Financing costs are super expensive. It's what's keeping a lot of the top neoclouds — their profit margins are very small, and it's because they're paying these huge, huge costs per chip every year. So by saying that there's a person willing to buy this hardware in four years, it's certainty, and with certainty there's less risk, and with less risk you get charged less.
Welcome to another Semi Doped podcast. I'm Vikram from Vik's Newsletter on Substack. With me is Wayne Nelms, CTO of OR. OR is building a financial exchange for trading GPU compute as a standardized commodity. Think of it like a futures market for compute power, and more recently memory too. So we're going to talk about it in this episode. So Wayne, how are you doing?
I'm doing well. How are you?
Pretty good, pretty good. So I heard you got back from PTC in Hawaii. How was that?
It was amazing. Just got back late last week. Unfortunately, the flight back was a little rough. There were some stops for weather, and I think the brakes on the plane weren't working, but got back safe and sound to cold New York. So yeah.
Yeah. Was the weather warmer in Hawaii?
Oh, way, way warmer. It was t-shirts and actually Hawaiian shirts and khakis every day. So...
Yeah, now you're back to your heavy jackets, I suppose.
Awesome. All right, so how did you get into the idea of starting a financial exchange for GPU compute? This sounds like a great idea. But how did you even come across this thing? Give us some background on how you got into this.
Yeah, sure. So I can give you the full rundown. So I graduated not only just last year — I graduated in December of 2024, a semester early, from MIT. And there I was studying math and computer science at school. I was on the tennis team; I was a big tennis player in high school as well. And then I got into poker in college. And so pretty much after I graduated, I was trying to figure out what to do. I started working in trading. So I was a quant trader doing equity options at Susquehanna, or SIG, in New York. And I loved the job, but at the time I saw all these companies, all these Bitcoin miners that were evolving their business models into GPU as a service. So, you know, Iris and Terawulf, you know, then CoreWeave, most famously, I guess. And I kind of saw their stocks take off, and I was wondering what's going on here? I have to figure out what is a semiconductor, what is going on. So I took some time, did a bit of research, and went to one of my closest friends at MIT from school. We had known each other for four years. His name is Kush Bubaria. He's the current CEO of OR. But we were just trying to figure out what to do. And at the time, I was doing a little bit of stuff on the weekends with a friend in private equity, kind of talked to him about what they were doing every day. And he told us, both Kush and I, that they were looking into data center deals, but they didn't really know what was going on. They didn't know how to measure risk. And this theme kind of was the genesis of what OR is today. It's how do you categorize risk in the data center space, in the neocloud industry? You know, what is the value of compute? And then we kind of took that idea and ran with it, using my financial background, and Kush has a background in venture and startups. So taking our unique set of skills and then coming together to try to solve this issue.
That's awesome. Actually, that's an important question — like, what is the value of compute? Because now you have this era of agentic AI, and you have cloud bots that are like taking over all of people's computers, I suppose. Now people are trying to buy Mac minis. So compute has somehow landed up at the edge all of a sudden. So the question is always what is it on the cloud, and what's the value on the cloud, and things like that. So yeah, this is a very important question. So basically, what financial infrastructure exists for AI compute today, or is there no such thing? What's out there?
Yeah. So we're still in the very early stages of all this stuff, of the financialization or commodification of compute. I would say there's a few startups that are focused on trying to standardize it. So you have people that are actually focusing on the offtake, like SF Compute, those sorts of models, and they have standardized compute in a sense, right? They have minimum benchmarks for your compute that you can resell on their platform. And then there's also a few businesses like ours that are tracking prices of compute across the country, based on provider, based on chip type, networking, etc. But in terms of actually being able to trade it, which is the next step, right? We can track it, build indices, and look at how numbers move. But if you can't trade it, if you can't hedge your actual operating expenses with this number, I would say the value is minimal. So what we do at OR is we actually allow you, as either a data center, colo operator, neocloud, or on the flip side, right, if you're buying a lot of compute, if you're, you know, a lab, university, hospital, we allow you to actually take positions against this number or on this number, and that can help you materially hedge some of your operating expenses. So I can give you a brief rundown, a brief example. So...
Let's say you're a lab, and let's say you know you have a long-term contract, of course, you know, three to five years, let's say. But let's say you have some inference demand, and that demand is very spiky. You can't really predict the inference demand. So you need to buy some compute on the spot market, let's say, and the spot market meaning right now, I need it today. A lot of the time, if you're buying a lot of compute at once, it can drive the price up. That's just how markets work, right? Supply and demand. What people need right now is a way to hedge the price. So if I'm buying a lot of compute at once, I want to be able to protect or put a ceiling or a cap on how much I'm paying. So what we do is we allow you to buy a future on our platform. It's fully cash settled. Meaning when you buy a future on OR, you don't actually get the GPU hour, but what you do get is a way to hedge your cash risk. So if the price of the index moves up because you're buying a lot more compute, you actually make money on the hedge. So in effect, you're capping your upside and your payment. And on the flip side, you know, if you're a data center and you're selling a lot of compute, you can sell a future on our platform and put a floor to the price that you're selling all of your compute for. So it works in both ways, and we've seen a lot of people being interested in this product because in every single commodity space, if it's oil and gas, you know, electricity, this sort of product exists because there's a lot of people buying today or tomorrow, and they care about, you know, putting a cap or kind of a collar, we call it in finance, on their prices, their costs, and their revenue.
Awesome. All right. So for like basically GPU pricing, let's take an example. So I know it varies depending on what kind of surrounding resources you have with the GPU compute that you're going to rent out, for example, on the spot market today. Like I want to run a little, you know, fine-tuning run on a model, and I need some compute to do that. So let's say we pick, I don't know, an H100 or a B200 or something, you know, take your pick. What is the pricing example like? Is it $100 an hour? Is it $10 an hour? How does that spot market number look like between a few providers? If you have any numbers to put out there, I think it helps understand, even for me, where this is coming from.
Yeah, sure. So we track a variety of providers. So let's just say H100, SXM — that's one of the models that we think about the most as a benchmark. So, you know, on marketplaces, a lot of the marketplace pricing is a bit lower on the lower end. So we have roughly around $1.70 to $1.80 today for an hour on an H100 GPU.
But as you move up the stack, so I'm saying, you know, neocloud-level pricing, so call it the CoreWeave, Lambda, Crusoe, Nebius, you're getting around, you know, three to four dollar range. And then if you go up to, you know, the AWS, GCP, you're looking, you know, five to six to seven dollar range. And so, you know, part of the problem when you think about standardizing compute is how do you standardize across all these providers, because each of them is kind of providing a different level of service potentially, a different level of quality, uptime, you know, sort of these benchmarks and metrics. So what we do is we exclusively think about the marketplace. So closer to the, you know, the quote-unquote lower end of the spectrum, at least in terms of cost, because we think that these marketplaces are a more realistic example of how a spot market should work, right? There's a lot of people that list compute, there's a lot of people that buy compute. The market is very dynamic. When you start looking at these hyperscalers, the truth is the prices don't change very much, right? They might decrease their cost or increase their cost every quarter, once a quarter, twice a quarter potentially. And we don't think there's a true reflection of where daily markets are moving. You know, as a product that tracks GPU hours, we want to have our index be as dynamic as possible and, you know, allow you to have real exposure to the truest form of on-demand compute. So we track marketplace compute.
Gotcha. So for the same H100, depending on where you're running the actual hardware, the price actually varies, right? Is there a reason for that? That's my first question. And the second question I have is, how do you standardize a fixed price if everybody is charging different rates for the same hardware?
Yeah. So a lot of the differences come down to the hardware itself as well as the regionality, or where is the compute or the GPU actually located. But these are problems that have been solved in the electricity markets as well, right? So in electricity markets, the way it works is that you have hubs, and each hub is a location, you know, either in the US or across the world, which has a different price for the same electricity. And for compute, for powering a GPU, you know, electricity is a huge input cost. So you can only expect that this regionality bias in electricity should also be analogous to the regionality bias in compute. So what we do is we consider only the compute in a certain region of the United States, and instead of kind of trying to standardize across the entire country or the entire world, we say, look, we only will track a specific type of compute in a specific region. And that allows us to maintain kind of uniformity across our product. It allows your hedge to be very specific, and as a result, hopefully makes your hedge a way better product for you and your business.
Gotcha. So going back to the example of where I was going to rent an H100. So now through OR, let's say I can purchase a certain amount of compute at a certain price that is determined by the OR index that you determine through the spot price market. Is that right?
Okay. And so basically, now that I have this bunch of compute, now let's say the next quarter rolls around and the compute prices go up, and now for the same — because I kind of bought compute at a certain lower price — I'm doing better off now because I was able to get in early and get my fixed rate. Is that how it works, right?
That's roughly correct. So yeah, how it pretty much works is that the futures — the way it works is you're buying it today for a time in the future. So what you can say is, look, I will be buying a lot of compute in, let's say, February of 2026. So what you can do is you can go on our platform, there's a February 2026 future, and it has a price, let's say $2. You can buy that future for $2 right now. And what it does is it exposes you to the spot prices in or across February. So if we get to February 1st and the price of our index is $21, and then on the 2nd of February it's $22, and then let's say it goes up and down around $25 across the entire month. You're buying compute across February because that was when you needed compute. However, because you bought our future for $2, you effectively have locked in your cost. So even if compute prices went to $10 or $20 — you know, it would ideally never happen — but if it went up significantly, it wouldn't matter to you because you already bought the OR future. So what we're doing is effectively putting a price cap on any compute that you would be buying in the future.
Gotcha, gotcha. And so what determines the spot pricing, let's say of $2 in the beginning of February? What are the factors that kind of go into this pricing to begin with in the markets? The market set this value, right? Somehow.
Yeah. I mean, it's purely supply and demand, right? And it all boils down to that. So just how many people are offering H100 hours in February, and then what is the demand from both, you know, the research, the training, and the inference needs of the community of AI users, right? And that's kind of how every single market works. And that's why we love and we've tried to bring all those financial concepts to compute, because at the end of the day, there's huge supply constraints in this market and also huge demand. And in any sort of resource of human history in which this sort of dynamic existed, financial markets have closely followed. So we just believe that, you know, we're at a stage in which there needs to be a financialization of compute, and we're trying to build the first kind of, take the first steps to make that possible.
