Long Strange Trip: CEO to CEO with Brian Halligan
00:20:48 · Chase Lochmiller
[00:20:48] S0: At this point, you know, Crusoe is very focused on activating megawatts for intelligence, and that requires all sorts of energy resources and all sorts of different use cases. At times, it's natural gas. You know, we're we're focused on trying to find and and and catalyze innovative new energy solutions that can be things like Blue Energy. They're trying out new nuclear, you know, energy production technologies. We have another partnership with Oklo, that we announced recently, where we're trying to bring to life the first, SMR powered AI factory at the Idaho National Lab. We've done stuff with behind the meter wind, with behind the meter gas, with behind the meter solar, with batteries. We've many different things. We've gone to Iceland, right, where Iceland's this sort of like geological phenomenon that it exists, but it's this land of fire and ice where you have very abundant, low cost geothermal energy, very low cost hydropower, and not a lot of people and and very, you know, cheap, efficient cooling. And and bringing the demand for power there via compute has been our philosophy of, and and that's really what we mean when we say take an energy first approach to building compute infrastructure. Norway is another great market for us, where they have a lot of low cost hydropower that's being underutilized and can be built bigger than it is today. And so we don't think it's a silver bullet. We think compute infrastructure will be ubiquitous to serve the infrastructure layer of intelligence that's ultimately gonna power, global economic development. And so you need to use different energy solutions in different cases. And and so we're really excited about a lot of the innovation taking place in the in in the industry. And, you know, Blue Energy is another great example of, like, you know, can we leverage the same steam turbine technology that, you know, is is needed for gas generation to basically... That... That's needed for both gas generation and nuclear energy to be a sort of bridge between gas and and and baseload firm nuclear power. All of these things we think are great. We think a lot of the development in new battery technology is really exciting. Everything from what's happening with companies like Redwood Materials that are repurposing used electric vehicle batter...
00:23:16 · Chase Lochmiller, Brian Halligan
Batteries and used lithium ion batteries to long duration storage solutions like Form Energy and Antora, that we think will be cost effective hundred hour batteries. And so there's like a spectrum of whole different things, and we kind of view it as like we sort of need them all. The demand from intelligence for energy is voracious. And it's really going to accelerate the human progress as it has more access to energy. [00:23:43] S1: One of the things I find fascinating about you and your company is you started essentially in the Bitcoin mining industry with stranded energy, putting that to work. And it was a really good business. You were cranking with it. And it sounds like from the beginning, you sort of foresaw there's going to be a need for AI infrastructure, and you started building it out. And it rhymed. At some point, you decided we're going to sell off this. Most of our business is this Bitcoin mining business. When did you make that decision? How did you tell your board? Take us behind the scenes on when that became a good idea in your head. That that was a bold move.
00:24:27 · Chase Lochmiller
[00:24:27] S0: Well, so the thing was is that I think externally, a lot of people thought this was a pivot for us. But internally, and if you talk to any of our investors or any of our employees that have been here for a very long time, they'll tell you that, like, this was sort of the plan all along, was this energy first approach to computing. You know, I had been doing research in AI as a quant portfolio manager. And then I was at Stanford for grad school and working in AI and doing a lot of work in computer vision and natural language processing. And I was pretty AI pilled early days. Like, when you know, I was reading things... Like like, one of the more influential things on me was actually Tim Urban had this blog called Wait But Why. And he wrote this blog. Yeah. It's great. It's fantastic. Just like everybody should go out and read that. But there were a couple pieces that were particularly influential on me. There was one that was about the path to superintelligence. And in it, he espoused this thesis that these multilayer neural networks were essentially really good nonlinear modeling mechanisms that, as they had more data and more compute, they just got better and better at modeling whatever thing that you're trying to model, whether it's natural language or computer vision or whatever it was. And so I was sort of playing that out and felt like AI was this meta science that wasn't going to improve one thing. It was going to improve everything. Right? So I had this professor at Stanford named Andrew Ng, and he has this famous saying where he says, AI is like electricity. Right? Electricity is not good for one thing. Right? If you ask someone, what what is electricity good for? It's it's not like, oh, electricity is good to, like, turn on the lights. It's like, yes, it's good at turning on the lights, and it's also good at powering my computer and my car and heating my stove and, you know, all these different things that we really take for granted with electricity. And so it's not just like one thing that, you know, that AI is good for. It's really good for everything. And and, you know, I was... I had, like, high conviction of this fairly early days. And, you know, I felt like if data and compute were the bottlenecks, ultimately compute becomes an energy problem. And when you you sort of play that out and think about that at scale...
00:26:51 · Chase Lochmiller
And and and we'd sort of seen this unfolding actually with Bitcoin mining. And so the original thesis for the business was like, find stranded energy that is otherwise being underutilized and bring a market to that stranded energy with compute. Instead of trying to bring that stranded energy to a market where it can be bought and sold, bring the market to the energy. And that's what led us to our initial stranded energy resource was natural gas that was being flared in the oil field, where oil companies were drilling for oil. They produced natural gas as a byproduct. And then absent having a pipeline, they just light it on fire. And it was this huge environmental liability. It was a waste product. They were paying money to dispose of it. It's a disaster. And it's still happening at scale today. But, you know, our solution to that was we would bring these small modular generators and these small modular data centers, and we would we would basically utilize that gas on-site and use it to power compute infrastructure, which, you know, our first monetization engine of that was Bitcoin mining, right? It was the easiest way to take energy and turn it into money. And we sort of created this great flywheel effect with the Bitcoin business. And there were definitely criticisms early days when I was like, Hey, I'm trying to do this AI cloud platform. And people were like, What are you talking about? Why are you spending any resources on that? You have this great thing. You're making all this money in Bitcoin. What are you talking about? Why would you focus on anything else? And we tried a ton of stuff, and we failed, and we screwed a bunch of stuff up in terms of building out our first AI clusters. And we're trying to figure out what the product looked like and who we were going to serve. And I called a bunch of my friends that were researchers at universities and saying, hey, do you have a use for these GPUs that I'm powering from this gas in the middle of the Bakken oil field in North Dakota? And there were just many different iterations of trying to get that business to life. But the magic moment was really the launch of ChatGPT.
00:29:02 · Chase Lochmiller
So 11/30/2022, think that was just like the moment that changed everything and just it led to all these new startups that were demanding the exact type of compute infrastructure that we had been building and the exact type of solutions that we had been building from a storage, compute, networking, etcetera. Now, I had always felt that Bitcoin and AI... And I still feel this way, that they actually go together very well because AI data centers do not consume energy in the same way that Bitcoin mining does. Bitcoin mining is very like... You can almost give it a fixed draw amount of power and say, Stay at that power level and just consume that. And it will be very flat line. And then you can move that up and down depending on demand response and what available energy is. AI has a bit more variance in terms of its overall power utilization. And you see this in certain when you're training big workloads. For inference and training, but you also have seasonality with cooling. You tend to have higher PUE during the day when you're having to run your air conditioning units and your chillers more aggressively because it's hotter than at night when you have more natural cooling. Depending on workloads that are being run on the clusters, things can ebb and flow a bit.
00:30:19 · Brian Halligan
[00:30:19] S1: Okay, but I'm going to come back to You did... So, it was it was on... You were thinking of it was an AI company all along. You made a decision to sell off what was a good business. Why didn't you just keep it? When did you broach that with the team? When did that become a good idea in your head? Did you get pushed back on it?