20VC: The Future of Datacentres: What You Need to Know | Why Everyone Gets GPU Depreciation and AI's Energy Costs Wrong | Who Really Makes Money From AI & Why Most Moats Don't Exist with Chase Lochmiller

Summary

Chase Lochmiller says Crusoe was built around the idea that AI compute would become constrained by energy, leading the company to pursue low-cost, distributed data-center locations rather than centralized internet-style hubs. He frames Crusoe as a vertically integrated “AI supermajor” that monetizes three layers — data centers, GPUs, and tokens — and says the GPU is the most valuable asset in the stack. The episode goes deep on Abilene, Texas, where Crusoe has had to solve long-lead supply chain problems, vertically integrate electrical manufacturing, and design a 1 GW-scale campus around a cohesive RDMA fabric. Lochmiller argues 2026 will be the “era of inference,” where time-to-token, throughput, latency, and KV-cache management matter more than giant training clusters. He also pushes back on claims that data centers drive up power bills or consume lots of water, saying modern AI factories are close to zero-water and can even lower local energy prices through added generation investment.

Notes

• Crusoe’s core thesis is that AI infrastructure will be distributed where energy is cheap and abundant, not centralized in legacy hubs.
• Crusoe says it monetizes three product layers: data centers, GPUs, and tokens.
• Lochmiller treats the GPU as the highest-value asset inside the data center.
• A mountaineering mindset is a stated company value for resilience, safety, and contingency planning.
• He links the metaphor to operating in both physical and digital environments with failure planning.
• Crusoe says it was founded as an AI platform from the start, with compute, data, and energy as linked bottlenecks.
• He says the original premise was that compute and data would bottleneck AI, and energy would bottleneck compute.
• The pivot point for Crusoe’s AI focus was ChatGPT’s launch on 11-30-2022.
• Crusoe says AI workloads do not need legacy centralized geographies because inference time is mostly compute time.
• The company’s model is built around low-cost geographies and distributed capacity.
• Abilene’s first two buildings were a little over 200 MW of compute capacity, built under a one-year target.
• A medium-voltage power distribution center had a 100-week lead time, creating a major supply-chain bottleneck.
• Crusoe says vertical integration let it build the same component in 28 weeks instead of 100.
• Abilene is designed as a 1 GW-scale computer with a unified RDMA fabric across the campus.
• He sees 2026 as the era of inference, token utilization, and time-to-token.
• Energy and skilled labor are the main near-term constraints on AI factory buildouts.
• He says U.S. power prices are globally competitive, even if labor is cheaper in China.
• Policy is framed as friction to navigate, not a fatal blocker.
• He says modern AI data centers are close to zero-water and use closed-loop cooling.
• He argues data center buildouts can lower local energy prices by adding generation and better utilizing infrastructure.
• Crusoe expects to contribute more than a third of Taylor County tax revenue.
• He estimates roughly 50% of planned data centers may never get built.
• Permits, entitlements, land, interconnection, and air permits are key failure points for projects.
• Data centers have become a political flashpoint because of AI job fears.
• He says the factual case for data centers is positive on jobs, energy, water, and taxes.
• Crusoe portrays data centers as a driver of blue-collar job growth and reindustrialization.
• GPU-hour pricing is described as commodity-like, with compute futures already emerging.
• Crusoe uses a portfolio of long-term contracts, shorter-term contracts, and software-like services to balance margin and risk.
• Crusoe’s strategy is to stack value across data centers, GPUs, and tokens, analogous to an oil-and-gas value chain.
• He compares Crusoe’s vertical integration to upstream, midstream, and downstream oil and gas businesses.
• Managed GPUs and managed compute clusters are the highest-margin layer right now because supply is tight.
• GPU rental contracts are typically take-or-pay.
• Lochmiller says the market is still early and model capability keeps increasing.
• He ties AI infrastructure spending to scientific discovery, especially materials and drugs.
• Crusoe uses a six-year depreciation schedule for its assets today.
• Managed AI services are meant to extend the monetization life of underlying chips and infrastructure.
• Older chips can still be monetized if they deliver cheaper intelligence.
• He says Hopper utilization pricing is still higher three years after Crusoe’s 2023 purchases.
• Customer forecasting is being pulled further out because AI infrastructure lead times are long.
• Crusoe wants to reduce time-to-token with small modular manufactured AI data centers.
• He says scaling laws still hold and AI capability continues to advance.
• He points to AI systems solving Millennium Prize problems as evidence of progress.
• He says AI is becoming a critical engine for work and productivity despite uncertainty about job structure.
• He says AI already reaches human-level or better productive capacity in some areas.
• Dollars per token is useful, but token efficiency matters too.
• Inference operations are optimized around throughput, time to first token, and time to last token.
• Vertical integration lets Crusoe stack margin across the process and flex pricing.
• GPU utilization is critical because an idle GPU is wasted capital.
• KV-cache management across HBM, DRAM, NVMe, and object storage is central to inference efficiency.
• Not all inference providers are equivalent, even with strong open-source tooling in the stack.
• Crusoe sees more opportunity to collaborate across the stack than to compete.
• He says open source and closed source will coexist, with different economics and use cases.
• Open source matters for data sovereignty and for unlocking private data.
• He expects a mixed future of frontier models, open source, and custom domain-specific systems.
• Managed inference abstracts infrastructure complexity across GPUs, data repositories, models, and routing.
• He prioritizes family time by taking red-eyes and protecting a morning block at home.
• He says data centers should be celebrated locally because of broad community benefits.
• He says large investments create broad benefits that outweigh inequality concerns in the abstract.
• He argues the value created by abundant intelligence will exceed Crusoe’s market cap by multiples.
• Lochmiller says most moats are illusion and ephemeral.
• Speed and adaptation matter more than entrenched defenses during rapid technological change.
• He thinks Crusoe may fit public markets because of the capital intensity of AI factories and large GPU clusters.
• He says the 2018 version of himself would be shocked that the energy-first thesis worked.
• Crusoe is approaching 2,000 employees.