INTERVIEW

AI’s Biggest Cost Lacks Standard Pricing 

Five Minutes With Simeon Bochev, CEO of Compute Exchange

By now, you’ve heard how expensive AI is to train and run. McKinsey estimates the boom could require nearly $7 trillion in data center investment by 2030.

Numbers that big are hard to wrap your head around.

Even harder to grasp: buyers of computing power have almost nothing to benchmark against. There’s no Expedia for compute the way there is for flights. It’s hard to know if what you’re paying is “fair” or “market rate.” It’s similar to calling around plumbers for quotes to fix a busted pipe.

And how do you keep track of how much compute you’re using? Does that match your invoice?

Today, there’s very little standardization.

That gap is what Simeon Bochev is betting on. His startup, Compute Exchange, is building a marketplace to treat compute more like a commodity. I spoke with him about how this market could evolve, why he teamed up with DRW’s Don Wilson, and where he thinks compute belongs on the spectrum between oil, electricity, and currency.

As AI costs balloon and more companies scramble for scarce GPU resources, opaque pricing lets bad actors exploit uninformed buyers. The market is ripe for the kind of transparency common in today’s commodity exchanges.

This interview has been edited for clarity and length.

How did you connect with Don Wilson?

I was invited to a dinner hosted by Don to talk about the commoditization of compute. The CEOs of most of the neo clouds were there.

The reason my CEO wasn't there is because he's in Vermont and I was his leadership person in the office, in the headquarters in San Jose. And so I went and Don and I really hit it off. And we really talked about how there are these structural issues in the market.

I described one of them, which is pricing transparency, right between Amazon and Apple. But the issues go well beyond that, and I don't want to go into a monologue, so I'll skip the issues for now. But we identified a lot of structural issues on both the supply and the demand side.

And so we said maybe a market can solve that. And so I ended up being Don's point of contact within Lambda about building this market. And as we went down that path more and more, I'm like, actually, why don't we do this together? And so, with the blessing of Stephen (Balaban, CEO) at Lambda, we started Compute Exchange in April of 2024.

And now, more than a year and a half later, we launched the market in February, we've had over 80,000 GPUs in supply in about five months. To go from zero to 80,000 GPUs, I challenge anyone else to find someone that's grown that fast. And it's all in GPUs that we don't own. We don't take balance sheet risk on. We open them up to the market.

I see. So, you're the auction? You are the place to source GPUs.

We are the market. Our goal is to be the place where all compute in the world transacts.

So, think about how oil transacts largely in one place, whether you're buying a barrel to speculate as a market maker or you're a country buying a billion for, or maybe not a billion, but a million for strategic oil reserve. They all transact in an exchange. That's what we aim to be.

Don Wilson thinks that compute demand will outstrip oil in 10 years. Of course, he has vest interested with stakes in Compute Exchange, but that’s a bold statement. Do you see compute as oil? I tend to think of it as freight.

Sam Altman calls it the currency of the future. I talk about oil. Some people talk about electricity. I don't think there's an easy, this is the exact parallel, which is why part of the onus is on us to educate.

Here's how I look at it: You've got physical hardware that goes into data centers. That hardware, an H100, has a serial number. It can be the number 1 or a million and excluding any defective units, they’re the same thing. You put that hardware inside of a much larger, complex system of hardware. A server or a rack, networking, storage, liquid cooling, all this stuff coming together. That thing then goes inside of a data center where you pump in a lot of power.

Then, there are differences in the data centers, the power cost, where it's coming from, is it carbon friendly or not, et cetera. And then there's the software layer on top of that. That whole thing we call the AI stack. That thing is what we are trying to build a market out of.

It's not the GPU, right? It's not ‘Go resell an H100 alone,’ because that doesn't help anyone, outside of maybe a few countries that don't have access to that compute. And so the analogy of this thing doesn't really exist anywhere else in the world. One factor of it is power, right? And yeah, you can model power grid demand—and as GPUs consume more power, great. You can use that as a proxy for some energy demand. But that’s not the whole story. Just like oil isn’t the whole story. With oil, you have a limited range of octane grades. Most people don’t care whether it comes from Saudi Arabia or West Texas. Candidly, when they fill up their car, they care whether it’s jet fuel or unleaded.

We ultimately hope to get compute to that level of simplicity, but it's not there yet. So, I would say there's not a simple analogy, but there are elements of these other commodities markets that work.

I used to cover white-collar crime—SEC enforcement, financial shenanigans, all that. So, I’ve seen firsthand how a lack of transparency can play out, whether it’s with level three assets or elsewhere. You’re standardizing this market. And I have a feeling that at some point, there are going to be some stories uncovering shady behavior, where someone’s selling compute in a way that’s not exactly transparent.

Without using names, I’ll give you an example. We know some very smart young people who started a company and were being advised by someone they trust. That person is the co-founder of a company that sells GPUs.

Obviously, that person holds a position of power over them as a trusted advisor. And the price they secured through this person’s company was, at the time, about 30 to 40% above market.

To me, that’s icky. Imagine if I’m your parent, but I overcharge you by 40% for something.

Maybe I’m trying to teach you a life lesson—but that’s not what’s happening here. I’d love for independent folks to step in and find these things because it would make the market more efficient. My goal isn’t to point a finger and say, you’re a bad actor. They’re doing what they have to do to sell their compute.

At the same time, in this case, it shouldn’t fall on the users to figure out why it’s a bad deal. It should be obvious.

Even if I’m a sophisticated buyer, we say you shouldn’t need a PhD in infrastructure to make sense of this. But right now, you kind of do.

Would you consider naming who is setting the lease price?

We don't want to call out anyone. For example, we don't want to say a particular company closed the deal at $4.26, right?

That’s not our job. And I don’t know of another exchange that calls out the buyer and provider by name. Their goal is to aggregate the data and make it available to the market.

We work with Silicon Data, and I’d say they’re the right ones to go a level deeper for us. Their role in providing pricing information is to help create a more efficient market.

Any final thoughts?

I see a large part of my job—and our job as a company—as education.

We put the pricing intelligence out there; we’re not monetizing it. The only way we make money is through transaction volume. The more people who buy in the market, the more we earn from a fixed transaction fee. If we can help make the market more efficient, that reduces a lot of friction.

I’ll be brief and just highlight two challenges.

On the provider side: they need to sign multi-year contracts so they can raise capital to buy chips and deploy them.

But customers hate long-term contracts. Why would they want to lock themselves into something that’s depreciating over time, especially with a fixed commitment?

And for startups, compute is their biggest cost—even bigger than headcount. If we had a liquid market, we could enable derivatives to lower the cost of capital for providers, without forcing buyers into long-term contracts. That gives buyers more optionality.

Ultimately, that’s what we’re trying to build.