The Programmable Data Center — Evan Conrad, SF Compute

“So our impression is that NVIDIA can't make it, but benefits tremendously from it existing. The reason why they can't make it is because they would then not be independent enough for people to trade on it.”

Claim: SF Compute's impression is that NVIDIA cannot create the market, but benefits tremendously from the market existing.

“So our impression is that NVIDIA can't make it, but benefits tremendously from it existing. The reason why they can't make it is because they would then not be independent enough for people to trade on it.”

Claim: SF Compute believes NVIDIA cannot create the market because NVIDIA would not be sufficiently independent for people to trade on it.

“So uh when we started we were this company called Junark and then we were just trying to sublease our cluster. Um and w we started there in the technical stack um and a way to think about it is it's like you're in a hotel and you're trying to sell that hotel room and then somebody asks you hey like are you gonna clean my room? And then you say oh, I guess I probably also need the hotel room as well.”

Claim: SF Compute began as a company called Junark that was trying to sublease its compute cluster.

“my supplier on this market is somebody that looks closer to a fund, or a cloud that has access, in our case, a cloud that has access to the data center, and we're going to level up. And so what we've spent the last several years doing is building the entire operating capacity of a cloud in order to offer, in order to actually deliver the compute at a reliable level. We call that physical settlement. We're taking a term from”

Claim: SF Compute has spent several years building the full operating capacity of a cloud to deliver compute reliably.

“level. We call that physical settlement. We're taking a term from”

Claim: SF Compute uses the term “physical settlement” for reliably delivering compute.

“In some markets, like in oil and also in power, there's like a delivery network. In oil, you have to have pipelines. If you would like the oil to be delivered to you, there has to be a way for it to get to you.”

Claim: Physical delivery in oil and power markets depends on a delivery network, and oil must travel through pipelines to reach its recipient.

“there's parts of the orders that you fill out, the buy orders and sell orders, that are there because the boiler takes a while to cool off. And you can stop the boiler. boiler and then it will you can start it back up again within some period of time and so that has to get factored into the order book. Um this genre thing is the same thing that happens in compute markets. So we co-designed physical settlement and the order book at the same time and that meant that we get to actually handle the real constraints of compute which might include something like oh I'm okay with being interrupted, but only if there's like storage on the thing that I can recover from really qui”

Claim: Power-market buy and sell orders account for the time boilers need to cool down and the delay before they can be restarted after being stopped.

“boiler and then it will you can start it back up again within some period of time and so that has to get factored into the order book. Um this genre thing is the same thing that happens in compute markets. So we co-designed physical settlement and the order book at the same time and that meant that we get to actually handle the real constraints of compute which might include something like oh I'm okay with being interrupted, but only if there's like storage on the thing that I can recover from really qui really quickly. Um very similar types of problems.”

Claim: SF Compute co-designed physical settlement and its order book to represent real compute constraints, including a buyer's willingness to accept interruption only if the workload can be recovered quickly from storage.

“Uh not very d mature or deep or liquid.”

Claim: The computer-capacity market is not very mature, deep, or liquid.

“So what we've been doing is first we had to go get a bunch of supply online in the first place, which just takes a while.”

Claim: SF Compute had to bring a substantial amount of supply online first, and doing so takes time.

“And then we had to build the entire platform in order to do this.”

Claim: SF Compute had to build its entire platform to support this capacity marketplace.

“And we just now got that working. And so now we move capacity over onto the order book,”

Claim: At the time of the conversation, SF Compute had just gotten its platform working and was moving capacity onto the order book.

“which makes it liquid.”

Claim: Listing capacity on the order book makes that capacity liquid.

“And so 5% of our total capacity is on the order book. And we're consistently moving more and more.”

Claim: At the time of the conversation, 5% of SF Compute's total capacity was on the order book, and SF Compute was consistently adding more.

“And then we're bringing on more capacity that's all on the order book.”

Claim: SF Compute is bringing new capacity online with all of that capacity listed on the order book.

“the order book. Um the nice part is that um I maybe I should have said the the volume on the actual order book is small. But the um the liquidity on the order book is high. So for that small amount of capacity that you have, um you have um like very concentrated liquidity. So um”

Claim: SF Compute's order book has low volume but high liquidity, concentrating liquidity in the relatively small amount of listed capacity.

“I would say, for someone who is using this is a customer who buys capacity from us and they spend $100 to buy that capacity will make back $25 on average from selling it back to somebody else.”

Claim: On average, a customer who spends $100 buying capacity from SF Compute can earn $25 by reselling that capacity to another buyer.

“And that's the number that we or this is one of the metrics that we try to track because we want the liquidity to be useful to the commercial participant and you can measure liquidity in a bunch of different ways. You can say like, oh, I just want tons of makers or something.”

Claim: SF Compute tracks customers' resale returns as one measure of whether its liquidity is useful to commercial participants.

“And you can measure it in like high-minded ways. We try to measure it by like, is this solving the problem um for the customer.”

Claim: SF Compute evaluates liquidity in part by whether it solves the customer's problem.

“Um so I think standardising the contracts in general is still pretty far away. Um”

Claim: Standardization of compute-capacity contracts remains a considerable distance away.

“Or basically what we learned over the last several years was w we just needed to operate the whole thing. Otherwise you cannot standardize the thing. And so to get closer to something that can be standardized, you first need to make all the operations of the thing the same.”

Claim: Standardizing the system requires first making its operations uniform.

“Or basically what we learned over the last several years was w we just needed to operate the whole thing. Otherwise you cannot standardize the thing. And so to get closer to something that can be standardized, you first need to make all the operations of the thing the same.”

Claim: SF Compute concluded from several years of experience that it had to operate the entire system in order to standardize it.

“To give an example of what can go be different is a single operator can h offer two different types of virtual machines. One type of virtual machine um might have a f forty percent performance um degradation versus another types of virtual machine.”

Claim: Two virtual-machine types from the same operator can differ substantially in performance; one type might perform 40% worse than the other.

“And forty percent is a lot.”

Claim: A 40% performance difference between virtual-machine types is substantial.

“It will affect, so in this particular case I'm talking about G_P_U_ kernel launches and how long it takes for that to happen, um which is a process that happens quite frequently when you're doing sort of either training or inference.”

Claim: The cited virtual-machine performance difference concerns GPU kernel-launch latency, which is incurred frequently during model training or inference.

“Uh that will basically the end result is that you get a worse M_F_U_ but um like is the intermediate level Yeah, is it the intermediate level basically?”

Claim: The performance difference ultimately results in worse MFU.

“So what we did instead um was we created an order book that had multiple different constraints. And you can say I wanna have um you know something that's between this and that um it runs from this time to this time um it's like uh X_ to Y_ it's a properties that you can attach to your order, and then we match that against sell orders that have those properties”

Claim: SF Compute's order book lets buyers specify multiple capacity constraints, including property ranges and time windows, and matches their orders with sell orders that meet those properties.

“um in a big matching engine. Um and the reason why we did it this way is it allows us to generate like a crunch.”

Claim: SF Compute's order book uses a large matching engine to match capacity orders by their specified properties.

“rate liquidity without having to standardize before you know what the technology can do. Um and because the technology was shifting so quickly, um we wanted to basically be able to do that. The trade-off of this is that you have less concentrated liquidity, and so the prices are um going to ha there's gonna be more spreads between the the buy and the sell.”

Claim: SF Compute chose a flexible, multi-constraint order book because the underlying technology was changing rapidly.

“rate liquidity without having to standardize before you know what the technology can do. Um and because the technology was shifting so quickly, um we wanted to basically be able to do that. The trade-off of this is that you have less concentrated liquidity, and so the prices are um going to ha there's gonna be more spreads between the the buy and the sell.”

Claim: SF Compute's flexible order-book approach produces less concentrated liquidity and wider bid-ask spreads.

“rate liquidity without having to standardize before you know what the technology can do. Um and because the technology was shifting so quickly, um we wanted to basically be able to do that. The trade-off of this is that you have less concentrated liquidity, and so the prices are um going to ha there's gonna be more spreads between the the buy and the sell.”

Claim: SF Compute's multi-constraint order-book approach can generate liquidity without requiring standardization in advance of learning the technology's capabilities.

“but on the other hand, it means there's lots of arbitrage to make, um if you can do that correctly. So if you have a product”

Claim: The less-concentrated liquidity creates substantial arbitrage opportunities for participants who can exploit them correctly.

“product that can use lots of different types of capacity, um then you might really like S_F_ compute, because you will be able to use lots of different types of capacity and you'll make more money. Um you can take the uh the spread.”

Claim: A product able to use many types of compute capacity might benefit from SF Compute by using varied capacity and earning money from the price spread.

“A bunch of different things. Um, one is dimensions categories that I will say is like is there a workload running on it right now, or is it capacity in the future. So is it Sure. capacity right now? Um capacity right now, in order for you to sell that, um you have to interrupt your current workload. So you have to think, do I care if my current workload is interrupted or not. Capacity that's sold that's starting a month from now, um you can buy and sell that all the time, and so you get”

Claim: Capacity scheduled to become available a month later can be traded readily and is more liquid than capacity whose sale interrupts an active workload, although future availability remains a concern.

“A bunch of different things. Um, one is dimensions categories that I will say is like is there a workload running on it right now, or is it capacity in the future. So is it Sure. capacity right now? Um capacity right now, in order for you to sell that, um you have to interrupt your current workload. So you have to think, do I care if my current workload is interrupted or not. Capacity that's sold that's starting a month from now, um you can buy and sell that all the time, and so you get”

Claim: Selling currently available data-center capacity requires interrupting its active workload, making the impact of interruption relevant to the sale.

“um liquidity on that really easily. Um there is some concern like oh no will I have capacity next month, but it's very different than if you're interrupting a workload right now. Um the way that the interruption happens matters. So an example is if you're on a cluster and the cluster doesn't ha have as amazing storage as you want it to have, um so an example is like do you have a wack or a vastus system under the hood, um and do you offer that and like what's the size of that. Um uh this changes how you respond to um being interrupted. being preempted or removed off the cluster.”

Claim: A workload's response to being preempted from a cluster depends in part on the cluster's storage system and storage capacity.

“Meanwhile, the location of that storage matters. So if you're able to move that data from one cluster to another quickly, then maybe you might be able to spread across multiple different clusters and you can be interrupted over here and then pop back up over there.”

Claim: If data can be transferred quickly between clusters, a workload may be able to move across clusters and resume on a different cluster after preemption.

“To make that work, you also need like a fast transit path between one cluster another. Typically, this is implemented with a WAN.”

Claim: Moving workloads between clusters requires a fast inter-cluster transit path, typically a wide-area network.

“yeah, a wide area network. So um basically cables that connect the clusters together. A private internet is another way to think about it. Um uh lot of clusters that are being spun up today are like uh they're not attached to um traditional CPU clusters were um Next to the customer for latency reasons, in GPUs they're next to the power. Um wherever the t like you know s you put the data centre next to the power rather than next to the customer.”

Claim: Many GPU data centers are being located near power sources rather than customers, whereas traditional CPU clusters are located near customers to reduce latency.

“Um this means that a lot of stuff is getting put in places that previously they weren't expecting data centres there. And so you may not have um as um excellent of a transit path um from one to the other. So this this starts to matter. Um other examples are um weird sh it where we theoretically support, but don't actually um uh”

Claim: Building data centers in new locations can mean that transit paths between clusters are less capable, increasing the importance of inter-cluster connectivity.

“Like this is a thing that we're thinking about but haven't done basically. Um the early G_B_ two hundreds broke, the early H_ one hundreds broke, um and so if you were buying H_ one hundreds or you're buying G_B_ two hundreds, you might really want the ones that were deployed after a certain time because you're not getting the first G_D_ ones off the rack. And so if you're a super sophisticated customer, you may be willing to pay more um because”

Claim: Because early H100 and GB200 deployments experienced failures, sophisticated customers may pay more for later-deployed units instead of first-off-the-rack units.

“So like uh the first ones off the rack maybe had some fault or some problem, and so if you were the ones who got the first ones um or you know first ones from um the the first shipment of them um then yeah they would they would be broken or um in some way.”

Claim: First-shipment hardware units may have had faults, so customers receiving them could have received broken or otherwise problematic equipment.

“Um so when you get a cluster um a good cloud, what it does is it tries to tries to kill all the GPUs, um before they ever get in the hands of the customer. Um and then um the GPUs that are broken or the nodes that are broken, um you get R_M_A_d. You return it, yeah. you return it and you get a new one and there's a lead time to that, but whatever, you'll get you'll get some back. Um and then the customer never has to experience this. Um the um thing I'm referring to is that”

Claim: A cloud provider can burn in hardware by testing it and returning faulty GPUs or nodes for replacement before customers receive them, shielding customers from those failures.

“um like in early h100s there were like firmware bugs that may not even be related to the specific like they may not be nvidia they might be like super micro has a firmware bug in the um in the box that you bought um and it applies to all of them and it doesn't matter like um the solution to that was you need to get them to update their firmware um or they need to ship a new thing in every new box”

Claim: Some early H100 system failures were caused by firmware bugs in the server box, potentially from a vendor such as Supermicro rather than Nvidia, and required a firmware update or replacement boxes with corrected firmware.

“um like is in the broader market, um clusters are financed three to five years. Um so the the the payments are all get over that period of time.”

Claim: Data-center clusters are financed over three to five years, with payments spread across that financing period.

“Um there is also like blocks of capacity i in our order book um which can extend any arbitrary time.”

Claim: The speaker's order book includes capacity blocks with durations that can be extended for an arbitrary period.

“So we care more about longer out, because we're trying to build a forwards curve that has a lot of liquidity on it. Um the”

Claim: SF Compute prioritizes longer-dated compute markets to build a liquid forward curve.

“since A, you can buy an hour out. You can buy right now to the next hour. Or you can buy right now to the next three years, anywhere in between. That's wild. I know. Um and so the um you know or a five-year, actually technically, um the”

Claim: Customers can purchase compute for delivery horizons ranging from the next hour to as far as five years.

“Uh that's a batch inference basically, where I can be interrupted at any point, I don't really care. Um then My ability to be cut off is probably like I need an hour of time to know that I'm going to be cut off and then I'm good. Or I can even be notified in, you know, a minute or two.”

Claim: Interruptible batch-inference workloads may require roughly an hour's notice of interruption, though some can tolerate notice of only a minute or two.

“But if I have live inference customers who are expecting me to be online, then. You can interrupt me, but in a month, maybe two months, maybe three months. Um and then I can adapt to that um capacity over that period of time. Um but you cannot interrupt me in a minute. Um and so those two things are priced differently and the reason why they're priced differently is because what is the price you would pay um in order to interrupt your batch job. It's very different than the price you would pay to interrupt your um live job.”

Claim: For live-inference workloads whose customers expect availability, interruption may require one to three months' notice, allowing capacity to be adapted over that period; a one-minute interruption notice is not acceptable.

“Um and then I can adapt to that um capacity over that period of time. Um but you cannot interrupt me in a minute. Um and so those two things are priced differently and the reason why they're priced differently is because what is the price you would pay um in order to interrupt your batch job. It's very different than the price you would pay to interrupt your um live job.”

Claim: The price of interruptible compute differs between batch and live workloads because customers value avoiding interruption differently for each workload type.

“job. If you have a line job, you're you are probably paying an S_L_A_ when you break for your customer. Um does that make sense?”

Claim: Interrupting a live job may breach the customer's SLA and carry a cost.

“These are the market for breaking leases basically. And then there's also the market for future leases or like for leases that are um”

Claim: Compute markets include a market for breaking existing leases and a distinct market for future leases.

“is in power markets. There's actually two markets. There's like the day ahead market and there's the live market. Compute kind of has the same thing where it has capacity that is coming online. Capacity is about to arrive to you and then capacity that is currently in use. And these have different price dynamics.”

Claim: Compute markets distinguish between capacity that is coming online and capacity already in use, analogous to day-ahead and live power markets, and the two have different price dynamics.

“Oh this is an ex both at different times. Yeah, I mean both it depends on who you are. Um”

Claim: Compute forward-curve shape varies by buyer and can differ across time.

“We see uh like a hump.”

Claim: SF Compute observes a hump-shaped compute price curve across delivery horizons.

“Uh well, we're more saying like if there's a block of compute and it's approaching now um it can't be sold in the past and so it has to price in order to clear at some price”

Claim: As a block of compute approaches its delivery time, it must be priced to clear because the opportunity to sell it cannot be carried into the past.

“and so i initially you see this like drop where um stuff that's sold right now anything that's on the market right now um is something that hasn't been sold and so therefore the price will drop. Um and then it goes up um to where it wants to be. Um”

Claim: Unsold compute available for immediate delivery tends to have its price reduced to clear, after which the price rises toward its target level.

“And then maybe compute in the future is worth less because people don't want like people would like the compute now. And so they're sort of paying for the shorter lead time and they pay less for the compute that's off into the future.”

Claim: Future-dated compute may be priced below near-term compute because buyers prefer shorter lead times.

“This is what I think a normal healthy curve looks like.”

Claim: The speaker considers a curve in which near-term compute is more valuable than farther-future compute to be a normal, healthy compute price curve.

“The place where I'm like, oh man, no shit, um is that um right now, if you took a B_ three hundred and you trying to run inference on it, um you can make forty dollars an hour or so um selling that inference. Any model basically. Um and so that's if you use it at a hundred percent. And maybe if you only have like fifty percent utilization or something, so like twenty five dollars an hour.”

Claim: The speaker estimates that a B300 used for inference can generate about $40 per hour at full utilization and about $25 per hour at roughly 50% utilization, for inference on any model.

“So twenty five dollars an hour. Meanwhile you can buy a B_ three hundred at like five to six dollars an hour, Has to change.”

Claim: The speaker says a B300 can be purchased for about $5–$6 per hour, compared with the higher hourly inference revenue estimate, and concludes that one of those prices must change.

“Yeah, so um a bunch of people are basically looking at this and thinking um if I think prices are going to go up, I might be willing to pay more than what I previously thought in the future um because I might think okay, I can buy a five right now. But if I can't buy any more, then maybe I will be willing to buy even more in the future just to lock it in if it's gonna come online later. So we see the other side as well. Um”

Claim: Buyers who expect compute prices to rise may pay more for future-dated compute to secure supply when they cannot purchase enough compute in the present.

“Kind of depends on who you are.”

Claim: A buyer's willingness to pay more for future compute depends on the buyer's circumstances or identity.

“W uh we care about different generations of chips because”

Claim: SF Compute distinguishes between chip generations in its compute market or pricing considerations.

“Uh yes, I think people totally care about the generation of chips. Um uh I think part of this is because of things like B_R_a_M_ um and part of this is because of like um uh larger N_D_ link domains uh and things like that.”

Claim: Customers care about chip generation, in part because of technologies such as HBM and larger NVLink domains.

“I think like when you were saying oil, even oil isn't oil. Oil is actually just like a bunch of different types of oil.”

Claim: Oil consists of multiple different types rather than being a single uniform product.

“Whatever. I think it's the same thing. Um you're gonna get a bunch of different types.”

Claim: The speaker considers compute analogous to oil because both come in different types.

“Yes, we reduce a lot of the dimensions, but nonetheless there are still dimensions. Yeah. Like it's still gonna exist.”

Claim: The company has reduced many dimensions of the compute-market problem, but some dimensions remain.

“we actually made SKUs like a a concept um that and the goal is to make it so that somebody else can make their own SKU. So if I wanna buy Blackwells, I can say I'm willing to buy Blackwells at”

Claim: The company designed its SKU concept so that other people can create their own SKUs.

“Six dollars an hour less, and you can match me against G_B_ three hundreds or you can match me against B_ three hundreds, and I don't care which black well, as long as it's less than six dollars an hour. Um”

Claim: A buyer can define a SKU that accepts Blackwell capacity below a specified price—illustrated as six dollars per hour less—and allows matching with GB300 or B300 capacity without specifying which Blackwell, provided the price meets the threshold.

“does. So some customers like um just can work on black holes, whatever it is. Some customers are super-hyper-specific and they need this exact thing, and they're gonna pay more for that because the market's can charge 'em more.”

Claim: Some customers accept broadly specified Blackwell capacity, while customers requiring an exact configuration pay more because the market can charge a premium for that specificity.

“Yes, um uh with the caveat that having a named query is the new thing we're doing um but the query system as a concept is how the order book works right now.”

Claim: The company’s query system currently underlies its order book, while named queries are a new feature in development.

“Um, mainly just the requirement for physical settlement.”

Claim: The speaker identifies physical settlement as the main requirement to understand in the compute-trading market.

“Um, we think that if you we s we spent three years learning why you need physical settlement. Um and um our entire conversation today is um that it is th these things are not as fungible as you think. We're gonna try and make 'em fungible, that's the industry's job to do that, um but without that you run a lot of risks. Um if you trying to build”

Claim: The company spent three years learning why physical settlement is necessary for compute markets.

“Um, we think that if you we s we spent three years learning why you need physical settlement. Um and um our entire conversation today is um that it is th these things are not as fungible as you think. We're gonna try and make 'em fungible, that's the industry's job to do that, um but without that you run a lot of risks. Um if you trying to build”

Claim: Compute capacity is less fungible than people assume, and the industry should work to make it more fungible.

“Um, we think that if you we s we spent three years learning why you need physical settlement. Um and um our entire conversation today is um that it is th these things are not as fungible as you think. We're gonna try and make 'em fungible, that's the industry's job to do that, um but without that you run a lot of risks. Um if you trying to build”

Claim: Compute markets carry substantial risks without physical settlement.

“build too much financialization before you've the industry has had time to build out the the physical settlement then I think what you have is you have a market that's very easy to manipulate and a market that has a lot of basis risk where the price you may be hedging against actually has nothing to do with the price that the market you can actually buy the capacity for.”

Claim: If compute-market financialization expands too far before physical settlement is established, the market will be easy to manipulate and will have substantial basis risk.

“a market that's very easy to manipulate and a market that has a lot of basis risk where the price you may be hedging against actually has nothing to do with the price that the market you can actually buy the capacity for.”

Claim: When compute markets have substantial basis risk, a hedging price may be unrelated to the price at which capacity can actually be purchased.

“Yeah. So to be clear, we think um cash settlement as a future concept is a great idea. Um we just think it should be derived from physical settlement.”

Claim: Cash settlement could be valuable in the future, but it should be based on physical settlement.

“Not yet, totally. people are trying to make cash settlement before you make the physical settlement.”

Claim: Some market participants are attempting to establish cash settlement before physical settlement exists.

“Um and lots of those go very poorly. Um like it would be kind of bad if you made”

Claim: Many cash-settled products perform very poorly.

“It would be bad if you started selling the derivative on a stock that doesn't exist yet.”

Claim: Selling derivatives on a stock before the underlying stock exists would be harmful.

“So critically, LIBOR was manipulated.”

Claim: LIBOR was manipulated.

“yes, um well our concern is there's a very similar problem, um where you can manipulate an index. Um our impression is why you can manipulate that index is because a lot of the indexes are based on a very small amount of data that has um Not a lot of correlation with the actual price of buying a new reserved contract. That's what we're concerned about.”

Claim: The speaker's concern is that indexes may be manipulable because many are based on very little data that has little correlation with the actual price of buying a new reserved contract.

“Uh, so we're big fans of silicon data and compute exchange. Um I think they were also very early to this, um and I think they are doing like a reasonable shot at the a path that we think um that will likely end up being partners with. Um”

Claim: The speaker's company is a strong admirer of Silicon Data and Compute Exchange, considers it an early entrant, and thinks its approach may lead to a partnership.

“So um uh the general gist is they're doing cash settlement, we're doing physical settlement. Um we think both can exist and should exist. Um and that they're sort of mutually beneficial to each other.”

Claim: Silicon Data and Compute Exchange use cash settlement, while the speaker's company uses physical settlement; the speaker believes both approaches can and should coexist.

“Um and so totally fine. Um so I think they are one of the earliest in the space and I think they are um very clearly took the bet um well before anybody else did and I think that's like very honourable to do. Um, um you mentioned uh jump”

Claim: The speaker considers Silicon Data and Compute Exchange among the earliest entrants in the space and regards its early commitment as honorable.

“So the categories of folks that we've seen is there's folks that are trying to set better standards. So semi-analysis, I think, is working to try and create better standards.”

Claim: Some companies are working to establish better standards in the compute ecosystem; the speaker identifies SemiAnalysis as one such effort.

“There's folks that are trying to do residual value insurance and so on.”

Claim: Some companies in the compute ecosystem are working to provide residual-value insurance.

“There are folks that are trying to do deposit lending where they'll front. your credit risk for the cloud.”

Claim: Some companies are developing deposit-lending services that front customers' credit risk for cloud services.

“There's an ecosystem of companies that are reforming around how do we resolve the financial bottlenecks of compute, um and it is cool to see.”

Claim: Companies are forming an ecosystem to address compute's financial bottlenecks, which the speaker views positively.

“Um we do all the things the cloud does without being the cloud. Um as in we don't own the hardware. Um the reason why we don't own the hardware is because if we did, we would be able to manipulate the forwards market. So We have to be Switzerland, we have to be independent,”

Claim: The speaker's company provides cloud-like functions without owning hardware because hardware ownership would let it manipulate the forwards market; the company therefore needs to remain independent.

“But in order to do physical settlement, you have to do site visits, you have to visit the data center,”

Claim: Providing physical settlement requires the company to conduct site visits, including visits to the data center.

“you have to, you have to design the bill of materials and ensure that You have to basically do like supply chain consulting where you have relationships with different OEMs that ensure that things get delivered on time.”

Claim: Physical settlement requires designing the bill of materials and managing supply-chain relationships with OEMs to ensure equipment is delivered on time.

“Then you have to do cluster architecture. Uh so you're sort of designing the cluster itself. Um then you're um rolling out racks or you're working with partners to roll out racks and then supervising that. Um then you're doing uh back-plane management. So you're taking over the um uh P_D_Us and the switches and the D_P_Us and um U_F_M_ and B_M_C_ and you're operating um the whole um harness around the nodes you're selling.”

Claim: The speaker's company designs cluster architecture, deploys or supervises rack deployment, and operates supporting infrastructure around compute nodes.

“but there's all this shit around the bare metal. Um and in order to make that bare metal node work, you have to operate everything else. Then we”

Claim: Operating a bare-metal compute node requires the speaker's company to operate the surrounding infrastructure as well as the bare-metal hardware.

“We're we're not in the data centre business sort of. Um what we do work with is So we offer supercomputing as a service.”

Claim: The speaker's company offers supercomputing as a service.

“And what this means is folks come to us and say, you know, I've got X, where X is I've got, you know, X megawatts here. And maybe sometimes they mean I've got a data center. That's like I would like to turn into a GPU cluster.”

Claim: Customers approach the speaker's company with power capacity or a data center that they want converted into a GPU cluster.

“We will turn you into the cloud. We're the supercomputing part. You're the data center clerk. Let's partner together and then we will sell and all make money.”

Claim: The speaker's company partners with data-center owners by providing the supercomputing layer while the partner supplies the data center, then they jointly sell the service and make money.

“Sometimes it's people who have something that could turn into a data center, in which case we're going to basically project manage that up to the layer. And then we will work with those folks to turn their powered land or whatever they have into a data center by working with partners. And then we take over at the supercomputing part.”

Claim: When customers have land or other assets that could become a data center, the speaker's company project-manages development with partners and then takes over the supercomputing layer.

“This requires us to have some sort of data center expertise in-house. house, but we don't self perform builds. Um so yeah.”

Claim: The speaker's company needs in-house data-center expertise but does not perform data-center construction itself.

“but and we like uh we we deeply partner with folks on both sides.”

Claim: SF Compute says it deeply partners with participants on both sides of its market.

“primarily the U.S. but we are operate around the world.”

Claim: SF Compute operates primarily in the United States and also operates around the world.

“we're deploying clusters all around the world,”

Claim: SF Compute is deploying compute clusters around the world.

“and when we do that we're working with people who themselves are trying to be clouds all around the world. Um they look like sovereign compute players. Um and our pitch to sovereign compute is amazing. Uh let's help you set up a cloud in your own country, um and we will bring the San Francisco off take.”

Claim: SF Compute offers sovereign-compute players help setting up a cloud in their own country, while bringing San Francisco offtake.

“and when we do that we're working with people who themselves are trying to be clouds all around the world. Um they look like sovereign compute players. Um and our pitch to sovereign compute is amazing. Uh let's help you set up a cloud in your own country, um and we will bring the San Francisco off take.”

Claim: SF Compute works with organizations around the world that are trying to become cloud providers and appear to be sovereign-compute players.

“We'll bring the San Francisco expertise, we'll set up the clusters, um that's our whole business.”

Claim: SF Compute says its business is bringing San Francisco expertise and setting up compute clusters for partners.

“Nobody I can name, basically.”

Claim: The SF Compute speaker says there are essentially no partners they can publicly name.

“Uh we will have some announcements um in a week or two um that we expect to come on.”

Claim: SF Compute expects to make some announcements within one or two weeks of the conversation.

“Uh yeah, so our impression is that Nvidia can't make it, but benefits tremendously from it existing. The reason why they can't make it is because they would then not be independent enough.”

Claim: SF Compute's impression is that Nvidia benefits greatly from the market's existence but cannot create the market itself.

“Uh yeah, so our impression is that Nvidia can't make it, but benefits tremendously from it existing. The reason why they can't make it is because they would then not be independent enough.”

Claim: The speaker says Nvidia could not create the market because it would lack sufficient independence for participants to trade on it.

“for people to trade on it. If you don't trade on it, um then you don't have enough liquidity for lenders to finance against the forwards market.”

Claim: The speaker says that without trading on the market, liquidity would be insufficient for lenders to finance against the forwards market.

“This is also why Crusoe and Korweave and Lambda, why the hyperscalers can't make it, why base ten and fireworks and modal can't make it, because they have too much demand on that side.”

Claim: The speaker says Crusoe, CoreWeave, Lambda, hyperscalers, Baseten, Fireworks, and Modal cannot create the market because they have too much demand on the market's demand side.

“This is a thing that benefits the entire industry because it makes the financing costs go down.”

Claim: The speaker says the market benefits the entire industry by reducing financing costs.

“But nobody can make it unless they give up the ability to do all the big inference money and all the big cloud deals.”

Claim: The speaker says a company can create the market only by giving up its ability to pursue large inference revenue and large cloud deals.

“You have to kill your business in order to make this.”

Claim: The speaker says a company would have to sacrifice its existing business to create the market.

“We've closed several hundred million dollars of volume.”

Claim: SF Compute says it has closed several hundred million dollars in transaction volume.

“We have a significant number of new clusters on the way.”

Claim: SF Compute says a significant number of new compute clusters are in development or deployment.

“shipped a whole bunch of new features um that make S_F_ compute very viable for platform as a service companies. Um an example is um we allow customers to subdivide their InfiniBand fabric um without having to ask us. So many clouds, when you go to buy them, they will give you one InfiniBand fabric, um and then if you want to reep”

Claim: SF Compute says it has shipped numerous features intended to make its service viable for platform-as-a-service companies.

“shipped a whole bunch of new features um that make S_F_ compute very viable for platform as a service companies. Um an example is um we allow customers to subdivide their InfiniBand fabric um without having to ask us. So many clouds, when you go to buy them, they will give you one InfiniBand fabric, um and then if you want to reep”

Claim: SF Compute allows customers to subdivide their InfiniBand fabric without requesting assistance from SF Compute.

“shipped a whole bunch of new features um that make S_F_ compute very viable for platform as a service companies. Um an example is um we allow customers to subdivide their InfiniBand fabric um without having to ask us. So many clouds, when you go to buy them, they will give you one InfiniBand fabric, um and then if you want to reep redraw the lines of that infinite band fabric. Uh y you need to contact them and they will then do something for you. Um we now can do that on the fly for customers, um which I think is really great, um because it interfaces with the market really well, um where you want to be able to redraw things in order to scale up and down your clusters um Would we are”

Claim: The speaker says many cloud providers supply a single InfiniBand fabric and require customers to contact them to have its configuration redrawn.

“redraw the lines of that infinite band fabric. Uh y you need to contact them and they will then do something for you. Um we now can do that on the fly for customers, um which I think is really great, um because it interfaces with the market really well, um where you want to be able to redraw things in order to scale up and down your clusters um Would we are”

Claim: The speaker says on-the-fly fabric reconfiguration fits the market because customers need to adjust configurations as they scale clusters up or down.

“redraw the lines of that infinite band fabric. Uh y you need to contact them and they will then do something for you. Um we now can do that on the fly for customers, um which I think is really great, um because it interfaces with the market really well, um where you want to be able to redraw things in order to scale up and down your clusters um Would we are”

Claim: SF Compute says it can reconfigure customers' InfiniBand fabric on the fly.

“Yeah, we're virtualising the data centre and making it programmable.”

Claim: SF Compute describes its work as virtualizing the data center and making it programmable.

“So all the stuff that you would have gotten if you had owned the hardware, we're trying to virtualise virtualize that.”

Claim: SF Compute says it aims to virtualize the capabilities customers would have if they owned the hardware.

“If you would like to own a GPU cluster and you want the residual value from it, you want to have it for your portfolio, we would love to work with you in order to build that GPU cluster.”

Claim: SF Compute invites people who want to own a GPU cluster and retain its residual value in their portfolio to work with the company.

“After we will take over and we will uh do the superconducting portions of that and ensure the things go right, we own the S_L_A_ and then if it turns out that that was a bad idea, that you need out of that cluster, we can immediately sell that cluster to somebody else, so it doesn't hit your balance sheet.”

Claim: SF Compute says it will take over the cluster's superconducting portions, oversee their operation, and hold the service-level agreement after the customer builds the cluster.

“After we will take over and we will uh do the superconducting portions of that and ensure the things go right, we own the S_L_A_ and then if it turns out that that was a bad idea, that you need out of that cluster, we can immediately sell that cluster to somebody else, so it doesn't hit your balance sheet.”

Claim: SF Compute says that if a customer needs to exit a cluster, it can immediately sell the cluster to another buyer, preventing it from affecting the customer's balance sheet.