Chase Lochmiller on AI Data Centers, Energy and GPU Economics

Harry Stebbings begins with mountaineering. Crusoe co-founder and CEO Chase Lochmiller says the company has made ‘think like a mountaineer’ a core value: conditions change, so teams need resilience and must prepare for unexpected setbacks.

An expedition has a goal, but weather, equipment, illness or an avalanche can disrupt the original route. Lochmiller argues that a sound plan includes alternatives and the resolve to endure difficult stretches, rather than assuming everything will proceed as expected.

Safety is another part of the analogy. In mountaineering, he says, reaching the summit is optional but returning safely is mandatory. He applies that principle to construction safety and to protecting Crusoe’s digital platform.

Lochmiller says he has seen mountaineers take unnecessary risks and believes careful planning can prevent some tragedies. Stebbings then asks about difficult personal change, rather than only the risks of a climb.

Before 2018, Lochmiller usually knew his next step before leaving the previous one: school, college, work and graduate study each had a plan. When he left to climb Everest, he had no defined next chapter, despite having already achieved some financial success.

The blank slate felt both challenging and empowering. He considered managing his own money and enjoying life, but also asked what difficult and interesting problem he might be excited to pursue every day. That period helped give rise to Crusoe.

Stebbings asks whether financial security changes a founder’s appetite for risk. Lochmiller says knowing he could still cover basic needs if a venture failed made it easier to consider larger risks and ambitions. He presents this as his own experience, not a rule for every founder.

Lochmiller had once expected a life in theoretical physics, with research at MIT and Los Alamos. He says he found the pace of discovery slower than he wanted, felt a loss of purpose, and moved into quantitative finance after a recruiter explained the demand for mathematical problem-solving.

After achieving financial security, he again asked what to pursue. Stebbings frames Crusoe’s movement from Bitcoin to AI as a possible pivot and asks when a founder should persist versus change direction.

Lochmiller says building an AI platform was part of the original ambition, even though most early resources went to monetizing otherwise wasted energy through Bitcoin mining. His thesis was that compute and data would constrain AI, and energy would constrain compute.

He describes resource allocation as a probabilistic exercise: hold a view of the future, update it with evidence, and recognize that distant outcomes are less certain. Crusoe developed an AI platform alongside its Bitcoin work rather than devoting all resources to it immediately.

The company worked with researchers to understand infrastructure needs and launched Crusoe Cloud with paying customers in early 2022, he says. The November 2022 launch of ChatGPT then convinced him that demand for AI computing infrastructure would be much larger.

Lochmiller recalls Bitcoin mining moving from laptops and GPUs toward specialized hardware and much cheaper facilities. He says miners learned that they could accept less redundancy than conventional data centers, substantially reducing costs for that particular workload.

He believed AI infrastructure would also evolve, although its requirements differ from mining. As chips became more power-dense, he expected data-center designs to change; Crusoe’s work with Nvidia helped inform its planning, he says.

He argues that some AI workloads need not all run in traditional centralized hubs. In his view, computation can dominate the time to serve a neural network, making locations with abundant, inexpensive energy more attractive for certain workloads.

After ChatGPT, Crusoe increased investment in designing AI-specific facilities and sourcing energy, he says. Stebbings asks where the immediate constraint lies in the infrastructure supply chain.

Lochmiller’s answer is that there are too few places where GPUs can be installed and powered for AI workloads. He compares the supply chain to whack-a-mole: the limiting component changes over time, requiring developers to solve successive bottlenecks.

For the first two Abilene buildings, he says Crusoe committed to roughly a year for a little over 200 megawatts of compute capacity. One medium-voltage power distribution component had a supplier lead time of about 100 weeks, longer than the construction schedule.

Because Crusoe had invested in electrical manufacturing, its team sourced components and produced the equipment internally in 28 weeks, Lochmiller says. He presents vertical integration as a way to gain flexibility and visibility into cost when a bottleneck threatens delivery.

Asked about margins, he says electrical manufacturing is not normally an exceptionally high-margin business, even if scarcity has increased margins. For Crusoe, the main benefits are availability, timely delivery and the ability to design equipment around the workload.

Lochmiller describes the Abilene concept as a campus-scale computer: a very large interconnected GPU cluster designed to run a cohesive workload. For inference, he says, smaller clusters can be useful, making the time between deciding to build and delivering tokens especially important.

Asked about today’s biggest constraints, Lochmiller names power available for computing and skilled labor. Electricians, welders, plumbers and construction workers are finite, and an attractive energy location must also have enough workers willing and able to build there.

Stebbings asks about the relative costs of the United States and China. Lochmiller says he regards US power prices as globally competitive while acknowledging cheaper labor in China; he does not offer a detailed comparison of particular tariffs or projects.

On regulation, Lochmiller says the issue is something developers must navigate rather than eliminate. He supports thoughtful policy for large investments, while acknowledging that compliance can slow construction; he sees more factory-made components as one possible response.

Local communities, he says, care about jobs, energy bills, water, air quality and whether a project contributes to local services. These concerns shape the policy conversation because infrastructure is built in specific places rather than in an abstract national market.

Lochmiller argues that public discussion overstates water consumption for modern AI facilities. As an example, he says one roughly 140-megawatt Crusoe building in Abilene uses about as much water annually as ten single-family homes. That estimate is his description of a specific company design, not evidence that all data centers use little water.

He explains that the facility uses water in a closed cooling loop: water circulates through racks and a chiller rather than being continuously consumed in the same way as an evaporative system. He identifies staff and landscaping as uses of additional water at the site.

On electricity prices, Lochmiller claims that data-center investment has typically reduced local energy costs by encouraging new generation and spreading network costs across more demand. He does not present a study in the interview; effects on bills depend on the project, generation, grid and how costs are allocated.

Asked about legitimate local concerns, he acknowledges construction traffic, dust and noise. He also claims benefits from spending at local businesses, permanent jobs and tax revenue, including substantial projected revenue for local services and schools. Those benefits and figures are his company’s assessment.

Stebbings asks how many announced data-center projects might never go online. Lochmiller says that an estimate of 50% seems reasonable to him, while treating it as a judgment rather than a measured forecast.

He lists the things that can derail a project: land acquisition, permits, development entitlements, a large-load interconnection agreement with a utility and, for new generation, an air permit. Each step introduces uncertainty before a project can deliver capacity.

Lochmiller says data centers have become politically polarizing partly because they are the visible physical manifestation of AI. He acknowledges that people have legitimate questions about whether they will have work and be able to earn a living as the technology changes.

He argues that data centers create skilled-trade work, economic activity and tax revenue, and repeats his view that they can benefit communities. These are the developer’s claims; job effects, water use, energy prices and public costs can differ across places and over time.

Stebbings turns to the economics of GPU capacity. Lochmiller compares the price of a GPU-hour to a commodity with changing market prices, and notes the emergence of platforms seeking to trade computing capacity.

Crusoe manages the risk of financing hardware through a mix of contracts, he says. Long-term leases to creditworthy customers can provide cash flow and repay the investment over the contract period, while shorter contracts may pay more but leave renewal and utilization risk.

Managed inference and serverless fine-tuning are examples of higher-level services the company offers. Lochmiller says these can help developers adapt open models to domain-specific work or serve applications as they grow, typically on shorter contracts with higher margins for Crusoe.

He summarizes the company’s offerings as three products: data centers, GPUs and tokens. The mix matters because revenue and risk can arise at different points in the infrastructure chain rather than from renting hardware alone.

Lochmiller compares this strategy to integrated oil and gas companies operating upstream, midstream and downstream. He argues that margins can shift between layers: a lower commodity price may reduce returns in one part of the chain while improving economics in another.

At the time of the interview, he says managed GPU clusters have particularly high margins because supply is scarce. This is his observation about market conditions at that moment, not a promise that those margins will persist.

A take-or-pay agreement means a customer commits to pay for capacity whether or not it uses all of it. Lochmiller says Crusoe’s GPU rental agreements typically use that structure. It changes utilization risk but does not by itself eliminate customer or renewal risk.

Stebbings asks what would happen if consumer demand weakened. Lochmiller remains optimistic that AI will find uses across many parts of the economy and in scientific discovery. He treats that breadth as a reason to be bullish, though those future uses and their economics remain uncertain.

On chip depreciation, Lochmiller says Crusoe currently uses a six-year accounting cycle. The company also wants multiple ways to earn from its long-term investments in data centers, energy and chips, which helped motivate its managed AI services.

Managed services hide the underlying chip from the end customer. Lochmiller expects demand both for cutting-edge silicon and for cheaper alternatives, and believes older, slower chips can still provide useful intelligence at lower cost.

He therefore thinks a portfolio of services may extend the useful economic life of chips beyond the six-year accounting schedule. The statement is a business forecast: accounting depreciation and the future resale or service value of hardware are separate questions.

Lochmiller recalls concern in 2023 that Nvidia Hopper GPUs might not be valuable after three years. He says prices for using those chips were higher three years later than when they were new, and credits application developers with finding ways to make older compute valuable.

Crusoe builds demand forecasts through conversations with customers rather than simply buying every chip available, he says. A growing challenge is asking customers to commit further in advance as infrastructure lead times lengthen.

To shorten delivery, the company is pursuing smaller modular, manufactured AI data centers for certain workloads. Lochmiller says more just-in-time capacity could help startups that otherwise have to predict their compute needs years ahead.

Stebbings asks whether the future has ever felt so uncertain. Lochmiller says he sees uncertainty about the form of change but is confident that AI will become important to work and productivity. He is optimistic about continued progress in model capabilities.

In illustrating his optimism, Lochmiller points to apparent advances on hard mathematical problems, including a claim about Millennium Prize problems. He does not establish a verified solution in the interview, so that claim should not be treated as proof that any particular prize problem has been solved.

Stebbings raises uncertainty about agents, professional work and the pace of change. Lochmiller agrees that the future of work is uncertain in detail, but predicts that AI capabilities will change productive work substantially. The two perspectives differ in emphasis, not in their expectation of change.

The discussion turns to whether the lowest-cost producer of intelligence wins. Lochmiller says cost per token is a useful indicator, but tokens do not all carry equal value. The task and the model’s efficiency matter when comparing what a customer actually receives.

He also points to throughput and latency, including time to first token and time to the end of a response. Cost, software and infrastructure all influence those measurements, and he argues that owning multiple layers gives a provider more room to adjust its economics.

A GPU is an expensive asset, so keeping it busy matters. Lochmiller highlights memory management and the key-value cache: serving systems can reuse stored attention-related states, but the cache can grow beyond memory on a single accelerator and must be managed across storage layers.

He describes coordinating GPU memory, system memory, fast local storage and cluster storage to move data efficiently. In his view, handling this memory hierarchy well helps increase utilization and improve latency and throughput; implementation details depend on the workload.

Not all inference providers are alike, he says. Open-source software such as SGLang and vLLM provides general building blocks, but providers can differ in workload optimization, tools, features and user experience.

Asked about Fireworks, Lochmiller praises its work and says it serves some workloads very well. He sees both competition and opportunities for collaboration across the infrastructure stack, depending on which layer or customer need is involved.

He returns to the oil-and-gas analogy: companies that compete to produce a commodity may also sell transport capacity or services to each other. Similarly, he believes Crusoe can work with firms that compete with it elsewhere in AI infrastructure.

On open versus closed models, Lochmiller estimates that customers spend more money on closed frontier models while generating more tokens with open models. He offers this as his market perspective, not as a verified industry-wide measurement.

He expects open models to matter for data sovereignty and customers who want control over their models. He also sees potential in private data, though he says organizations can incorporate such data into either closed frontier systems or models they run themselves.

Lochmiller predicts a combination of approaches rather than one dominant model strategy. He cites domain-specific post-training and says frontier labs are also trying to expand their models’ knowledge and capabilities. These are possibilities in a fast-changing market, not established outcomes.

Looking five years ahead, he describes managed inference as an easier way for companies to access intelligence, including custom models adapted to private data. The service, as he envisions it, would hide much of the complexity of GPUs, data repositories, model choice and query routing.

Asked how he combines work and parenting, Lochmiller says family time requires explicit priorities. He sometimes takes less convenient flights, including overnight flights, to avoid missing a bedtime, and tries to spend weekends on activities with his children.

He also blocks a specific hour in the morning when he is home: colleagues know he is unavailable while he makes breakfast and gets his children ready for school. The protected time is his practical answer to a demanding travel schedule.

Lochmiller repeats his belief that data centers should be welcomed for their economic and community contributions. This is the position of a developer with a direct interest in such projects; local residents may experience and assess their effects differently.

Asked about wealth inequality, Lochmiller says he is not worried about inequality itself and argues that technological investment can benefit many people even when value is concentrated. Stebbings raises the unusual concentration of market value; their exchange presents competing concerns rather than resolving the distributional question.

One view Lochmiller says he has changed is that many supposed long-term competitive moats are durable. During rapid technological change, he believes advantages can be temporary and that moving quickly and adapting matter more than assuming a position is permanently protected.

Would Crusoe be public by the end of 2028? Lochmiller says he does not know. The company needs substantial capital to build facilities and deploy chips, and he sees advantages in accessing public markets eventually; he does not commit to a date.

Asked whom he would like on the board, he names Michael Dell, whom he admires for his business instincts and personal qualities, while acknowledging that Dell is not going to join. Stebbings then asks what the founder’s 2018 self would find most surprising.

Lochmiller says the scale at which Crusoe’s energy-first AI infrastructure thesis developed would have seemed hard to believe in 2018. He recalls how difficult it once was to imagine a company with 150 people and says Crusoe was approaching 2,000 employees at the time of the interview.

He closes by expressing gratitude to employees who chose to spend years of their careers building the company. Their decision to invest time and effort alongside Crusoe is something he says he does not take for granted.