# How Compute Markets Work

Evan Conrad · AI Engineer NYC 2026 · October 14

[Watch recording](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10186s)

[02:49:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10188s) How are we doing?

[02:49:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10191s) So this is a talk about compute markets, but it's kind of more of a talk about why to make

[02:49:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10197s) a compute market.

[02:49:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10199s) If

[02:50:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10200s) you don't know me, I'm Evan.

[02:50:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10201s) We run a company called San Francisco Compute.

[02:50:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10204s) We make a physically settled order, physically settled forwards market.

[02:50:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10208s) So if you were here for the last talk, the last talk was basically why all the operators of GPU clusters don't really work very well.

[02:50:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10217s) When we say physically settled, we mean we're the operator.

[02:50:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10221s) We are both the operator of the hardware in the same way that a cloud is the operator

[02:50:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10225s) of the hardware, and then we are also the market at the same time.

[02:50:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10229s) It's kind of like a grid for compute.

[02:50:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10231s) So the general gist of this is that in order to run and operate a supercomputer, we think

[02:50:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10238s) problems start right at the design of the supercomputer, right in the design of the

[02:50:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10243s) bill of materials.

[02:50:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10244s) It's control of the backplane fabric as well.

[02:50:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10248s) It's also control of bare metal orchestration.

[02:50:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10250s) It's also services on top, VMs and so on.

[02:50:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10253s) There's a lot of different parts of the stack that you have to do effectively.

[02:50:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10257s) Otherwise, the folks who have a data center and would like to make more money on that

[02:51:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10264s) data center by becoming the GPU cloud end up failing for the customer.

[02:51:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10270s) So that's what we do.

[02:51:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10272s) We offer what's called supercomputing as a service.

[02:51:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10275s) We go to folks who want to become a cloud.

[02:51:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10276s) that's like folks who want to become what NVIDIA would call an NCP,

[02:51:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10281s) or venture firms that want to have clusters for their portfolio

[02:51:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10284s) and want to be able to reallocate that capacity

[02:51:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10287s) when a company like Instinct is taking off,

[02:51:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10290s) or AI companies who would like the residual value on the compute

[02:51:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10293s) because they've realized that's really important to their financials

[02:51:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10296s) but don't want to operate a supercomputer.

[02:51:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10298s) So that's physical settlement.

[02:51:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10300s) And then on top of that, we run a live order book

[02:51:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10302s) for capacity that you can buy reserved,

[02:51:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10304s) served, either right now or in the future.

[02:51:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10309s) This together is not at all what we thought we were going to be doing.

[02:51:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10314s) Originally, we were this company called JunLark that was trying to make kind of like Suno or

[02:51:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10319s) UDO, but prior to Suno or UDO.

[02:52:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10321s) At the time, what you needed to do to realistically start a lab was you needed to raise a big

[02:52:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10328s) deposit, a huge one, which was $10 million at the time.

[02:52:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10332s) We thought that was a crazy amount of money, so we thought, no, that's stupid.

[02:52:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10336s) We're going to raise 500k and obviously what we're going to do is rent from somebody else.

[02:52:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10341s) Obviously today it's not $10 million, it's a billion dollars.

[02:52:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10347s) And nonetheless, we had raised our 500k, we buy a big cluster, and then we're underwater

[02:52:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10353s) on our cluster every month, and if we didn't sublease our cluster, we would have gone bankrupt.

[02:52:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10358s) If you have heard of GPU brokers, we were the first GPU broker, so the first year the

[02:52:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10364s) company was us trying to sell our cluster, otherwise we would go bankrupt.

[02:52:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10367s) So I would wander around with these actual pieces of paper at NeurIps, which both say

[02:52:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10373s) H100 and A100 at the same time, because that's how I made the posters so quickly.

[02:53:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10381s) And so this talk is not really about necessarily how a compute market works, it's how a compute

[02:53:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10388s) market gets to exist.

[02:53:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10391s) We like to think about it like there's this pamphlet, and on the pamphlet, on the front

[02:53:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10395s) side, it's like, man, starting a Neolab sucks, there's all these different financial and

[02:53:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10399s) economic problems.

[02:53:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10400s) On the back of the pamphlet, it says compute markets.

[02:53:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10404s) It's not super clear how one gets to the other until you open the pamphlet and then

[02:53:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10409s) And holy shit, there's so much stuff in this pamphlet.

[02:53:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10411s) And it keeps going.

[02:53:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10412s) There's so many folds.

[02:53:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10413s) It's a difficult and challenging problem.

[02:53:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10415s) Because you need to solve two miracle problems to get working.

[02:53:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10419s) You have to both make a market, which is a uniquely difficult problem because it has

[02:53:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10423s) a cold start problem.

[02:53:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10424s) You also have to build a cloud.

[02:53:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10425s) And you have to not become a cloud.

[02:53:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10427s) Which is a whole separate challenge.

[02:53:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10430s) But these are the problems that the labs are facing.

[02:53:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10433s) And it's about to be the problems that everyone is about to be facing.

[02:53:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10435s) So if you work at a start-up, these are your problems now.

[02:53:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10438s) Right now, everybody is sold out of compute.

[02:54:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10441s) This is probably not going to change for a while.

[02:54:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10445s) So that means you're going to have to buy a new cluster, a net new one, and there's

[02:54:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10449s) a lead time.

[02:54:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10450s) You have no way to pay for that cluster to come online now.

[02:54:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10453s) You're going to pay for it to come online later.

[02:54:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10454s) There's nothing in the market that can realistically get that cluster online for you now.

[02:54:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10459s) Meanwhile, you're going to pay your entire bank balance as a deposit.

[02:54:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10463s) The reason you're paying your entire bank balance as a deposit

[02:54:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10465s) is because the lender has no idea how to underwrite AGI.

[02:54:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10469s) How in the world are you supposed to do that?

[02:54:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10471s) You've got two-person startups in San Francisco, or maybe New York,

[02:54:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10475s) who say, I'm going to build superintelligence.

[02:54:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10479s) No lender knows how to underwrite superintelligence.

[02:54:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10482s) So they have no idea what the asset is worth.

[02:54:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10485s) Meanwhile, your cluster is very likely to be delayed.

[02:54:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10488s) That's because the lender needed to underwrite against the credit risk of the customer,

[02:54:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10491s) and so they need to know who is the customer first

[02:54:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10494s) before you deploy the cluster.

[02:54:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10496s) So right when the customer signs

[02:54:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10497s) is the riskiest point in the entire operation.

[02:55:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10500s) You wanted it to be the case

[02:55:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10501s) that you built the cluster first and then you sell it.

[02:55:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10504s) This also helps reduce all sorts of different types of problems

[02:55:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10507s) where clusters break.

[02:55:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10509s) Meanwhile, most of the people who are setting up clusters

[02:55:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10511s) are people who have the actual asset

[02:55:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10513s) that makes the thing cheap.

[02:55:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10514s) Like you've got the data center space already.

[02:55:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10517s) What that doesn't get is supercomputing expertise.

[02:55:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10520s) as we saw in the last talk, if you've watched that one.

[02:55:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10525s) Meanwhile, you're going to run into the exact same problem

[02:55:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10527s) that Junelark did,

[02:55:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10528s) which is you're going to either way overbuy

[02:55:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10530s) or you're going to way underbuy.

[02:55:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10532s) That's because AI doesn't have margins

[02:55:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10534s) that look like SaaS companies.

[02:55:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10535s) And so you don't have very much margin for error.

[02:55:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10537s) And so the chances that you either underbuy

[02:55:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10540s) and somebody else wins because they did it correctly

[02:55:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10542s) or you overbuy and then you just become stability AI

[02:55:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10545s) or you blow up, you end up like character AI,

[02:55:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10548s) are really high.

[02:55:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10550s) So, these are all the same problem. They're all just different types of risk.

[02:55:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10553s) They're technical risk, they're financial risk, they're credit risk.

[02:55:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10556s) They're all one big bundled risk problem, which is the problem that we end up solving,

[02:56:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10560s) and it's also the problem that we think compute markets broadly, the entire industry solves.

[02:56:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10565s) So we'll walk through each one of these and walk how that gets to a compute market,

[02:56:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10568s) because if you don't understand these, it's kind of unclear why you would want a compute market,

[02:56:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10572s) and they almost seem like this problem that somebody thinks,

[02:56:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10575s) oh, wow, compute markets are cool, so I'm going to go start one.

[02:56:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10577s) We think they should be derived from the actual problems that people face,

[02:56:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10581s) which is, first, everyone is sold out.

[02:56:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10583s) Why are they sold out?

[02:56:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10584s) Because the lender couldn't finance the cluster

[02:56:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10586s) unless you had a long-term off-take agreement.

[02:56:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10588s) In CPUs, this wasn't the case,

[02:56:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10590s) because CPU buyers are SaaS companies.

[02:56:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10594s) SaaS companies have high margins,

[02:56:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10595s) and because they have high margins,

[02:56:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10597s) you can build a product that rewards you with high margins for your CPUs.

[02:56:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10608s) Think about it this way.

[02:56:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10610s) If you're a SaaS company and your underlying cloud doubles their prices, you don't go bankrupt.

[02:56:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10615s) You hate your cloud, you're mad, but you don't go bankrupt.

[02:56:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10618s) If you're an AI company and your underlying cloud doubles their prices, you go bankrupt.

[02:57:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10623s) And what that means is that you're actually price sensitive, and because you're actually

[02:57:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10627s) price sensitive, there's nothing keeping you at any particular cloud, so you can't sell

[02:57:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10632s) on demand because your customers will immediately leave to whoever is cheaper.

[02:57:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10636s) And so that means that in order to finance GPU clusters, you have to have long-term take-or-pay

[02:57:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10642s) contracts, hopefully with someone who has investment-grade credit.

[02:57:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10646s) However, if you don't have liquidity on those, if you're selling reserved and your customer

[02:57:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10650s) can't get out, then that also means that some other customer who wants the capacity right

[02:57:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10654s) now because they're like instinct and their product is taking off, can't pay for anything.

[02:57:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10658s) They have to finance a new cluster which has a lead time because the physical world just

[02:57:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10661s) exists and it's going to take a while.

[02:57:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10663s) Sometimes you even have to build the damn data center.

[02:57:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10666s) And then, because you're financing a new cluster, you're going to pay a deposit.

[02:57:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10669s) Because almost nobody in AI has any reasonable credit.

[02:57:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10673s) And because this is a brand new technology, all these companies are new.

[02:57:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10676s) Meanwhile, even as you start to scale up that ladder, you're still going to run into credit

[02:58:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10681s) issues all the way through.

[02:58:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10682s) It's the number one bottleneck of AI at the moment.

[02:58:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10684s) It's not power.

[02:58:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10685s) You hit credit way before you hit power.

[02:58:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10691s) So, what that means is every inference company, every neo lab, they are bottlenecked by how

[02:58:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10696s) much cash they have in the bank account.

[02:58:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10698s) So they go back to the venture capitalist who says, oh, I'm going to fund at the model

[02:58:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10702s) layer, the application layer, the inference layer, because I don't want to fund all of

[02:58:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10705s) that capex stuff.

[02:58:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10707s) And so the venture capitalist then funds all the capex stuff by giving the new neo lab

[02:58:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10714s) or new inference company, the check that they then send to the cloud, which is the deposit

[02:58:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10720s) for the cluster, and that's how you underwrite the customer.

[02:58:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10725s) So the clusters then are delayed because the cloud couldn't actually deploy the cluster

[02:58:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10732s) until they had the customer in hand, and so that was the riskiest point.

[02:58:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10736s) So therefore, what should have happened was the cloud had bought the cluster ahead of

[02:59:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10742s) time, and then waited until it was landed in the data center, and then sold it.

[02:59:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10747s) And that means you have way less likelihood of delays.

[02:59:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10750s) Meanwhile, the cluster is very likely to break when you get the cluster as the customer because

[02:59:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10755s) you don't have liquidity.

[02:59:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10756s) And so if the failure rate is higher than what you expect, and you don't have the ability

[02:59:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10760s) to transition a node in, then you don't have anywhere you can pull nodes from.

[02:59:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10764s) Which means that if you need new nodes to make up for the failed ones, you have to RMA

[02:59:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10768s) them.

[02:59:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10769s) which means you have to get a new node.

[02:59:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10770s) And then that gets us back to several steps before

[02:59:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10773s) where you have the lead time problem again.

[02:59:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10776s) And so everything breaks for everybody all the time.

[02:59:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10780s) Once you have gotten through all of that as a new AI lab,

[02:59:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10782s) your problems are you, you know, you, nobody was sold,

[02:59:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10786s) like everyone was sold out.

[02:59:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10788s) You then had to wait a lead time.

[02:59:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10789s) You paid this big deposit.

[02:59:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10791s) The thing was delayed and then it broke.

[02:59:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10792s) And finally you got through all of that.

[02:59:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10794s) And the likelihood that you predict your demand correctly

[02:59:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10796s) is very low.

[02:59:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10798s) And so you're either going to be way

[03:00:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10800s) over-provisioned or way under-provisioned, and you have no way out.

[03:00:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10802s) There's no way to sell your capacity to somebody else, and there's no way to buy more capacity if you need it.

[03:00:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10806s) So, compute markets.

[03:00:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10809s) To solve these problems, you need to solve a bunch of different problems all at once.

[03:00:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10813s) You need to underwrite the value of the GPU cluster rather than underwriting the credit risk of the customer.

[03:00:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10820s) This is how it could work.

[03:00:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10822s) It's not guaranteed to work like this.

[03:00:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10824s) In other industries, what happened is everything just consolidated.

[03:00:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10828s) You got one big credit-worthy customer that does the whole damn industry.

[03:00:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10832s) That's what the path of AI is currently on track to do.

[03:00:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10836s) It's instead of going to be a liquid market where anybody can participate,

[03:00:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10840s) it's going to look like TSMC.

[03:00:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10842s) You're going to get one buyer or a few buyers that are the clouds.

[03:00:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10846s) You're going to get a few model labs that are the AI labs.

[03:00:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10849s) And then you won't get other startups.

[03:00:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10852s) This is what we think is going to happen.

[03:00:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10854s) And this is why three years ago, when we were doing our AI lab, we had a chance.

[03:00:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10858s) We could have said, no, we're not going to stick it out for several years just trying to keep our company alive.

[03:01:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10866s) We're just going to go do a new AI lab.

[03:01:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10868s) But we saw this problem.

[03:01:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10870s) And we said, oh, shit.

[03:01:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10872s) If we don't do anything about this, then we will just hit this three years from now.

[03:01:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10877s) And here we are three years from now hitting this problem and where everyone is going to hit this problem.

[03:01:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10882s) If you watched the last talk, it's very hard to solve this problem unless you are operating

[03:01:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10888s) both the market and the clusters at the same time.

[03:01:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10893s) This is Dylan Page.

[03:01:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10894s) Dylan is the co-chair of the CNCF infrastructure group.

[03:01:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10896s) That's the organization that makes Kubernetes.

[03:01:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10898s) He formerly worked at Lambda doing infrastructure there.

[03:01:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10901s) And he joined San Francisco Compute because GPU Compute isn't sufficiently reliable for

[03:01:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10908s) the market operator to be divorced from the physical infrastructure.

[03:01:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10910s) It doesn't work if you aren't the one who is buying the thing.

[03:01:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10913s) Sorry, it doesn't work if the person you are buying from isn't operating it.

[03:01:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10916s) To give an example of why this doesn't work is I'm sure some of you are flying in from

[03:02:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10920s) overseas or out of town and you have a hotel room.

[03:02:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10926s) Wouldn't it be great if you could sell that hotel room to somebody else?

[03:02:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10930s) I go to you and I want to buy that hotel room and I say, can I get room service?

[03:02:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10936s) Can I get my room cleaned?

[03:02:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10938s) I've got a problem with my room, who do I call?

[03:02:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10941s) If you don't operate the hotel, you can't fix any of those problems.

[03:02:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10945s) That's why brokering clusters doesn't work.

[03:02:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10948s) You have to operate the cluster like the hotel.

[03:02:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10950s) You have to run the whole damn thing.

[03:02:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10952s) Otherwise, you can't actually make a market.

[03:02:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10954s) It doesn't work.

[03:02:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10955s) You can't do physical settlement.

[03:02:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10957s) So why does nobody do physical settlement?

[03:02:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10960s) The problem is there's a very strong incentive to never do physical settlement.

[03:02:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10964s) That is because if you build out physical settlement for compute, what that literally

[03:02:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10968s) means is you are building out the entire operating capacity of a GPU cloud.

[03:02:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10972s) And if you are building out the entire operating capacity of a GPU cloud, you should just be

[03:02:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10977s) a GPU cloud.

[03:02:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10978s) Because it looks to be the case that you will make an ungodly amount of money.

[03:03:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10982s) You will sign billion dollar deal after billion dollar deal.

[03:03:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10986s) And so, fuck it.

[03:03:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10988s) Let's not make a GPU cloud, or let's not make a compute market.

[03:03:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10990s) Let's instead just become a GPU cloud, which is what basically everybody does, because

[03:03:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10995s) the numbers are astronomical and it's a very reasonable thing to do.

[03:03:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=10998s) The reason why we ended up building physical settlement was because we knew that if we

[03:03:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11002s) didn't start this, we would then not be able to start an AI lab, which was the real damn

[03:03:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11006s) thing that we wanted to do in the first place.

[03:03:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11007s) We wanted to make a lab.

[03:03:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11008s) We didn't get to make a lab unless this exists.

[03:03:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11012s) So when we started, it was 10 million as the deposit that you had to pay to start a new

[03:03:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11018s) lab.

[03:03:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11018s) This is how you get over the credit risk problem.

[03:03:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11022s) Then the cluster sizes got bigger, the sort of minimum ante into the space got bigger,

[03:03:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11027s) then it was 100 million, now it's a billion.

[03:03:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11029s) We don't think it can continue.

[03:03:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11031s) We don't think there is a world in which you can get a $10 billion round right out the

[03:03:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11035s) bat before you have revenue or traction or your product or anything.

[03:04:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11040s) And what that means is that if you don't solve this problem, we are currently on track to

[03:04:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11045s) not have any start-ups in AI.

[03:04:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11047s) Now, one of the things that you could think could happen is maybe there can be new inference

[03:04:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11053s) companies and new model labs.

[03:04:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11055s) And the way that they solve the credit problem is that instead they just ship the risk to

[03:04:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11059s) the application layer.

[03:04:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11060s) They say, look, I need to get over my credit problem, so what I'm going to do is I'm going

[03:04:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11068s) to only sign customers who are investment grade credit, you know, Fortune 500, I'm going

[03:04:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11073s) to only sign customers on five-year contracts for tokens.

[03:04:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11077s) I'm going to only sign customers for tokens that require a big deposit up front.

[03:04:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11084s) That is the end of everything.

[03:04:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11088s) you have inference layer getting bottlenecked by credit, what happens is you can't get tokens

[03:04:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11094s) unless you have credit or you have a big deposit.

[03:04:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11097s) the problem of that is that tokens are used by the whole damn economy.

[03:05:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11101s) There is no rule that says that startups have to exist.

[03:05:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11105s) If every startup is suddenly bottlenecked by credit in order to access intelligence,

[03:05:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11111s) what happens is existing firms just ship everything.

[03:05:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11114s) They themselves get to make the product because they've got credit,

[03:05:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11119s) and then they make the revenue from that product, and then it makes their credit better.

[03:05:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11122s) And that's a flywheel that makes consolidation happen throughout the entire economy.

[03:05:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11127s) If you don't fix this bottleneck, there aren't any startups.

[03:05:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11130s) This is why we started doing this.

[03:05:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11132s) Because at one point, what we think is going to happen is you're going to end up in three

[03:05:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11138s) to five years, you're going to look out and say, I want to start a new startup.

[03:05:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11142s) To get there, you need to buy so many tokens in order to compete with the existing players.

[03:05:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11150s) You need to buy $100 million of tokens.

[03:05:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11154s) We think this is what you're going to have to do in order to start a new company because

[03:05:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11157s) other people are spending hundreds of millions of dollars of tokens.

[03:06:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11160s) If you just draw the line out, this is what happens.

[03:06:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11162s) So you need a big giant deposit because you're two people who have no idea what they're doing

[03:06:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11167s) and so nobody trusts you.

[03:06:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11169s) So you're going to go and raise $100 million from your venture capitalist.

[03:06:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11173s) Meanwhile, nobody wants to work with you because you're a two-person startup.

[03:06:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11177s) Then you're going to be asked to pay for that hundred, or you're going to be asked to pay

[03:06:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11181s) for a five-year contract on that hundred million if you're not paying it all as a deposit.

[03:06:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11185s) And then you just don't get new model labs, you don't get new start-ups.

[03:06:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11188s) You get a consolidation in the industry.

[03:06:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11191s) So we, at one point, were this little lab.

[03:06:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11196s) And what we did is we lived in San Francisco and we knew all the other people who were

[03:06:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11201s) also just two little folks who had started a start-up and wanted to be able to participate

[03:06:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11206s) in the most important technology of our time.

[03:06:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11210s) So I want these people to exist.

[03:06:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11212s) I want this person to exist and I want this person to exist.

[03:06:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11215s) I want this startup to exist and this startup to exist and this startup to exist.

[03:06:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11219s) This one and this one and this one and this one and this one and this one and this one and this one.

[03:07:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11223s) If you get these companies going, maybe they'll become big folks that build your cancer cures

[03:07:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11229s) or they build something that nobody thought of or something that wasn't just sort of AI slopped out.

[03:07:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11235s) This is kind of the problem that compute markets solve.

[03:07:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11238s) And if you don't solve it, we don't think there's going to be any startups.

[03:07:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11242s) Cool.

[03:07:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11244s) That's it.

[03:07:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11245s) That's it.

[03:07:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11245s) Yeah.