Chase Lochmiller on AI Data Center Economics

Chase Lochmiller spent his twenties training to become a theoretical physicist at MIT and Los Alamos National Lab, chasing what he calls a monastic life of discovery. He decided the field moved too slowly, switched into quantitative finance, made money, and in 2018 walked away to climb Mount Everest with nothing planned for afterward (08:18). That blank slate, he says, is where the idea for Crusoe Energy came from. Crusoe is now one of the companies building the physical plumbing of the AI boom: the data centers, the chips inside them, and the software layer that turns both into billable tokens. The company just closed a $3.9 billion Series F at a $30.9 billion valuation (01:16).

Lochmiller runs the company on a philosophy borrowed from the mountains. Crusoe has written "think like a mountaineer" into its official values: expect things to go wrong, keep a plan B and a plan C, and treat safety as non-negotiable, which matters more now that the company is pouring concrete and wiring high-voltage power at industrial scale.

"Getting up to the top is optional, getting down is mandatory. We try to practice that safety culture here at Crusoe." — Chase Lochmiller [06:40]

Crusoe did not start as an AI company in practice, even though Lochmiller says that was always the destination. In its early years the business monetized stranded energy, mostly natural gas that would otherwise be flared at oil wells, by running Bitcoin mining rigs on it. That detour taught the company something useful: how cheap a data center becomes once you strip out redundancy. Lochmiller says the Bitcoin industry cut data center costs by roughly 98% by moving from colocation-grade facilities, built to the "five nines" reliability standard, down to bare, no-frills buildings he calls "chicken coops" (15:53). Crusoe launched a cloud platform for AI researchers in early 2022, but the moment that reordered the company's priorities, he says, was November 30, 2022, the day ChatGPT launched (14:33).

Why Crusoe Builds Its Own Parts

The case Lochmiller makes for vertical integration (owning more of the supply chain rather than buying each piece from a vendor) is concrete rather than ideological. When Crusoe committed to building its first data center campus in Abilene, Texas, in one year, against competing bids of two and a half years, it hit a bottleneck: a power distribution center, the equipment that steps high-voltage power down to something a server rack can use. The standard market lead time for that part was 100 weeks. Crusoe's team instead sourced components and manufactured the units itself, cutting delivery to 28 weeks (19:14). Lochmiller borrows a phrase from Elon Musk to describe the discipline behind this: the "idiot index," the ratio between what raw materials cost and what the finished product sells for. Owning more of the chain does not mean fat margins on manufacturing, which he says is not a high-margin business even now. What it buys is speed, availability, and a clear view of true costs across the whole stack.

That same logic reshaped where AI infrastructure gets built. Traditional internet infrastructure clustered around hubs like Northern Virginia because network latency mattered. Lochmiller argues AI workloads behave differently: most of the time to answer a prompt is spent computing inside the data center, not traveling to it. That breaks the old anchor to centralized hubs and opens the door to building wherever power is cheap and abundant.

"The infrastructure to support AI wasn't gonna be centralized. It was gonna be distributed where energy was low cost and abundant." — Chase Lochmiller [17:16]

Lochmiller spends real time pushing back on what he calls misinformation about data centers. On water: modern AI buildings use closed-loop cooling, and he says one of Crusoe's 140-megawatt buildings in Abilene uses about as much water annually as 10 single-family homes, mostly from staff using sinks and bathrooms (25:05). On electricity prices: he argues data centers typically bring local energy costs down, not up, because new generation investment gets spread across more megawatts and the same transmission infrastructure (26:13). Where he is less bullish is on the pipeline itself. Asked what share of currently planned data centers will actually get built, he puts it at roughly 50%, citing failures in permitting, land acquisition, and utility interconnection agreements as the usual killers (28:29).

The AI Super Major

Lochmiller frames Crusoe's business as three products sold across one value chain: data centers, GPUs, and tokens (34:36). He draws the analogy directly from oil and gas, where companies like Exxon and Chevron are "super majors" spanning upstream drilling, midstream transport, and downstream refining. Exxon famously does not hedge its oil exposure, because vertical integration is itself a hedge: when oil prices fall, upstream margins shrink but downstream margins on gasoline and plastics expand. Lochmiller sees the same seesaw happening across electrical infrastructure, chips, and inference services, and says margin right now is highest in managed GPU rental, because supply is so short (36:23).

That shortage shows up starkly in chip economics. Crusoe depreciates its GPUs on a standard six-year cycle (38:43). When the company placed large orders for Nvidia's Hopper-generation chips back in 2023, the market doubted they would hold value past year three. Three years later, Lochmiller says, the rental rates customers pay for those same Hopper chips are higher than when the chips were brand new, because demand for compute, even older compute, has outrun supply (40:17). The company's bet is that layering managed services like inference and fine-tuning on top of aging hardware can stretch the useful economic life of a chip well beyond its accounting depreciation schedule.

Underneath that bet sits a genuinely technical problem: keeping expensive GPUs busy. Lochmiller calls the GPU the single most valuable object in the building, so idle GPU time is value burning for no reason. The fix involves managing something called the KV cache, short for key-value cache, a lookup table that stores intermediate results from a neural network so it does not have to redo the same math twice. As that cache grows past the fast on-chip memory (HBM), it has to be shuffled across slower system memory and solid-state storage without stalling the chip, which Lochmiller treats as a core engineering advantage for Crusoe's managed inference product (46:16).

"Most moats are a illusion. Most moats don't exist. Most of them are ephemeral." — Chase Lochmiller [55:00]

That line lands as the thesis statement for the whole conversation. Lochmiller has watched the market obsess over defensibility, the classic venture capital question of what protects a company's margin over time, and concluded that in AI infrastructure almost nothing stays protected for long. Open-source models are already eating usage even as closed models keep more dollars: Lochmiller says customers spend more money on closed-source frontier models but generate more total tokens on open-source ones, because open models are cheaper to run at volume (49:27). His own answer to the moat question is not a technology but a habit: move fast enough, and adapt efficiently enough, that the next bottleneck gets solved before it becomes someone else's advantage. It is the same instinct he describes on a mountain, where the plan matters less than the willingness to revise it when the weather changes.