# The Dark Pools of Compute

Affaan Mustafa, Alejandro Soumah Fernandez · AI Engineer NYC 2026 · October 14

[Watch recording](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11709s)

[03:15:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11709s) All right, I think we can get started here.

[03:15:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11712s) Hey, everyone.

[03:15:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11713s) My name is Affaan, and I'm a lead for compute trading at Jump Trading, which is a proprietary

[03:15:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11718s) trading firm, alongside my co-lead over here, Alejandro.

[03:15:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11723s) So today we're going to be looking over some of the market structure as it applies to compute

[03:15:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11728s) trading, what compute trading even is.

[03:15:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11731s) what we can see it becoming.

[03:15:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11734s) And I think we'll start with some macro,

[03:15:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11737s) look at microstructure,

[03:15:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11738s) and then talk a little bit about

[03:15:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11740s) where we see things in policy.

[03:15:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11749s) So the title of the talk is

[03:15:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11752s) The Dark Pools of Compute,

[03:15:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11753s) but some of you guys are in finance,

[03:15:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11754s) so you might be confused as to why I'm using this terminology.

[03:15:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11757s) What do I mean by dark pool here?

[03:15:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11759s) Essentially what I'm referring to

[03:16:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11760s) is an unlit market. Right now, most compute trades in blocks, physical compute here. And

[03:16:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11767s) these are private quotes, we have a lot of brokers, and these are long bilateral contracts.

[03:16:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11771s) Now, at the end of the day, these are really commercial agreements. So we're using the

[03:16:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11776s) term trading here to kind of like look at how it's transacted between parties. These

[03:16:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11780s) commercial agreements tend to be some mixture of an MSA, an SLA, and an order form, cluster

[03:16:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11786s) specific with hardware specific language in each order.

[03:16:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11794s) But when we're looking at pricing and looking at some of these other metrics, a lot of the

[03:16:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11798s) time most of the talk is around the SKU, so what the chip itself is, the baseboard.

[03:16:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11803s) So H100, H200, V300, etc.

[03:16:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11807s) And so a lot of the negotiation here happens kind of outside the public view.

[03:16:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11813s) There's no exchange for physically traded compute that's widely adopted and used for everyone goes there to take a look and trade compute.

[03:17:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11823s) These are done in chats.

[03:17:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11826s) They're out of the public view.

[03:17:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11827s) There's public marketplaces.

[03:17:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11829s) There's price references.

[03:17:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11830s) There's a lot of on-demand

[03:17:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11832s) places. However, for these actual tenured, so on a given term for a fixed amount of capacity

[03:17:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11839s) with a certain total contract value, these trade over Slack messages, X messages, Telegram group

[03:17:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11845s) chats, so a bunch of OTC kind of channels that are all over in these kind of, I want to call them

[03:17:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11850s) like Neo chat. And some of you might be familiar with commodities and how a lot of, you know,

[03:17:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11856s) There's chatting happening on Ice Chat, Bloomberg Chat, etc.

[03:17:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11860s) This is kind of where the markets here lie.

[03:17:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11865s) And so I want to look a little bit here on the macro market structure

[03:17:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11868s) and a little bit on the players here as well.

[03:17:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11871s) And so this was a recent kind of A16Z.

[03:17:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11875s) Some of you guys might have seen this, so I'm putting this out there

[03:17:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11877s) because you can go on A16Z's actual X and take a look at this graph yourself.

[03:18:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11882s) What we're looking at is that the cost of a given capability

[03:18:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11884s) has fallen sharply.

[03:18:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11887s) The bill, of course, still depends on the model usage, capacity,

[03:18:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11890s) and other factors, such as speed,

[03:18:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11892s) is a big thing right now with the new ChatGPT models.

[03:18:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11897s) Cheaper tokens unlock more usage,

[03:18:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11899s) and a team can do more work for the same budget.

[03:18:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11901s) So what we're seeing is that a lot of teams

[03:18:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11904s) are simply just using way more tokens.

[03:18:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11906s) So we have this latent demand that hasn't really been hit yet,

[03:18:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11908s) and people are just going to use more with the same budget

[03:18:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11911s) that they were using before when models cost more.

[03:18:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11914s) and they're just spending more tokens.

[03:18:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11917s) So applications that were too expensive become possible.

[03:18:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11920s) However, as we're kind of talking about compute here,

[03:18:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11922s) the capacity underneath still has to be built and paid for,

[03:18:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11926s) even as the cost of a particular task comes down.

[03:18:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11928s) And so the important distinction is between the price of the intelligence,

[03:18:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11932s) the number of tasks people want to run,

[03:18:53](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11933s) and the amount of hardware needed to serve them.

[03:18:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11938s) And so this is where we get to kind of the Javon's paradox

[03:19:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11941s) and a little bit of some of the stuff that we've been looking at lately.

[03:19:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11944s) where we had a very interesting thing where, you know,

[03:19:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11948s) intuitively you'd assume that compute and inference

[03:19:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11951s) kind of like rises together, correlates very, you know,

[03:19:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11955s) very closely, as you can see.

[03:19:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11956s) You can see there's a little bit of a decoupling that happened,

[03:19:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11959s) and token prices declined further,

[03:19:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11961s) but compute prices remain robust.

[03:19:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11963s) And so, like, what can be an explanation for this now?

[03:19:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11965s) There's a lot of, like, potential explanations.

[03:19:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11968s) One here is the whole Javon's paradox,

[03:19:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11971s) is that even though token prices and inference

[03:19:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11973s) is getting extremely efficient to serve

[03:19:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11975s) on these new baseboards and GPUs,

[03:19:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11978s) people are just consuming a lot more.

[03:19:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11980s) And so the demand for compute rises

[03:19:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11982s) and people just consume a lot more tokens,

[03:19:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11984s) even if the cost per million for input and for caching

[03:19:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11989s) and for output becomes a lot more efficient.

[03:19:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11992s) And so making the tasks cheaper,

[03:19:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11995s) encouraging the additional use

[03:19:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11995s) to outweigh resource saving for tasks.

[03:19:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=11997s) For inference clouds, that's going to be more applications,

[03:20:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12000s) serving more tokens, bigger market share. And for AppLayer individuals, that means you guys can serve to a lot more customers, anyone who's utilizing the token infrastructure.

[03:20:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12011s) structure.

[03:20:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12013s) And so what really goes into the macro structure, now like this is very bare bones here, but

[03:20:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12020s) you've got your chips, which we've been talking about, you have power, you have buildings,

[03:20:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12025s) grid connections, and then of course you need a lot of capital to get these build outs going

[03:20:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12029s) and that's kind of been the whole talk here.

[03:20:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12032s) And so chips, power, buildings, money, they all have to come together before you can even

[03:20:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12037s) get to an hour of actual usable GPU per hour compute.

[03:20:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12041s) And so the IEA April update projects that the global data center electricity is going

[03:20:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12045s) to rise from 485 terawatt hours in 25 to 950 in 2030.

[03:20:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12051s) And so a lot of the talk right now is around build outs about getting the power grid set

[03:20:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12056s) up.

[03:20:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12056s) Will we have enough power grid here?

[03:20:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12058s) Is there other energy sources that we can tap into?

[03:21:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12061s) Where are we able to get the land to actually be able to do this?

[03:21:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12065s) And so, co-location sites and data center operators and people who are doing the financing, the build-outs, or the lenders themselves, are all kind of looking into this.

[03:21:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12071s) And that type of everything that's upstream, essentially, of the compute is going to then flow down into the compute costs and also usage.

[03:21:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12081s) And so, what does this, like, actual macro map look like underneath a quote?

[03:21:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12085s) Chip and system vendors supply the GPU service and networking.

[03:21:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12089s) Sites and operators supply the power space assembly and management.

[03:21:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12094s) right here, GPUs, memory service, networking.

[03:21:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12097s) Capacity providers, these are

[03:21:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12098s) hyperscalers, neoclouds, and any

[03:21:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12100s) other operator that's operating similar

[03:21:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12102s) to how a neocloud would operate, reselling

[03:21:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12104s) bare metal capacity, and then the users.

[03:21:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12106s) So who's the actual end user? The startups,

[03:21:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12108s) the inference clouds,

[03:21:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12110s) labs, neolabs, and also anyone

[03:21:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12112s) on the app layer, and enterprises

[03:21:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12114s) for anyone doing open-weight

[03:21:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12116s) stuff as well.

[03:21:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12118s) And these boxes,

[03:22:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12120s) I've kind of boxed them out here, but

[03:22:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12123s) there's people

[03:22:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12124s) People can be all of the above.

[03:22:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12126s) A chip and system vendor, someone could do a full end-to-end build out and also be the

[03:22:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12130s) end user of everything.

[03:22:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12132s) And so there's a little bit of variation that occurs here.

[03:22:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12143s) And so as I was kind of mentioning about the inference clouds before, they might rent a

[03:22:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12148s) cluster and they'll serve tokens off of it and they'll use that revenue to pay off the

[03:22:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12154s) And so that might be a use case for the actual inference cloud themselves.

[03:22:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12157s) However, for hyperscalers, frontier labs, and larger NeoClouds,

[03:22:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12160s) they can make substantial commitments

[03:22:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12161s) because they have the ability to get into the large equity,

[03:22:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12164s) they have the ability to take on more debt,

[03:22:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12167s) and they have the ability to do long horizon kind of deals.

[03:22:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12171s) So 10-year deal, open AI, contract it out, et cetera.

[03:22:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12175s) So of that nature.

[03:22:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12176s) However, anyone in the middle market isn't actually able to do this.

[03:22:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12179s) Who's underwriting a Series B startup

[03:23:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12181s) with enough capacity to get in a $200 million plus contract

[03:23:06](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12186s) over three years for compute.

[03:23:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12188s) How many equity rounds will they have to raise, right?

[03:23:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12190s) Will they have to use their entire equity round

[03:23:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12192s) to fund just the actual cluster

[03:23:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12194s) and then they don't have any remaining capital

[03:23:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12195s) left to use for actual costs?

[03:23:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12200s) And so we're looking into a little bit of the financing chain here

[03:23:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12202s) and how that plays in.

[03:23:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12203s) So lenders are financing the build-out

[03:23:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12206s) against contracted revenue even in many cases for inference clouds.

[03:23:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12209s) And so they want to finance the build-out.

[03:23:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12210s) They want confidence that their loan will get repaid.

[03:23:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12213s) I mean, like, how are you going to underwrite a startup,

[03:23:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12215s) like a private company that doesn't have, like,

[03:23:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12217s) much else going for them right now?

[03:23:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12218s) Like, what are the, you know, they're not publicly traded.

[03:23:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12220s) There's no that much cash flows, balance sheet, income statement.

[03:23:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12222s) They might even be pre-profit, pre-revenue, et cetera.

[03:23:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12225s) And so what they're looking at is, like,

[03:23:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12226s) can the person get contracted revenue over a certain scale?

[03:23:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12230s) And so Core Reef's public filings give large-scale example

[03:23:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12232s) of infrastructure financing supported by

[03:23:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12234s) take-or-pay customer contracts.

[03:23:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12236s) And so this diagram kind of shows the smaller,

[03:24:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12240s) illustrated example here

[03:24:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12241s) in the mismatch is that

[03:24:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12243s) hardware financing, they want very long terms,

[03:24:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12245s) but customers are only able to afford

[03:24:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12247s) or don't really want to get into a long-term

[03:24:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12249s) financing over 36 months.

[03:24:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12251s) They have to put up on after 15% to 25%

[03:24:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12253s) is what CoreWeave has quoted

[03:24:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12255s) that they've seen their median

[03:24:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12257s) prepayments to be, 15% to 25%

[03:24:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12259s) up front

[03:24:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12261s) on a 100 million TCV cluster.

[03:24:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12263s) That's $15 to $25 million

[03:24:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12265s) that a startup has to put up front.

[03:24:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12267s) That might be their entire round.

[03:24:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12270s) And so this is kind of a look at what I'm

[03:24:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12273s) referencing here. It's like you're raising $40 million, you need a $100 million contract

[03:24:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12277s) over three years, you put down 30% prepayments, this is an illustrative example,

[03:24:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12281s) you've got $10 million left to run the business, and so you might have this

[03:24:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12284s) kind of step function as you can see here, so that's your contract at capacity

[03:24:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12289s) and what you can actually, and the demand is below that.

[03:24:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12292s) And so now every time that you have to get into a new cluster, it's like tied to your

[03:24:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12296s) equity round or whatever else it may be, but you might have to cap out and turn away customers.

[03:25:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12300s) So what are your options? You either turn away customers until you can get into your next

[03:25:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12305s) equity round so you can put down another 30% prepayment so you can get

[03:25:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12308s) the next little bit of capacity going, or do you use on-demand to close the area

[03:25:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12312s) under the gap and bridge the shortfall? That's like an

[03:25:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12317s) option now. Now you have to look into the economics. What does that look like? What's the difference between

[03:25:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12321s) maybe instead what they do is they go to a lender to finance the prepayment component,

[03:25:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12324s) But if the yield on the prepayment component that they're getting financed is less than them just buying it out on demand at that time, if they're able to get into that amount of capacity, then does the economics make sense there?

[03:25:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12341s) And so here's some of the operating models that we've seen.

[03:25:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12344s) It's like you can build and operate the full stack.

[03:25:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12347s) Hardware, site, your team is going to arrange power permit, the assembly, the management.

[03:25:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12352s) You can also own servers in a co-location facility with more or less assembly and operation

[03:25:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12357s) handled by a partner.

[03:25:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12358s) On the other hand, an off-take contract buys defined capacity and the provider is going

[03:26:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12363s) to own and operate the asset and they might not give you complete root access to the bare

[03:26:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12367s) metal or there might be levels to how much access you get depending on if it's tenant,

[03:26:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12371s) multi-tenant, et cetera type of configurations that can also affect the pricing.

[03:26:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12377s) And so some businesses are going to value the physical access.

[03:26:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12380s) Some businesses are going to value confidential compute.

[03:26:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12385s) They might value certain topology changes

[03:26:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12387s) or the ability to reconfigure the system.

[03:26:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12388s) And so if that is something for your business,

[03:26:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12390s) then they might actually want to have more and more control

[03:26:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12393s) and build and operate the stack.

[03:26:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12394s) And the economics for them might make sense

[03:26:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12396s) given the opportunity costs they would have to give up

[03:26:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12398s) if they were to just take rental off-take contracts.

[03:26:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12401s) So they might be willing to pay more there.

[03:26:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12410s) And so here's how, like, ownership and rental, this is a very, like, crude graph, how ownership

[03:26:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12414s) and rental can compare as utilization arises.

[03:26:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12417s) And so what we're looking at is that usually, you know, an off-take contract or these off-take

[03:27:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12423s) contracts are getting to the cost of actually almost building out and financing your own

[03:27:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12427s) cluster.

[03:27:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12428s) And it's not, because, I mean, if you're spending $100 million over a three-year period,

[03:27:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12431s) then maybe if the economics makes sense, it would have made more sense to simply buy out

[03:27:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12435s) the compute operator yourself outside of the time cost and then obviously the operational

[03:27:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12439s) costs that might exist. So there's an equilibrium that exists that has to be followed where

[03:27:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12446s) the ownership costs outweigh the rental costs, and the rental costs should sit below the

[03:27:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12455s) ownership costs. But if you're utilizing at such a rate and in such variable kind of increments,

[03:27:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12460s) it might be best to just own the stack outright.

[03:27:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12467s) And so we're seeing this new dynamic

[03:27:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12469s) that the compute procurement

[03:27:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12471s) affects how much runway a growing company actually has.

[03:27:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12475s) And so this also gives investors

[03:27:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12477s) an interest in compute procurement.

[03:27:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12479s) So we're kind of seeing this.

[03:28:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12480s) Access and payment terms affect how much runway

[03:28:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12482s) their port codes are going to have

[03:28:04](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12484s) from a VC perspective.

[03:28:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12485s) So helping a company buy a suitable capacity

[03:28:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12488s) can improve the underlying investment.

[03:28:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12490s) And they're very connected, right?

[03:28:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12491s) So they're able to procure

[03:28:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12492s) and they're able to actually find compute capacity

[03:28:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12495s) for their port codes and fund that.

[03:28:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12497s) And so that also brings another interesting aspect

[03:28:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12500s) to the macro structure.

[03:28:21](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12501s) Into the micro structure is that

[03:28:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12502s) who helps place the capacity

[03:28:24](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12504s) and how does this offer get compared

[03:28:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12506s) and what risks these contracts can transfer.

[03:28:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12512s) So we have a lot of different players

[03:28:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12514s) around this GPU hour

[03:28:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12515s) that are helping kind of shape this market here.

[03:28:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12517s) You've got your price discovery individuals.

[03:28:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12519s) So people with indexes, data,

[03:28:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12522s) benchmarks that we're all looking at right now and referencing which would be

[03:28:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12526s) used to you know in the future launch the futures based on the benchmark it's

[03:28:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12531s) you know talked about in the associated contracts per spec per pending approval

[03:28:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12538s) and then there's the physical matching venues auctions aggregators so anyone

[03:29:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12542s) who can match capacity we're looking at essentially decentralized marketplaces

[03:29:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12547s) for on-demand compute. You're looking at liquid pools of compute that offer liquid pool services

[03:29:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12553s) and part it out themselves and have their own standardized contract. You're looking at bare

[03:29:17](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12557s) metal resellers. You're looking at neoclouds themselves doing larger capacity sales. And

[03:29:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12562s) then you're looking at individual brokers and individual operators here as well on the bare

[03:29:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12565s) metal reselling. You have destined brokers. They're looking at the bilateral quotes and execution.

[03:29:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12570s) They might know a lot of people. They might be able to move fast. They might be able to see quotes

[03:29:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12575s) have very similar specs in one place

[03:29:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12576s) and know that there's a buyer

[03:29:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12578s) that is extremely inelastic to price right now,

[03:29:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12581s) needs to compute immediately,

[03:29:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12582s) and they can get the RFS in time

[03:29:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12583s) so they can charge a decent spread on it.

[03:29:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12586s) And then what they do is they're going to have the source

[03:29:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12588s) negotiating, and then they're going to be carrying

[03:29:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12589s) and placing the risk.

[03:29:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12591s) They're going to have to operate that cluster for the person,

[03:29:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12592s) but it has to be worth that spread.

[03:29:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12595s) And you have looks in financial markets,

[03:29:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12597s) the exchanges that are coming that are based on the GPU

[03:30:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12600s) hour, the futures and derivatives.

[03:30:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12602s) People are going to be using it to take a look at how to maybe hedge out and look into more financial structures.

[03:30:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12607s) So hedge out risk, so hedge out balance sheet risk, maybe get into more interesting agreements.

[03:30:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12613s) We're seeing kind of agreements such as prepayment, payment facilities are coming online.

[03:30:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12618s) we're seeing people utilize futures

[03:30:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12620s) and are interested in utilizing futures

[03:30:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12622s) and other kind of

[03:30:25](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12625s) options to be able to

[03:30:27](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12627s) there's demand for this

[03:30:28](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12628s) to be able to hedge out the risk that they currently have

[03:30:30](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12630s) on the books. We have buy side

[03:30:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12632s) individuals looking for flex contracts

[03:30:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12634s) flexible contracts like hey

[03:30:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12635s) is there ability that if I lock myself into a 3-5 year

[03:30:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12638s) contract that I'll be able to

[03:30:39](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12639s) at least upgrade one of the new chips that comes out

[03:30:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12641s) or maybe I'll be able to roll it over or maybe let me sell it back to you

[03:30:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12645s) and then there's

[03:30:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12646s) credit and financing so equipment lenders, leasers

[03:30:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12648s) specialist credit providers.

[03:30:51](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12651s) And so the way we're seeing it right now

[03:30:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12652s) is that every deal is bespoke.

[03:30:55](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12655s) There's different spec sheets, there's formats.

[03:30:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12657s) The bill of materials may not match

[03:30:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12659s) the actual delivery,

[03:31:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12661s) or there might be a subject to change

[03:31:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12663s) if a part is missing or a part is out of stock,

[03:31:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12665s) and there's very long lead times

[03:31:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12667s) that some of these OEM and ODM

[03:31:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12669s) kind of providers have.

[03:31:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12671s) And the same term can mean

[03:31:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12673s) different things when you're negotiating very quickly.

[03:31:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12674s) If someone's talking about memory,

[03:31:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12676s) are we talking about, you know,

[03:31:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12678s) VRAM, are we talking about the CPU memory,

[03:31:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12680s) are we talking about additional storage here?

[03:31:22](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12682s) Like, what did you exactly mean there?

[03:31:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12683s) It affects quite a bit what type of loads can be run

[03:31:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12686s) and, you know, the actual end users' needs,

[03:31:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12689s) and also with pricing and delivery dates.

[03:31:33](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12693s) So there's a lot of back and forth

[03:31:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12694s) before anyone can really compare,

[03:31:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12695s) and this is where the concept of, like,

[03:31:36](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12696s) non-fungibility kind of comes into place

[03:31:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12698s) and how to partition that out,

[03:31:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12700s) what parts can be fungible,

[03:31:41](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12701s) what parts hold the most of the weight,

[03:31:43](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12703s) such as the skew, et cetera,

[03:31:44](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12704s) and what parts are still, like,

[03:31:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12706s) traded as if it was a product.

[03:31:47](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12707s) e-commerce, et cetera, where it's only for that particular cluster that this price applies.

[03:31:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12714s) And so prices can update daily or faster.

[03:31:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12716s) Quotes are still a good starting point.

[03:31:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12717s) Bids can get shopped.

[03:31:58](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12718s) Gradwarding has changes of what's on offer.

[03:32:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12720s) There's all these kind of latent factors, hidden factors as well,

[03:32:03](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12723s) why we're calling it a dark market.

[03:32:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12725s) These are some public H100 references to NAP-chotted today.

[03:32:09](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12729s) I updated them.

[03:32:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12730s) So take a look at what the public...

[03:32:11](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12731s) So those are the references for if you wanted to go on their websites to see.

[03:32:15](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12735s) This is just an example here.

[03:32:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12736s) of what a public on-demand or spot pricing looks like for an H100 today.

[03:32:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12743s) You can see various prices here.

[03:32:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12746s) People might use this or they might use their own internal,

[03:32:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12749s) what they've seen, to actually assume or use it to charge

[03:32:34](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12754s) what they want to charge on their actual cluster specification,

[03:32:37](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12757s) taking into account every other kind of factor that goes into it.

[03:32:42](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12762s) Why are these new structures needed and what are these microstructures?

[03:32:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12765s) Interesting.

[03:32:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12766s) It's mainly to solve for these mismatches, right?

[03:32:48](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12768s) Bridging capacity, more flexible commitments,

[03:32:50](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12770s) resale rights, and protection against price moves.

[03:32:54](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12774s) You know, you're selling a fixed-price service

[03:32:56](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12776s) from an inference cloud point of view.

[03:32:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12777s) You'd be selling a fixed-price service

[03:32:59](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12779s) while the GPU bill floats,

[03:33:00](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12780s) and that creates price exposure.

[03:33:02](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12782s) And so locking capacity for years can solve a part of that,

[03:33:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12785s) but it'll create idle hour exposure elsewhere as well,

[03:33:08](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12788s) unless you're always using it.

[03:33:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12790s) Reselling spare capacity only helps

[03:33:12](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12792s) if someone actually can buy the hours.

[03:33:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12793s) A financial hedge can address a reference price,

[03:33:16](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12796s) but can't deliver the cluster right now in terms of cash settlement

[03:33:18](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12798s) or guarantee a utilization.

[03:33:20](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12800s) So the useful structure has to start with the particular risk

[03:33:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12803s) that the business has and needs to manage.

[03:33:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12809s) And I'll go pretty fast on this policy section

[03:33:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12812s) because we're running out of time.

[03:33:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12815s) Everything that Affaan has said sits inside of a structure

[03:33:38](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12818s) that is regulated and is affected by policy.

[03:33:46](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12826s) In various ways. One way is export controls. One of the examples, in January, the U.S. Commerce Department moved qualifying H200 class license reviews for China and Macau to case-by-case review with conditions.

[03:34:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12841s) furthermore open weight models connect policy back to the u.s economics we started with the

[03:34:07](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12847s) u.s ai action plan explicitly encourages open source and open weight ai and recommends improving

[03:34:13](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12853s) the financial market for compute so that startups and academics can access large-scale computing

[03:34:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12859s) power so i'm going to tie this out right here but where's this opportunity like we have this

[03:34:26](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12866s) growing middle market. We need balance sheets that can bridge buyer and supplier needs. There needs

[03:34:31](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12871s) to be some level of standardization, a standard capacity contract, common specs, common terms,

[03:34:35](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12875s) and definitions. And so growing companies whose needs and commitment do not fit the contracts,

[03:34:40](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12880s) the hyperscalers, and Frontier Labs consign. Those companies have real workloads. We want them to

[03:34:45](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12885s) grow. They still face this financing problem, a timing problem, a specification problem. A supplier

[03:34:49](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12889s) might want the long-term commitment for the build-out while the buyer wants to expand in

[03:34:52](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12892s) smaller blocks as demand becomes clearer, as it's so volatile, front of curve. The buyer needs a way

[03:34:57](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12897s) to avoid paying for capacity it can't use yet, and so participants with strong balance sheets

[03:35:01](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12901s) that are uncorrelated with the market and the ability to assess those risks with a strong

[03:35:05](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12905s) credit profile can help bridge the gap between what buyers need and what suppliers offers.

[03:35:10](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12910s) That requires more than quoting the actual GPU hour. It means understanding the delivery,

[03:35:14](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12914s) workload, all of that, and so there's very few kind of players in the market, and a very well-suited

[03:35:19](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12919s) player, can be a firm that's uncorrelated with market, has a strong balance sheet,

[03:35:23](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12923s) strong credit rating, and incredible amounts of hardware talent that sits inside for high-performance

[03:35:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12929s) computing.

[03:35:29](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12929s) And so, just making a CTA here, let's build this together.

[03:35:32](https://www.youtube.com/watch?v=jiUC6IUMdvc&t=12932s) Thank you guys so much for attending our talk.