Interview, Fireside Chat
Crusoe CEO: Why Everyone Gets GPU Depreciation & AI Energy Costs Wrong
- AI infrastructure will evolve into a distributed model anchored in regions with low-cost, abundant energy, with the industry shifting from scaling tokens to maximizing infrastructure utilization for useful tasks in approximately two years, culminating in an inference and agent-scaling era defined in 2026.
- Significant operational delays are anticipated for data center projects, with an estimated 50% of planned facilities failing to launch due to logistical hurdles involving permits, land acquisition, and utility interconnection agreements.
- Financial structures will rely on long-term contracts with credit-quality customers for five-year repayment and cash flow generation, while shorter-term contracts offer higher margins with associated renewal risks, and GPU prices will fluctuate as compute becomes a traded commodity with futures.
- The economic lifecycle of hardware will extend beyond the standard six-year depreciation cycle through managed services and the monetization of older silicon, supported by the expectation that frontier silicon demand will persist alongside growing demand for older-generation chips.
- Revenue models will leverage "take or pay" contracts for GPU rentals to guarantee capacity payments, while token efficiency will vary by use case, with dollar-per-token serving as a key metric for identifying the lowest cost producer of intelligence.
- Operational efficiency will improve via the deployment of small modular manufactured data centers for just-in-time delivery and the vertical integration of power distribution, reducing lead times from 100 weeks to 28 weeks to build a one-gigawatt scale computer in Abilene using a single RDMA fabric.
- The competitive landscape is characterized by rapidly evolving technology that renders static advantages illusory, though private domain-specific data is expected to drive performance gains, prompting companies to host custom models alongside frontier models while open-source models may generate more tokens than closed-source counterparts.
- The company plans to reach 2,000 employees focused on infrastructure development, potentially enter public markets by the end of 2028 to access scaled capital for AI factories, and operate as a vertically integrated "AI supermajor" to hedge against commodity price cycles.
- Macro-economic expectations include the creation of massive blue-collar employment and economic growth rather than job replacement, alongside the emergence of a 2018-era AI cloud vision at an unimaginable scale, with consumers potentially relocated to better conditions for large-scale U.S. projects.