Crusoe CEO: Why Everyone Gets GPU Depreciation & AI Energy Costs Wrong
Summary
Crusoe’s founding thesis was that energy would become the bottleneck to scaling AI, so compute should migrate to wherever power is cheap and available. Bitcoin mining initially monetized excess energy while Crusoe built its AI platform; ChatGPT’s November 30, 2022 launch then made infrastructure demand unmistakable. “The infrastructure to support AI would not be centralized.”
The data-center shortage is ultimately a shortage of places where GPUs can be connected, powered, and put to work. Crusoe promised roughly 200 MW in Abilene within one year versus the next-best 2.5-year proposal, then manufactured a medium-voltage distribution system internally in 28 weeks when suppliers quoted 100. Vertical integration is less about component margin than “availability” and the ability to route around shifting bottlenecks.
Chase Lochmiller argues that the loudest community objections to AI data centers—water use and higher power prices—are contradicted by Crusoe’s operating evidence. A roughly 140 MW Abilene building uses about as much annual water as 10 homes because GPU cooling is closed-loop; he says energy prices have typically fallen where large facilities stimulate new generation. The real near-term costs are traffic, dust, and noise, while Crusoe expects to provide more than one-third of local tax revenue and more than double school funding.
Crusoe manages commodity risk by selling three products across the AI stack: “data centers, GPUs, and tokens.” Five-year GPU contracts with reliable customers provide payback and cash flow, shorter contracts carry higher margins and renewal risk, and managed training or inference can monetize the same infrastructure at still higher margins. The analogy is an oil supermajor whose profits migrate between upstream, midstream, and downstream as commodity prices change.
The interview’s sharpest challenge to consensus is that GPUs may remain economically useful far longer than investors assume. Crusoe depreciates chips over six years, yet Hoppers bought in 2023 now command higher usage prices than when new; older, slower chips can still deliver cheaper intelligence when services abstract away the hardware. “People underestimate the ingenuity of applications and developers in turning computing power into value.”
The real inference race is not simply dollars per token, because tokens differ in value and customers also optimize latency, throughput, and utilization. Since GPUs are “the most valuable thing in the entire data center,” idle time is burning money; KV-cache orchestration across HBM, DRAM, NVMe, and object storage becomes a core economic capability. Lochmiller expects customers to spend more on closed models but generate more tokens through open-source models.
Lochmiller has become skeptical that durable AI moats exist amid accelerating technical change. “Most of these moats [are] an illusion”; the lasting advantage is moving quickly and adapting to the changing chessboard. The host introduced Crusoe as having raised $3.9 billion in a Series F at a $30.9 billion valuation; Lochmiller sees public-market capital as ultimately advantageous but is unsure whether Crusoe will be public by the end of 2028.
Deep dive
1. Crusoe treated energy as AI’s constraint before AI became the headline
Lochmiller’s founding logic ran backward from intelligence: AI and machine learning would transform every industry; computation and data would become the constraints on that transformation; and energy would become computation’s bottleneck. Crusoe therefore searched for inexpensive, available power that could be monetized through computing.
Bitcoin was not, in his telling, a discarded first act. It was the best available mechanism for monetizing excess energy while Crusoe developed an AI platform, worked with researchers at MIT and Stanford, and learned what infrastructure AI users actually valued.
Crusoe Cloud reached its first paying customers in early 2022. ChatGPT’s November 30, 2022 launch was the Bayesian update: “AI was here,” use cases would proliferate, and demand for AI-compute infrastructure would be much larger than previously assumed.
The mountaineering culture frames that evolution: plan A through plan D, resilience through an expedition, and safety above summit fever. “Climbing to the top is optional, but descending is mandatory”—a lesson Lochmiller applies to physical construction and digital security alike.
2. AI compute can leave legacy hubs and follow cheap power
Bitcoin mining supplied the architectural precedent. Miners progressed from laptops to GPUs, traditional “five nines” colocation and specialized ASICs, then realized the workload could tolerate far less reliability; replacing full data centers with “chicken coops” cut total facility costs by roughly 98%.
Years earlier, NVIDIA’s roadmap moved from 150–200-watt chips toward 300 and potentially 600 watts. Lochmiller concluded that rising power density would change the buildings and data-center designs needed to support compute.
AI also weakens the case for concentration in hubs such as Northern Virginia: much of a neural network request’s service time is computation inside the facility, not network access. That latency profile opens new geographies, allowing infrastructure to be “distributed where energy was cheap and available.”
3. Vertical integration turns supply-chain delay into delivery speed
The shortage is concrete: “There are simply no places where you can plug in GPUs, turn them on, and run AI workloads.” The limiting component changes over time, making the supply chain a game of whack-a-mole across electricity, equipment, labor, permits, and construction.
For Abilene’s first two buildings—slightly more than 200 MW—Crusoe committed to one year when the next-nearest offer was 2.5 years among 34 other developers. A required 34–34.5 kV distribution system carried a 100-week supplier lead time, so Crusoe vertically integrated electrical manufacturing and delivered it in 28 weeks.
Lochmiller rejected the suggestion that manufacturing’s chief benefit is excess margin. It provides availability, schedule control, cost visibility through Elon Musk’s “idiot index,” and a platform for designing from first principles rather than accepting whatever standardized equipment suppliers offer.
At Abilene, the goal is a gigawatt-scale computer operating as one cohesive GPU cluster over an RDMA fabric. In the “era of agents” and inference, smaller modular facilities become especially relevant: the operative metric becomes how quickly Crusoe can go from nothing to generating the first token.
4. Energy and labor constrain deployment more than a single component
Lochmiller named available compute power as the principal bottleneck, with skilled labor—electricians, welders, plumbers, and construction workers—close behind. Cheap energy is useless if the workforce cannot be brought to a remote site to turn it into an operating “AI factory.”
Harry Stebbings pushed on China’s cheaper labor, subsidized power, and ability to clear development obstacles rapidly. Lochmiller agreed labor is cheaper and rapid relocation or site clearance can advantage developers, but maintained that US energy prices remain globally competitive.
Regulation, in his view, is not simply a problem to eliminate: large investments require rules rather than a “Wild West.” The challenge is friction, which strengthens Crusoe’s preference for manufacturing repeatable infrastructure instead of treating every deployment as a grand bespoke construction project.
5. Community opposition mixes real disruption with disputed resource claims
Local politics asks practical questions: Will residents receive jobs and tax revenue? Will the project raise energy costs, consume water, or cause pollution? Lochmiller said developers must demonstrate that they are worthy members of the community, not merely defend aggregate AI demand.
His strongest rebuttal concerned water. Each of Abilene’s first eight buildings budgets roughly 140 MW, yet one uses about as much annual water as 10 single-family homes; most consumption comes from staff and landscaping because GPU-cooling water circulates through a closed-loop system.
On electricity, he said markets receiving large data-center investments have typically seen community prices decline as new generation spreads more megawatts through the existing transmission and distribution infrastructure. Crusoe supports having large facilities help deploy the additional generation they require.
The concession matters: construction brings traffic, dust, and noise. Lochmiller argues those temporary costs give way to permanent jobs and taxes; Crusoe expects to generate more than one-third of Taylor County and Abilene tax revenue and more than double the revenue directed to schools.
6. Three products hedge the AI infrastructure commodity cycle
Lochmiller views hourly GPU capacity as an emerging traded commodity. Crusoe therefore combines five-year contracts with reliable customers expected to pay through the period, shorter higher-margin but renewal-sensitive agreements, and managed services such as managed inference and serverless training that monetize infrastructure through software.
GPU leases are generally “take or pay”: a customer reserving capacity pays whether it uses that capacity or not. Stebbings questioned whether weakening ChatGPT demand might expose those obligations; Lochmiller’s rebuttal was that consumer chat is only one outlet for compute across science, materials, medicine, and the wider economy.
Crusoe’s three monetizable products are “data centers, GPUs, and tokens.” Lochmiller compares the structure to Exxon or Chevron: vertical integration provides a hedge to some extent because margin can migrate among electricity, facilities, chips, and services as prices change.
Managed GPU clusters currently carry “incredibly high margins” because supply remains scarce. Lochmiller nevertheless expects profitability to shift through the stack, just as lower crude prices can hurt an oil producer’s wells while expanding downstream margins in gasoline, plastics, and other finished products.
7. GPU longevity and inference efficiency challenge simple depreciation models
Crusoe depreciates chips over six years, but Lochmiller believes services can extend their economic life further. Frontier silicon will retain demand, while older, slower generations can supply lower-cost intelligence once customers consume an abstracted service rather than selecting the underlying chip.
Investors doubted whether the Hoppers Crusoe bought in 2023 would retain value beyond year three. Three years later, Lochmiller said usage prices are higher than the rates in effect when they were new; early financing had required very quick payback and significant collateral.
Dollars per token is useful but incomplete because “not all tokens are the same.” Customers also optimize tokens per second, time to first token, time to last token, and total cost; winning requires both owning efficient infrastructure and keeping its most expensive asset—the GPU—from sitting idle.
KV-cache management illustrates the hidden systems work. Efficiently moving reusable computation across GPU HBM, system DRAM, NVMe drives, and cluster object storage can materially improve utilization, latency, and throughput instead of repeatedly recomputing known token histories.
8. Open models gain volume while durable moats give way to adaptation
Lochmiller’s split forecast is that customers spend more money on closed models than open ones but generate more tokens through open source. Open models matter for data sovereignty and private data, though he expects companies to mix frontier services, owned models, and specialized models trained on private data rather than converge on one architecture.
Stebbings pressed whether inference specialists such as Fireworks displace Crusoe. Lochmiller conceded competition at several stack layers and praised Fireworks’ user experience, but returned to the supermajor analogy: firms can compete in one layer while selling to the same customers or providing capacity to competitors in another.
Managed inference ultimately sells easy access to intelligence while hiding routing, model choice, distributed GPUs, and multiple data stores. Startups are currently being asked to book capacity for 2028; Crusoe’s proposed answer is smaller modular data centers that reduce delivery time and provide clusters just in time.
Lochmiller’s larger change of mind is that “most moats simply do not exist.” In an accelerating market, advantages are ephemeral; endurance comes from speed and adaptation. He sees Crusoe as ultimately better off in public markets because of their access to capital, but answered “I’m not sure” when asked about an IPO by the end of 2028.