FOUNDER FOCUSIn this edition, Tech in Asia speaks with Evan Conrad, co-founder and CEO of San Francisco Compute, an AI computing marketplace that allows customers to buy, sell, and sublease computing capacity or procure it directly from the company. Conrad believes his startup can help companies cash in on excess computing capacity while reducing their exposure to a potential AI bubble. This interview has been edited for brevity and clarity.  Some people describe your company as the Airbnb of computing power. Is that a fair and accurate description? We don't describe ourselves that way. The difference is that Airbnb doesn't run the hotel. We run the hotel. We operate and run the clusters. A customer can come to us and buy a long-term contract. Then they can sublease that contract. We take over other people's clusters that they have bought or assembled, or maybe they just want to own a cluster because they want to make a financial return on it. It's not locked in. We don't want to lock you in. For most companies that are building some sort of AI lab, they go out and raise, say, US$50 million, and then spend US$49 million on compute. If they don't bring in the revenue, that's it. Everything is over. The money is lit on fire. There's no recourse. We built a company that allows you to get out of that. We are also building a data center right now. It will be ready in 2027. How can a startup at your scale build a data center? Where does the money come from? We've raised US$60 million at the topco level. But for large-scale, capex-heavy assets, we raise other types of capital at the project level. It's the same way a real estate developer might raise some initial money for the company, but then raise separate financing for each project. At the moment, we're developing a multi-gigawatt-scale data center campus in the US. Each building - just the powered shell, meaning the power plant and the building itself - costs roughly US$3.5 billion. If you add chips, it costs a lot more. But we're not raising that on our balance sheet. We're not raising it from venture capitalists. People think we're a lot smaller than we are. We're just very low-key. We operate at quite a large scale. One of the advantages of being in San Francisco is that we know everybody personally. Our CTO is the co-founder of Voltage Park, a multibillion-dollar GPU cloud. We're made up of people from the original Lambda Labs team, along with folks from Zoox, Hut 8, and AWS, covering different layers of the stack. When we do the asset-heavy part of the business, we work with other entities that take on that portion. Sometimes that means forming an SPV [special purpose vehicle] for a specific project. The majority of the equity investors in those projects aren't us. The downside is that we don't make as much money. How can this model survive if computing power is getting less expensive? Computing itself - though this may be a lagging indicator - could actually become more expensive over time. The rationale is that it takes a long time to deploy clusters, in part because it takes a long time to build data centers. That's one reason we started building a data center ourselves. As we expanded our clusters, we ran out of data center space. We can operate on top of what's called a colo facility, or a rented data center, but capacity is limited. We hired a data center development team, and right now, the entire industry is building a wave of new projects. The problem is that it will take years for those projects to come online. Meanwhile, demand is growing like crazy. I think, in the near term, you're going to see a pretty significant crunch. What are you looking for in areas like Southeast Asia? We're looking for colo facilities, so we can be your customer if you're building a data center. We'll put clusters in it while you operate the data center. We want to run the cluster inside the facility. The second thing is that we're aware there are lots of people around the world who want to offer GPU as a service. We partner with those operators. We'll take a share of the revenue, but we'll build the cluster for you. We'll bring Silicon Valley offtake to you. We work with major AI labs to deploy clusters in different regions. We're the people who turn you into a cloud provider. We're like GPU consultants. What many operators probably don't have is the expertise to build large-scale supercomputers. They probably also don't have Silicon Valley offtakers, meaning customers that can make a project financially viable at scale. That's because the largest customers right now are in San Francisco. What's your take on the potential AI bubble and where the industry is headed? To some extent, SF Compute was set up to help people avoid a bubble. If you think about Stability AI when it ran into trouble, it had overbought capacity and really needed to sell it. xAI is a great example as well. The only reason it didn't blow up was because it was able to sublease its clusters. Our pitch is simple: If you buy a three-year contract from us, we'll let you sublease it through our marketplace. We'll help you get out of that contract if you need to. That's where we think the bubble is. If you imagine what an AI bubble might look like, it's people buying GPUs that they ultimately don't need. You could also imagine GPU cloud providers taking on a lot of risk by speculatively building clusters. In reality, though, almost no one is building large-scale clusters on a purely speculative basis these days. The moment you sign, that's when they place the purchase order. That means the party actually taking on the risk is the customer. We're trying to prevent or reduce that risk, or at least prevent individual companies from being overly exposed to a potential bubble. |
