可编程数据中心——Evan Conrad，SF Compute
- SF Compute的核心押注是，要建立可交易的GPU市场，就必须运营交付算力所需的基础设施。 Conrad称，公司花了3年才弄清实物交割为何重要：“这些东西并不像你以为的那么可互换”。即便是2种VM配置，性能也可能相差40%；在交付尚不可靠时先搭建金融合约，容易引发操纵，也会让对冲价格脱离实际可成交价格。
- SF Compute称，目前仅有5%的容量挂在订单簿上，但转售流动性已能切实满足客户需求。 客户每花100美元，平均可通过转售容量回收25美元，这是流动性是否真正帮到商业用户的实际指标。Conrad另称，公司已成交数亿美元交易量，但订单簿上的成交量仍然不大。
- Swyx对B300推理收入的估算，揭示了本期最尖锐的定价矛盾。 Swyx估算，利用率达到100%时，每小时推理收入约40美元；利用率约为50%时，收入约25美元，而租用成本为每小时5–6美元。Conrad认同这是“严重超额收费”；预期算力价格上涨的买家，可能愿意多付钱锁定未来供应。
- SF Compute希望建立一条流动性充足的远期曲线，为客户放弃容量的权利定价。 买家可以购买从1小时到3年的容量；批处理任务或许能接受提前几分钟通知，在线推理则可能需要提前1至3个月。Swyx形容价格走势有个“鼓包”：临近交付却尚未售出的容量，会像临期牛奶一样打折；交付时间更近的容量，也可能比远期交付更贵。Conrad还指出，预期价格上涨的买家可能愿意为未来容量支付更高价格。
- Conrad认为，独立性是把算力交易转化为更低融资成本的前提。 在他看来，NVIDIA、云厂商和大型推理买家都不够独立，难以让人放心在其平台交易。SF Compute不持有硬件：“我们必须像瑞士一样保持中立”（“we have to be Switzerland”），由此形成的流动性才能成为贷款人融资的依据。
- 可编程数据中心为资产所有者提供GPU投资的运营层。 客户可根据容量需求变化自行拆分InfiniBand网络，SF Compute则负责运营基础设施并承担SLA。其主权算力方案把本地集群与旧金山的需求和专业能力结合起来：“我们要让它具备融资条件”（“We’re going to make it financeable.”）。
1. 实物交割离不开云厂商式的运营能力
Conrad讲起创业缘起：最初是转租一座集群，后来逐渐接手更多责任，就像卖酒店客房的人被要求负责清洁、客房服务和安保，“到了一定时候，你其实就是在经营整家酒店”。因此，SF Compute搭建了云厂商所需的运营能力，确保算力能够可靠交付。“实物交割”借用了期货市场的概念：交易一份合约，和真正收到石油或大豆，是两回事。
订单簿和交付系统是围绕算力的实际约束共同设计的。Conrad拿电力市场举例：订单会纳入锅炉冷却时间等限制；算力客户可能只有在配备能让工作负载快速恢复的存储时，才接受中断。合约标准化“还差得很远”。这段经历让他们得出一个结论：“我们得把整套东西都运营起来”——先实现一致、可靠的运营，才能形成标准化产品。
Conrad起初称这个市场“还不成熟，深度和流动性也不足”，随后区分了订单簿成交量偏小与已挂牌容量内部流动性集中的情况。目前，订单簿上只有总容量的5%，但更多容量正在挂牌。客户每花100美元，平均可通过转售收回25美元。他称公司已成交数亿美元交易量；他更看重的流动性检验标准是：“这有没有解决客户的问题？”
2. 硬件差异让算力无法成为单一、可互换的产品
Swyx以大宗商品分级作类比，但实际工程问题不小：不同运营商的VM配置，性能可能相差40%。Conrad将差异归因于GPU内核启动延迟，最终会拉低MFU。因此，要判断算力是否等价，不能只看GPU型号，还必须了解基础设施实际交付的性能。
SF Compute会按约束条件撮合买卖，不等固定标准成形。买家可以接受每小时6美元以内的任意Blackwell，系统再匹配GB300或B300；另一些买家则要求特定显存或NVLink域。Swyx认为，暴露这些差异会让做市更困难。Conrad认同，产品仍有多个维度：灵活度带来套利机会，指定条件则要付出更高成本。按名称查询的功能正在推出，基于查询的撮合已经在运行。
存储和网络连接决定了容量能否真正互换。Conrad提到WEKA或VAST存储系统、存储容量，以及数据能否快速迁移、让工作负载在其他地方重启。这需要高速传输通道，通常是WAN。GPU集群选址跟着电力走，而传统CPU集群会跟着客户走，以满足延迟要求，因此较新的选址可能缺少站点之间的强连接。
即便集群的部署批次也可能影响匹配，不过按批次撮合目前还只是讨论中的想法。Swyx问，常规老化测试为什么不能筛掉缺陷。Conrad区分了单颗GPU故障与可能影响多套系统的固件问题：早期H100遇到的问题，可能出在Supermicro机箱的固件，而非GPU本身。测试并更换坏硬件解决不了这类问题；需要更新固件，或在新机箱中修复。
3. 远期曲线同时为临期容量和中断权定价
Conrad回答“3到5年”时，最初谈的是更广泛市场中集群的融资周期。SF Compute的订单簿支持任意时长的容量块，购买周期从1小时到“未来3年及其间任何时点”，技术上还可以切得更细。目前，容量块需要全额预付，长期租赁则不需要。他的首要目标是建立一条流动性充足的远期曲线。
已经在运行工作负载的容量，与下个月才启动的容量不同；后者的交易不会打断当前任务。批处理推理或许能接受提前1小时、甚至1到2分钟通知；在线推理可能需要提前1、2或3个月，因为中断可能导致客户要求支付SLA赔偿。Swyx意识到这其实是一个“打破租约”的市场，点出了中断权为何会有不同价格。
被问到曲线呈现反向市场还是正向市场时，Conrad先确认Swyx指的是远期容量价格高于还是低于现货。Swyx说：“我们看到一个鼓包”，并用临期牛奶作比：即将交付、尚未售出的算力必须打折，因为“过去的东西没法再卖”。他认为，价格随后会回升至他们偏好的水平；远期容量的价格也可能更低，因为买家更看重交付时间较近的容量。Conrad认可这套解释，随后补充说，预期价格上涨的买家可能愿意为更晚交付的供应支付更高价格。
Swyx估算，B300以100%利用率出售推理服务，每小时可产生约40美元收入——“基本上什么模型都行”；利用率约为50%时，收入约为25美元，而租用成本是每小时5–6美元。Swyx说：“这几个数字里总有一个得变。”Conrad认同这是“严重超额收费”。因此，预期算力价格上涨的买家可能愿意为未来交付支付更高价格，以锁定供应。
4. 现金结算需要可信的实物市场作支撑
Conrad的警告来自3年来对交付问题的摸索：实物交割尚未建立，就急于把算力金融化，可能导致市场操纵和巨大的基差风险。用于对冲的价格，可能与客户实际能买到容量的价格相差甚远。按他的说法，让算力更具可互换性，是整个行业的任务。
Swyx提出货币和利率市场的常见模式：参与者约定一个类似LIBOR的定盘价。Conrad立刻指出，LIBOR曾被操纵。他担心算力指数也会依赖规模很小的数据集，且与购买新的预留合约价格相关性不高。仅仅约定一个基准，并不能解决底层测量问题。
Swyx认为，如果现金结算建立在实物交割之上，它可以成为一个很有前景的未来概念。Conrad则把现金结算合约比作《大空头》里的那种旁下注，可能与标的资产本身关系不大。他说，有些领域已经有人在做这类押注；Swyx则提醒，不要对一个尚未真正存在的标的做衍生品交易。
Conrad称，SF Compute非常看好Silicon Data和Compute Exchange，并将它们称为早期入场者。他认为，对方走现金结算路线、SF Compute走实物结算路线，两者可以相互促进，也可能展开合作。
周边生态正在形成不同分工。Conrad提到，SemiAnalysis在推动更好的标准；此外还有残值保险，以及为云厂商垫付信用风险的押金贷款。他承认自己对个别服务商了解有限，但更广泛的趋势是，围绕算力金融瓶颈的生态正在形成。
5. 独立性支撑可编程数据中心业务
Conrad的判断是，NVIDIA能从这个市场受益，却很难可信地运营它；Crusoe、CoreWeave、Lambda、超大规模云厂商、Baseten、Fireworks和Modal也存在类似利益冲突。它们都不够独立，难以让交易者放心在其平台上交易；没有交易，就没有足够流动性让贷款人以远期市场为依据融资。而充足流动性可能降低整个行业的融资成本。因此，SF Compute选择不持有硬件。搭建这个交易平台，意味着放弃利润丰厚的云和推理业务：“要做成这件事，就得先把自己的业务干掉。”
运营工作涵盖现场考察、物料清单、协调OEM确保设备按时交付、集群架构、机架部署和背板管理。如今，客户无需再让云厂商代劳，就能自行拆分InfiniBand网络，从而随着容量交易扩容或收缩。他们希望将“拥有硬件时本来能获得的所有东西”虚拟化，让数据中心变得可编程。
资产所有者可以提供数据中心、兆瓦级电力或已接通电力的土地；SF Compute会与合作伙伴协调开发设施并接手超算运营，但不自行施工。对于希望获取GPU集群残值的投资者，Conrad提供运营保障——“SLA由我们负责”。如果投资者需要退出，SF Compute可以立即把容量卖给其他人，因此这部分资产不会留在他们的资产负债表上。
这项业务主要在美国运营，也与希望成为本地云厂商的合作伙伴在全球部署。主权算力客户获得集群搭建、旧金山的专业能力，以及旧金山的算力需求方。Conrad将运营服务与融资直接相连：“我们要让它具备融资条件。”
完整逐字稿
1. Why NVIDIA can’t build this market
NVIDIA is obviously the 800-pound gorilla in the room. Why didn’t they make a supercomputer?
Our impression is that NVIDIA can’t make it, but benefits tremendously from it existing. The reason they can’t make it is that they wouldn’t be independent enough for people to trade on it.
2. Welcome back to Latent Space
Anyway, Evan, welcome back to Latent Space. Welcome back to AIE. Actually, is this your first time at AIE?
I think so, yeah.
This is a very special one because I feel like you’ve been working on compute trading and compute markets for a while, and now it feels like the moment. Jump basically announced that they’re getting into compute markets, which is why I’m excited to sort of cluster people. You met Oliver from SemiAnalysis, who is basically initiating coverage on compute markets. This is my read on it, so people are going to watch your—
Talk.
I think it’s a very good explanation of both your origin story and why you need to exist, and the hard problems you work on.
Yeah.
It’s done in a very simple, engaging manner. I’m very jealous of your writing skills, because you showed the flyer you had at NeurIPS. It was pretty straightforward and simple, and it was good.
Yeah.
Anyway, congrats.
Thank you. I appreciate that. And congrats on AIE. It seems like you all are rolling.
We’re rolling. I just came back from the site tour for next year’s AIE.
Give me a hint.
It’s at Javits Center, the same place they do Comic-Con.
Okay, wow.
Which—
Wow. Wow. Damn.
I think the finance focus is really hitting for folks.
Yeah.
We’re going to try to split things up. Here, we just lump everything together in finance: your finance, their finance, insurance is finance. But they’re all super different.
Yeah.
So we’ll probably break things out more minutely in the future, because people want to go deep on hedge fund stuff, banking, personal wealth, or—
Yeah, especially in New York.
Yeah. Anthropic was like, “We’re all about RIAs right now.” RIAs, or registered investment advisors.
Yes.
That’s so different from you.
Yeah, no, I’m aware of this.
Yeah, yeah. Okay, so for listeners who don’t know, we did a podcast 2 years ago where you talked about where SF Compute was.
Mm-hmm.
3. From selling a room to running the hotel
Today, you’re much more into this sort of physical settlement. You’re talking about your backplane—what do you call it?
The backplane.
The backplane, which obviously takes a lot of hardcore engineering. Can you give a status update on where SF Compute is, for listeners who last heard you on a podcast 2 years ago?
When we started, we were this company called Junnar, and we were just trying to sublease our cluster. We started there in the technical stack. One way to think about it is that you’re in a hotel and you’re trying to sell that hotel room. Somebody asks you, “Hey, are you going to clean my room?” And then you say, “Oh, I guess I probably also need the hotel room, or I need the cleaning closet.”
So you get access to the cleaning closet. Then they want room service, so you get access to the kitchen. Then they want security, so you get access to the security closet. You get this, and you do that, and you do this. At a certain point, you’re just running the whole hotel.
The moment you’re running the whole hotel, you say, “Oh, wait a second. My supplier in 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.” What we’ve spent the last several years doing is building the entire operating capacity of a cloud in order to actually deliver the compute at a reliable level.
4. Physical settlement: actually delivering compute
We call that physical settlement. We’re taking a term from the futures markets, from other markets. In oil or soybeans, physical settlement means you get delivered the oil and the soybeans.
Yes.
In some markets, like oil and power, there’s a delivery network. In oil, you have to have pipelines. If you’d like the oil to be delivered to you, there has to be a way for it to get to you.
Yes.
And for people interested in financial history, during COVID, oil went negative because people were saying, “I will pay you to take oil from me,” since it was sitting there in the port and they couldn’t pay for it.
Yes.
And so I will pay you to take my oil.
Yes.
That’s the weirdness of physical settlement.
Yes. There are a bunch of weirdnesses like this in other markets as well. In power markets and the grid, parts of the orders that buyers and sellers fill out are there because the boiler takes a while to cool off. You can stop the boiler and start it back up again within some period of time, so that has to get factored into the order book. This Junnar thing is the same thing that happens in compute markets. We co-designed physical settlement and the order book at the same time, which meant we could handle the real constraints of compute. That might include something like, “I’m okay with being interrupted, but only if there’s storage on the thing that I can recover from really quickly.” They’re very similar types of problems.
Yeah. People can refer to your talk, and you’ve done a couple of other podcasts already. I want to be sort of 201 level instead of 101 level, partly because I’ve been a derivatives trader, so I can give people a bit more detail than you normally get into. For example, how mature, deep, or liquid is the compute market today?
5. Liquidity and the GPU order book
Not very mature, deep, or liquid. First, we had to get a bunch of supply online, which just takes a while, and then we had to build the entire platform to do this. We just got that working, so now we’re moving capacity onto the order book, which makes it liquid.
Five percent of our total capacity is on the order book, and we’re consistently moving more and more. Then we’re bringing on more capacity that’s all on the order book. The nice part is that maybe I should say the volume on the actual order book is small, but the liquidity on the order book is high. For that small amount of capacity, you have very concentrated liquidity.
A standard metric for someone using this would be a customer who buys capacity from us: if they spend $100 to buy that capacity, they’ll make back $25 on average by selling it back to somebody else. That’s one of the metrics we try to track, because we want the liquidity to be useful to the commercial participant. You can measure liquidity in a bunch of different ways. You can say, “Oh, I just want tons of makers,” or measure it in high-minded ways. We try to measure it by asking, “Is this solving the problem for the customer?”
Yeah, got it. Last time we talked, there was a lot of conversation about standardizing the contracts. What’s the status on those things? You were talking with the CFTC or whatever.
I think standardizing the contracts in general is still pretty far away. What we learned over the last several years was that we just needed to operate the whole thing; otherwise, you cannot standardize it. To get closer to something that can be standardized, you first need to make all the operations of the thing the same.
To give an example of what can be different: a single operator can offer 2 different types of virtual machines. One type of virtual machine might have a 40% performance degradation versus another type of virtual machine, and 40% is a lot.
When you say performance degradation, are you talking about MFU-type things, or the amount of burning you’ve experienced, or what?
It will affect—so, in this particular case, I’m talking about GPU kernel launches and how long it takes for that to happen, which is a process that happens quite frequently when you’re doing either training or inference.
Okay.
The end result is that you get a worse MFU, but—
Is the intermediate level—
Intermediate level, basically.
Yeah, yeah. So what I’m trying to say is, for anyone who’s traded commodities before, there are different grades of commodity, right? It’s not just oil; it’s light sweet crude versus whatever exactly the opposite of light sweet crude is, or corn grade A or corn B. People are trying to—well, you have to establish the equivalents.
The equivalents of those things.
6. Trading nonstandard compute
Right, yes. And I’m wondering if they have been established.
What we did instead was create an order book that had multiple different constraints, and you can say, “I want something that’s between this and that.”
It runs from this time to this time. It’s like X to Y. There are a bunch of different properties you can attach to your order, and then we match that against sell orders that have those properties in a big matching engine. The reason we did it this way is that it allows us to generate liquidity without having to standardize before we know what the technology can do. Because the technology was shifting so quickly, we wanted to be able to do that.
The trade-off is that you have less concentrated liquidity, so there are going to be wider spreads between the buy and the sell.
Yeah.
But on the other hand, there’s lots of arbitrage to make if you can do that correctly. If you have a product that can use lots of different types of capacity, you might really like SF Compute because you’ll be able to use lots of different types of capacity and make more money. You can take the spread.
Okay. What dimensions matter in this grid of things?
7. Interrupting and moving workloads
One category is whether there’s a workload running on it right now or whether it’s capacity in the future. Is it capacity right now? To sell capacity right now, you have to interrupt your current workload, so you have to think about whether you care if your current workload is interrupted. Capacity that’s being sold starting a month from now, you can buy and sell all the time, so you get liquidity on that really easily. There is some concern about whether you’ll have capacity next month, but it’s very different from interrupting a workload right now.
The way the interruption happens matters. For example, if you’re on a cluster and it doesn’t have as amazing a storage system as you want, do you have a WEKA or VAST system under the hood? Do you offer that, and what’s the size of it? This changes how you respond to being preempted or removed from the cluster.
Meanwhile, the location of that storage matters. If you’re able to move that data from one cluster to another quickly, maybe you can spread across multiple clusters. You can be interrupted over here and then pop back up over there. To make that work, you also need a fast transit path between one cluster and another. Typically, this is implemented with a WAN, but—
Most of these clusters don’t necessarily have a wide area network.
Yeah, a wide area network. Basically, cables that connect the clusters together. A private internet is another way to think about it. A lot of clusters being spun up today aren’t attached to traditional CPU clusters, which were next to the customer for latency reasons. With GPUs, they’re next to the power. You put the data center next to the power rather than next to the customer.
A lot of stuff is getting put in places where they previously weren’t expecting data centers, so you may not have as excellent a transit path from one cluster to another. That starts to matter.
Other examples are things we theoretically support but don’t actually support yet. This is something we’re thinking about but haven’t done, basically. The early GB200s broke. The early H100s broke. If you were buying H100s or GB200s, you might really want the ones that were deployed after a certain time, because you’re not getting the first GB200s off the rack. If you’re a sophisticated customer, you may be willing to pay more for that.
8. Why hardware cohorts matter
Oh, so it’s about what cohort they were produced in, and they get better over time?
The first ones off the rack may have had some faults or problems. If you got the first ones, or the first ones from the first shipment, then, yeah, they would be broken or have some problem.
I assume that sort of process is like, look, there are defects everywhere. It just happens, right?
Totally.
And so you get a new shipment from TSMC or whatever, run your tests, figure out which ones are bad, and throw them out. Isn’t that how it works? I didn’t know about this element of cohorting where, say, Q4 2026 B300s are way worse than 2027 B300s.
Yeah, I’ll give an example. You’re describing burn-in. When you get a cluster, a good cloud tries to kill all the GPUs before they ever get in the hands of the customer. The GPUs or nodes that are broken get RMA’d. You basically go and replace them. You return them and get new ones, and there’s a lead time, but whatever, you’ll get—
Yeah.
—you’ll get some back. Then the customer never has to experience this. The thing I’m referring to is that, in early H100s, there were firmware bugs that may not even be related to the specific GPU. They may not be NVIDIA’s; they might be a firmware bug in the Supermicro box you bought. It applies to all of them. The solution was to get them to update their firmware or ship a new thing in every new box.
Yeah. Cornelis had this for the B200s that we got, and then you had to use DevCloud to go fix it.
Yes, correct. Different types of problems like this can happen.
9. How far out do compute contracts go?
Interesting. Okay, that seems small compared to the overall picture. I want people to understand the shape of this market. You mentioned preemption, and we talked about spot instances and how that’s an analogy. There’s also booking a month out. How far out do these markets go? Is real-time preemption more interesting to you, or is it further out?
To answer your first question: 3 to 5 years.
Wow.
We’re starting to see—
Month by month or quarter by quarter? How does this work?
Are you asking how the cash flow works?
The contracts. The futures contracts, right? Again, I’m a former contracts and derivatives trader.
In the market itself—
Yeah.
—like in the broader market.
Yeah.
Clusters are financed for 3 to 5 years.
Okay.
The payments—
—are over that period of time.
Yeah. There are also blocks of capacity in our order book that can extend to any arbitrary time. There’s some stuff we have to do for the cash flow on that. Right now, we just ask people to pay entirely upfront for a block. We don’t ask them to do that for a lease, for a longer-term contract. This gets into something I definitely need a whiteboard to explain.
But you had another question I was about to answer. I forgot what it was.
Do you want real-time, or do you want something further out? What’s more interesting to you: real-time or further out?
We care more about further out because we’re trying to build a forward curve that has a lot of liquidity on it. These are kind of the same thing and kind of not. You can buy an hour out, from right now to the next hour. Or you can buy from right now to the next 3 years, and anywhere in between.
That’s wild.
10. Batch versus live inference leases
I know. And you can go even finer, technically. There’s a difference that depends on each customer. If I have a batch inference workload where I can be interrupted at any point, I don’t really care. I might need an hour’s notice that I’m going to be cut off, and then I’m good. I can even be notified in a minute or two.
But if I have live inference customers who are expecting me to be online, you can interrupt me, but it has to be in a month. Maybe 2 months, maybe 3 months. Then I can adapt to that capacity over that period of time, but you cannot interrupt me in a minute.
Those two things are priced differently. The price you would pay to interrupt your batch job is very different from the price you would pay to interrupt your live job. If you have a live job, you’re probably paying an SLA when you break it for your customer. Does that make sense?
Yeah, I see. So this is actually the market for breaking leases.
These are markets for breaking leases, basically.
God.
And then there’s also the market for future leases, or leases that are—
Oh, I thought this was a subletting situation. This is like—oh, the lease break.
One way to think about it is that, in power markets, there are actually 2 markets. There’s the day-ahead market, and there’s the live market. Compute kind of has the same thing, where there’s capacity that’s coming online, or capacity that’s about to arrive to you, and then capacity that’s currently in use. These have different price dynamics.
11. The shape of compute prices
Okay, interesting. Backwardation or contango?
Backwardation or contango?
Do you mean, is compute further out priced higher than spot or lower than spot?
Does compute further out go like this or like that over 1 year out, 2 years out? I find it fascinating that you don’t know.
Yeah, I mean, both. It depends on who you are. I’ll make the case—wait, so you’re asking whether a unit of compute sold today, or maybe a month out versus 1 year out versus 2 years out, is priced on a curve that’s going up or going down?
We see a hump.
Oh my God.
Okay, so—
Explain.
I think this will make intuitive sense. If I have milk and it’s about to expire, I’ll lower the price on it.
Yes.
That’s basically what we see.
So it’s the end of the lease cycle, basically, right?
Well, we’re saying that if there’s a block of compute approaching now, it can’t be sold in the past, so it has to drop its price in order to clear. Initially, you see this drop: anything that’s on the market right now is something that hasn’t been sold, so the price will drop.
Mm-hmm.
Then it goes up to where it wants to be.
Mm-hmm.
Maybe compute in the future is worth less because people would like the compute now, so they’re paying for the shorter lead time. They pay less for compute that’s further out in the future. This is what I think a normal, healthy curve looks like.
Yeah.
The place where I’m like, “Oh man, no shit,” is that right now, if you took a B300 and tried to run inference on it, you could make $40 an hour or so selling that.
Any model?
Any model, basically. That’s if you use it at 100% utilization. Maybe if you only have 50% utilization or something, it’s more like $25 an hour.
Yeah.
So, $25 an hour. Meanwhile, you can buy B300s at like $5 to $6 an hour. One of these numbers has to change.
Yeah. It’s a huge overcharge.
It’s a huge overcharge. You’ve mentioned this in your talks a lot.
Yeah. A bunch of people are basically looking at this and thinking, “If I think prices are going to go up, I might be willing to pay more than I previously thought in the future. If I can buy some now, but can’t buy any more, then maybe I’ll be willing to pay even more in the future just to lock it in, if it’s going to come online later.” So we see the other side as well. It kind of depends on who you are.
12. GPU generations, specifications and SKUs
Cool. I really want to make this super clear. Do you care about the different generations of chips, or do you just care about overall FLOPs and spread them among whatever chips you want?
We care about different generations of chips because—
You care? I guess, do your customers care?
Yes.
Right. When you say different generations of chips, do you mean, like—
Like B300 generation?
Yeah, exactly. Or even AMD or whatever, right?
Yes, I think people totally care about the generation of chips. Part of this is because of things like VRAM, and part of it is because of larger NVLink domains and things like that.
So there’s no way to abstract all this into some generalized unit of compute that you can mush around however you like?
I think, like you were saying with oil, even oil isn’t just oil. It’s a bunch of different types of oil.
Yeah, like jet fuel or car fuel or whatever.
I think it’s the same thing. You’re going to get a bunch of different types of—
It is, but you have oil refineries that you can effectively use as black boxes: you put oil in and get whatever you need out. If compute is so heterogeneous, I guess the word is, then you have to expose the underlying complexity to your end customer, which is bad for you as a market maker.
Maybe.
You want to identify the one choke point—or I don’t know what the better word for choke point is—a Schelling point, where everyone agrees, “Okay, this is kind of the standard,” and we all agree to it. It’ll be somewhat inefficient.
Yes. We’ve been able to reduce the problems to a certain level.
Yeah, your job is dimensional reduction and all that.
We reduce a lot of the dimensions, but there are still dimensions. They’re still going to exist.
Do you know how many SKUs you have?
So—
Like 5 times 100 times 5,000? That equals, I don’t know. What’s the number?
We actually made SKUs a concept, and the goal is to make it so that somebody else can make their own SKU. If I want to buy Blackwells, I can say I’m willing to buy Blackwells at $6 an hour or less. You can match me against GB300s or B300s, and I don’t care which Blackwell, as long as it’s less than $6 an hour.
But you just said you care about memory and other stuff.
Totally, but not every customer does. Some customers can work on Blackwells, whatever they are. Some customers are super specific and need this exact thing, and they’re going to pay more for that because the market’s going to charge them more. That’s what we do.
So people use your APIs to programmatically define what they want, and then you match that with whatever composition of sellers might be able to fulfill it?
Yes, with the caveat that having a named query is a new thing we’re doing. But the query system as a concept is how the order book works right now.
13. Physical versus cash settlement
Okay, interesting. I think that’s a lot of the nitty-gritty. Is there anything you want to educate people on, whether financial people or engineers, about the state of the compute trading market as you’ve developed it? You guys are effectively defining it.
Mainly, just the requirement for physical settlement. We spent 3 years learning why you need physical settlement. Our entire conversation today is about how these things aren’t as fungible as you think. We’re going to try and make them fungible; that’s the industry’s job. But without that, you run a lot of risks. If you try to build too much financialization before the industry has had time to build out physical settlement, then you have a market that’s very easy to manipulate and has a lot of basis risk. The price you may be hedging against may have nothing to do with the price—
Yeah.
—you can actually buy the capacity for.
Yeah. For listeners, we talked about this briefly last time, but people are often shocked to find out that the market for credit default swaps is 12 times the size of the bonds they insure, something like that. The markets for mortgage reinsurance and things like that are larger than the underlying physical settlement. That’s why physical versus cash-settled matters.
Yeah. Cash settlements are exactly the kind of side bet you see in The Big Short.
Mm-hmm.
They may or may not have anything to do with the underlying thing itself, and they may be used as a proxy for other things overall. You don’t want that to go out of control.
Yes. To be clear, we think cash settlement as a future concept is a great idea. We just think it should be derived from physical settlement. As you said, it’s a side bet. It would be kind of like if you made a cash-settled future before you made physical settlement.
Which I think people are kind of doing already.
Not yet, totally. People are trying to make cash settlement before they make physical settlement.
You do that for all sorts of things, including startup stock.
Yes. You can buy and sell a made-up share on OpenAI, but it has nothing to do with the actual share.
Yes.
Lots of those go very poorly. It would be kind of bad if you started selling the derivative on a stock that doesn’t exist yet.
Yeah.
Well, one thing we do—I used to be a currency trader, and interest rates as well—is you have a board that gets together and says, “This is the London Interbank Offered Rate, LIBOR.” We all agree this is the monthly fixing, and we all trade to that fixing.
Yes.
Is that even meaningful at some point 3 years from now?
So, critically, LIBOR was manipulated.
Yes.
Our concern at the moment is—
12 banks. Yeah.
Yes. Well, our concern is that there’s a very similar problem where you can manipulate an index. Our impression is that you can manipulate that index because a lot of the indexes are based on a very small amount of data that doesn’t have a lot of correlation with the actual price of buying a new reserve contract.
14. The emerging compute market ecosystem
Okay. Interesting. All right, I think that’s the rough state of things. I think the other part of today is that you talked with Oliver, you’re doing an event with Jump, and you’re friends with Silicon Data. What are you learning from all these other players? Who are the major players you’d call out as doing interesting things?
We’re big fans of Silicon Data and Compute Exchange. I think they were also very early to this, and they’re doing a reasonable job on the path that we think we’ll likely end up partnering with.
Can you distinguish your different paths?
The general gist is that they’re doing cash settlement, and we’re doing physical settlement. We think both can exist and should exist, and that they’re sort of mutually beneficial to each other. I think they’re one of the earliest in the space, and they very clearly took the bet well before anybody else did. I think that’s very honorable to do. You mentioned Jump.
SemiAnalysis, Jump. Are these other players emerging? You were my first introduction to this, and then I see all these other guys coming in, and I’m like, “Okay, something’s happening here.” I don’t know the landscape as clearly as you do.
The categories of folks we’ve seen include people trying to set better standards. SemiAnalysis, I think, is working to create better standards. There are folks trying to do residual value insurance, and so on. There are a number of different companies in the space. I don’t know any of them super well, but I know this is a concept, and I think that’s important.
There are folks trying to do deposit lending, where they’ll front your credit risk for the cloud. There’s an ecosystem of companies forming around the question of how we resolve the financial bottlenecks of compute, and it’s cool to see.
I think effectively what you want in this ecosystem is buyers and sellers, an analyst, and, we said, insurance—or what’s the—
Residual value insurance.
Residual, and you want this healthy ecosystem out there.
Mm-hmm.
But obviously, you guys are owning the problem of physical settlement, which is really interesting. We can go into the hard-tech side, because you’re also hiring—
Yeah.
15. What it takes to deliver a GPU cluster
—for extremely cracked engineers. Describe your stack.
We do all the things the cloud does without being the cloud. We don’t own the hardware. The reason we don’t own the hardware is that, if we did, we’d be able to manipulate the forwards market. So we have to be Switzerland; we have to be independent, which creates a lot of strange things about the company.
In order to do physical settlement, you have to do site visits. You have to visit the data center. You have to design the bill of materials. You have to do supply-chain consulting, where you have relationships with different OEMs that ensure things get delivered on time.
Then you have to do cluster architecture, so you’re designing the cluster itself. Then you’re rolling out racks, or working with partners to roll them out, and supervising that. Then you’re doing backplane management, taking over the PDUs, switches, DPUs, UFM, and BMC, and operating your harness around the nodes you’re settling.
You, as an SF Compute employee?
Yeah.
Wow.
All right. When you buy a node from a cloud, there’s the bare metal, but there’s all this around the bare metal. In order to make that bare-metal node work, you have to operate everything else. Then we—
Because traditionally, the job of the data center is to extract some of this away from you.
Yeah. We’re the ones that do that.
Yeah, yeah. Okay, all right. But you’re not in the data center business. You’re not the Equinixes of the world.
We’re not in the data center business, sort of. What we do work with is—so, we offer supercomputing as a service. What this means is folks come to us and say, “I’ve got X,” where X is, “I’ve got X megawatts here.” Sometimes they mean, “I’ve got a data center that I’d like to turn into a GPU cluster. My options are I could sell the GPU cluster, I could sell the data center to somebody, or I could become the cloud.”
Then we go to them and say, “Great, we’ll turn you into the cloud. We’re the supercomputing part; you’re the data center cloud. Let’s partner together, and then we’ll sell and all make money.” Sometimes it’s people who have something that could turn into a data center. In that case, we’re going to project-manage that up to the layer, and then work with those folks to turn their powered land or whatever they have into a data center by working with partners. Then we take over at the supercomputing part.
This requires us to have some data center expertise in-house, but we don’t self-perform builds.
Yeah, so it’s somewhat modular because you have different scenarios you need to deploy under.
Yeah. Someone once described us as Palantir for compute, because each time it’s this new rodeo. But we deeply partner with folks on both sides.
And primarily the U.S.?
Primarily the U.S., but we operate around the world.
At some point, does SF Compute become a misnomer?
We’re deploying clusters all around the world, and when we do that, we’re working with people who are themselves trying to be clouds all around the world. They look like sovereign compute players. Our pitch to sovereign compute is, “Amazing. Let’s help you set up a cloud in your own country. We’ll bring the San Francisco offtake. We’ll bring the San Francisco expertise. We’ll set up the clusters.” That’s our whole business.
Ah.
We’re going to make it financeable.
Who are your biggest announced partners that you can name?
Nobody I can name, basically. We’ll have some announcements in a week or two that we expect to come out.
Sure, sure. Good luck. I think this is really interesting for every country investing in sovereign AI to figure out. It’s interesting that SF is still that brand name, and that it’s very marketable.
I think one of the interesting things is just thinking through this market.
Mm-hmm.
Nvidia is obviously the 800-pound gorilla in the room. Why didn’t they make SF Compute? Again, I know the answer’s going to be market manipulation, but let’s really get into the incentives of the multiple players.
Our impression is that Nvidia can’t make it but benefits tremendously from it existing. The reason they can’t make it is that they wouldn’t be independent enough for people to trade on it. If you don’t trade on it, then you don’t have enough liquidity for lenders to finance against the forwards market.
This is also why Crusoe, CoreWeave, and Lambda can’t make it; why the hyperscalers can’t make it; and why Baseten, Fireworks, and Modal can’t make it, because they have too much demand on that side. This benefits the entire industry because it makes the financing costs go down. But nobody can make it unless they give up the ability to do all the big inference money and all the big cloud deals. You have to kill your business in order to make this.
That’s why nobody does it. Structurally, they’re not set up to do that. You’re basically reducing friction, and that’s all your revenue. You’ve announced some benchmarks, some milestones. Actually, I forgot to say that up front. Can you repeat what they are and take a victory lap a little bit?
16. A programmable data center
We closed several hundred million dollars of volume. We have a significant number of new clusters on the way. The things I’m most excited about are the new features we’ve shipped that make SF Compute very viable for platform-as-a-service companies.
For example, we allow customers to subdivide their InfiniBand fabric without having to ask us. Many clouds, when you go to buy from them, will give you one InfiniBand fabric. If you want to redraw the lines of that fabric, you need to contact them, and they’ll do something for you. We can now do that on the fly for customers, which I think is really great because it interfaces with the market really well. You want to be able to redraw things in order to scale your clusters up and down.
Would you describe this as virtualizing the data center and making it programmable?
Yeah, we’re virtualizing the data center and making it programmable. All the stuff that you would have gotten if you had owned the hardware, we’re trying to virtualize that, which I think is exciting to us.
It’s funny, the last time this happened was with the cloud.
Yeah.
Which is really cool. And then finally, what is your call to action for people? Is it engineers? Is it, like, “Give me your supply”? What is it?
If you would like to own a GPU cluster and you want the residual value from it—you want to have it in your portfolio—we would love to work with you to build that GPU cluster. We’ll take over and do the supercomputing portions of it and ensure things go right. We own the SLA. If it turns out that was a bad idea and you need to get out of that cluster, we can immediately sell it to someone else so it doesn’t hit your balance sheet.
Very simple pitch. Well, thank you so much for your time. You’re doing a great service to the industry, and I’m just happy to be your friend.
I appreciate it.
All right. That’s it.