think we think we think will happen is that will happen is that will happen is that you'll find yourself in you'll find yourself in you'll find yourself in about 3-5 about 3-5 about 3-5 years, you'll look years, you'll look years, you'll look and say, "I want to and say, "I want to and say, "I want to create a new create a new create a new startup," and to startup," and to startup," and to get there, you get there, you get there, you need to buy need to buy need to buy enough tokens to enough tokens to enough tokens to compete with an compete with an compete with an existing partner existing partner existing partner or, for example, or, for example, or, for example, existing players. existing players. existing players. So you need to So you need to So you need to buy a hundred buy a hundred buy a hundred million dollars worth of tokens. We million dollars worth of tokens. We million dollars worth of tokens. We believe this is what believe this is what believe this is what you really you really you really have to do have to do have to do to start a new to start a new to start a new company because company because company because other people are spending other people are spending other people are spending hundreds of millions of hundreds of millions of hundreds of millions of dollars on tokens. If you dollars on tokens. If you dollars on tokens. If you just draw a line, just draw a line, just draw a line, this is what happens. this is what happens. this is what happens. So you need a So you need a So you need a big giant big giant big giant deposit because you are deposit because you are deposit because you are two people who two people who two people who have no idea what they have no idea what they have no idea what they are doing and so no one are doing and so no one are doing and so no one trusts you. So, trusts you. So, trusts you. So, you're going to you're going to you're going to raise $100 million raise $100 million raise $100 million from your from your from your venture venture venture capitalist.

capitalist. capitalist. Meanwhile, no one wants to Meanwhile, no one wants to Meanwhile, no one wants to work with you work with you work with you because you're a because you're a because you're a two- two- two- person startup. Then you'll person startup. Then you'll person startup. Then you'll be asked to pay be asked to pay be asked to pay for, um, this hundred, or you'll for, um, this hundred, or you'll for, um, this hundred, or you'll be asked to pay be asked to pay be asked to pay for a 5-year contract for a 5-year contract for a 5-year contract for this 100 million, um, for this 100 million, um, for this 100 million, um, unless you pay unless you pay unless you pay everything as a deposit. And then everything as a deposit. And then everything as a deposit. And then you simply won't you simply won't you simply won't get new get new get new model model model labs. You won't labs. You won't labs. You won't get new get new get new startups. Hmm, you'll startups. Hmm, you'll startups. Hmm, you'll get get get consolidation in the consolidation in the consolidation in the industry. So at some industry. So at some industry. So at some point we, um, were point we, um, were point we, um, were this little this little this little lab, um, and what lab, um, and what lab, um, and what we were doing was we we were doing was we we were doing was we lived in San Francisco, lived in San Francisco, lived in San Francisco, um, and we knew all these um, and we knew all these um, and we knew all these other people who other people who other people who were also just were also just were also just two little two little two little people, um, who people, um, who people, um, who had started a startup and had started a startup and had started a startup and wanted to be able to wanted to be able to participate in the participate in the most important most important most important technology of our technology of our technology of our time. Hmm, that's why I want time. Hmm, that's why I want time. Hmm, that's why I want these people to exist. these people to exist. these people to exist. I want this person to I want this person to I want this person to exist, and I want exist, and I want exist, and I want this person to this person to this person to exist. I want exist. I want exist. I want this startup to exist, this startup to exist, this startup to exist, and this startup... and this startup... to exist, and this to exist, and this to exist, and this startup will exist. startup will exist. startup will exist. This one, and this one, and this one, and this one This one, and this one, and this one, and this one , and this one, and this one, and this one, and this , and this one, and this one, and this one, and this , and this one, and this one, and this one, and this one. If you one. If you one. If you start these start these start these companies, maybe companies, maybe companies, maybe they'll become they'll become they'll become big players big players big players that develop your that develop your that develop your cancer cure, or cancer cure, or cancer cure, or they'll create something that no they'll create something that no they'll create something that no one has one has one has thought of, um, or something thought of, um, or something thought of, um, or something that's not just that's not just that's not just artificial intelligence artificial intelligence artificial intelligence thrown out there. This is the thrown out there. This is the thrown out there. This is the problem that problem that computing markets solve. And if you computing markets solve. And if you don't solve it, we don't don't solve it, we don't don't solve it, we don't think there will be think there will be think there will be any startups.