The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
00:15:07 · Chase Lochmiller
[00:15:07] S0: After the ChatGPT moment happened, we had sort of watched this evolution of the data center to support Bitcoin. And if you look at the history of Bitcoin mining, it's it's pretty fascinating. Right? It started with, like, hobbyists on their laptops in this sort of decentralized way, and then people started to use GPUs, and then people started to lease data center capacity from, you know, more traditional colocation facilities that have, like, five nines reliability. People started to develop ASICs. And then at some point, people realized, you know what? We don't need five nines reliability to mine Bitcoin. We can do it with three nines. Oh, wait. No. What about one nine? Like, how much does that reduce the actual data center cost? And they were able to eliminate something like 98% of the total data center cost with like these, like, chicken coop type data centers. You know, very, very low cost, no redundancy, but, you know, ultimately brought down the payback period for these investments in Bitcoin mining. And, you know, we felt always that something similar was gonna play out in AI. And we're long term partners of NVIDIA, we'd been working with NVIDIA. And I was looking with them at their roadmap, you know, many years ago, and I was like, okay. Our current generation of chip is, you know, a 100 and watts or 200 watts. And, you know, the next generation is gonna be like 300 watts. And then maybe, you know, there's gonna be a future generation that's maybe 600 watts per chip. And I was like, okay, well, you know, if... As the power density goes up, it's gonna change what the data center looks like to actually support these things. And we also had a perspective on the compute workload. The workload itself for AI, it didn't need to be in these centralized facilities. I think so long people have been anchored on these, like, centralized hubs for data center computing infrastructure. So Northern Virginia is, the hub that runs most of the Internet. Most web applications run out of Northern Virginia. But it didn't have to be that way with AI because, you know, so much of the time to serve a neural network is actually the compute that's happening in the data center, not the time to get to the data center. And so that, I felt like, opened up new geographies, and we felt like infrastructure to support AI wasn't gonna be centralized.
00:17:15 · Chase Lochmiller, Harry Stebbings
It was gonna be distributed where energy was low cost and abundant. And so when we think about things that we did differently, it was like investing more heavily in how to design and develop the data center to support AI, and investing heavily in how do we find and source energy resources to support the scaling of AI. [00:17:33] S1: I I wanna kind of unpack those a little bit and kind of going back to some form of structure that I I did plan, I promise. That's great. If we start on data centers, we then have Crusoe Cloud, Managed Inference, energy kind of runs throughout it all. I'll kind of pick that apart too. On the data center side, we always hear about the supply constraints today. How truly supply constrained are we?
00:17:54 · Chase Lochmiller
[00:17:54] S0: Where it's actually manifesting is there are not places to plug in GPUs. So that's ultimately the supply constraint. There's just not places where you can plug in GPUs and turn them on and run AI workloads. So how does that actually manifest across the supply chain? You know, the supply chain to support large scale AI data centers is sort of like a game of whack a mole. At different points, different bottlenecks come into play, and I think that's actually one of the critical aspects of being vertically integrate... Integrated is that we're able to navigate a lot of those critical bottlenecks and situations. I'll give you an example. When we set out to build the first two buildings in Abilene, it was a, you know, a little over 200 megawatts of compute capacity. We had sort of committed to doing this in one year, and the next closest bid was two and a half years. There were 34 other data center developers, and you know, but speed was of the essence. So I said, sure, we can do it in year. One of the key bottlenecks was what's called a power distribution center. Where, you know, you receive medium voltage power and then you... So at around 34.5 KV, and then you distribute it to low voltage transformers that ultimately go into the data center and power the racks. When we went out and canvassed the market for vendors that were supplying these medium voltage power distribution centers, the lead time for this one component was a hundred weeks. And I said, well, I don't have a hundred weeks. I committed to doing this in a year. So I went to our internal team, and we had consciously vertically integrated electrical manufacturing. We said, how quickly could we make this ourselves? The team did a bunch of work, sourced components to make these things, and we were able to do it in twenty eight weeks. And so being able to navigate those types of key bottlenecks by actually having the resources internally gives you ultimately a lot more flexibility to unblock these key bottlenecks, number one. And number two, it gives you a perspective on, like, what are... What is the true cost of a lot of these different things that are being built. Elon has this very like interesting framing of this.
00:19:52 · Chase Lochmiller, Harry Stebbings
He calls it the idiot index in terms of, know, what what is the what is the cost of raw materials, and then what is the end product relative to that cost of, you know, just the raw raw materials going in. By being vertically integrated, we have a perspective on, like, the end to end cost of everything going into building and operating an AI factory. [00:20:10] S1: How much extra margin do you juice out by having that internal build process yourself? Like, is [00:20:16] S0: the idiot index chasm? Look. Electrical manufacturing is not like an ultra high margin business. Like, today, it's actually spreads have gone up quite a bit. Margins have gone up quite a bit just because they're so short on supply. But it's not where like the critical alpha of... It's not the main reason we're doing it. It's like stacking margin. It's more [00:20:35] S1: Availability and getting out.
00:20:35 · Harry Stebbings, Chase Lochmiller
[00:20:35] S1: Availability and getting out. [00:20:36] S0: Availability, being able to deliver on time. And then also, it gives us a critical innovation platform. By being able to, you know, start from first principles and say, okay, we can bring in raw materials and we can build anything we want in the world. What is the thing that we want to build to ultimately serve the workload? And so when you look at the campus in Abilene, the inspiration for that was like, okay, we wanna build a one gigawatt scale computer where you can interconnect a giant cohesive cluster of GPUs that can operate one cohesive workload across the entire campus. And what does it look like to actually build that thing from from the ground up? That's how, you know, the cores are designed. That's how the GPUs are designed. That's... You know, all of it is designed to run on one cohesive RDMA fabric. When we think about scaling of inference, you know, 2026 is really about the scaling of utilization of AI or utilization of LLMs and foundational models for useful tasks. You know, it's been this era of, you know, scaling of agents and and and tokens, token utilization. And, you know, during that moment of scaling infrastructures, people are finding utility in in these AI models, it's driving a lot more demand for inference. It's the era of the agent, the era of of inference. And for that, you don't need a a gigawatt scale computer to serve these models. You can do it with a much smaller cluster. What matters is actually how quickly can I go from not having anything to being able to deliver tokens? Like, so
00:22:00 · Harry Stebbings, Chase Lochmiller
[00:22:00] S1: what is my time to token? I love the analogy you said there of it's like whack a mole. What is the biggest supply constraint that we have today? [00:22:07] S0: Energy is definitely a key constraint. Power that's available for compute is a key bottleneck. Alongside that, labor is a very key bottleneck. There are a finite amount of skilled trade workers, electricians, welders, plumbers, construction workers in The United States. And then when you overlap, trying to find if I have access to energy in this unique place, can I get the labor to that place to actually bring the AI factory to life? You know, that that that is a big challenge. [00:22:37] S1: With the greatest of respect, if power and labor are the two biggest constraints or challenges, The UK is totally fucked. We have energy prices that are four times higher than anyone else, and we don't have labor. And we have humans. They just don't work. You and The US have both, but significantly more expensive than China, if we're honest. Power is not significantly more expensive. Really? No. You don't think that the government is subsidizing data centers in China with much cheaper energy to power that? I'm asking him, not telling. [00:23:07] S0: There are subsidies going in place, but in The United States, we are globally competitive with with power pricing. Labor, I'll give you. Labor in China is far cheaper than United States.