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

What It Covers

Chase Lochmiller, CEO of Crusoe Energy (valued at $30.9B after a $3.9B Series F), breaks down the full AI infrastructure value chain — data centers, GPU compute, and managed inference — while correcting widespread misconceptions about energy costs, water usage, chip depreciation, and where durable competitive advantages actually exist in AI infrastructure.

Key Questions Answered

• •Vertical Integration as Supply Chain Weapon: When building Crusoe's Abilene campus, a single power distribution component had a 100-week market lead time. By manufacturing it internally, Crusoe delivered in 28 weeks. Vertical integration across electrical manufacturing, data centers, GPUs, and inference isn't primarily a margin play — it's an availability and speed-to-deployment strategy that unlocks the ability to win contracts competitors cannot fulfill.
• •Three-Layer Revenue Model: Crusoe monetizes AI infrastructure across three distinct products — data centers (sold to operators), GPUs (rented on take-or-pay contracts, typically five-year terms to credit-quality customers), and tokens (managed inference, shorter-term, higher-margin). Blending these margin profiles — similar to how Exxon hedges oil exposure through vertical integration — creates a naturally balanced portfolio that buffers against commodity price swings in any single layer.
• •GPU Depreciation Is Misunderstood: The industry standard six-year depreciation cycle for GPUs is widely doubted, but Hopper GPUs purchased in 2023 now command higher rental rates than at launch — three years later. Early lenders demanded rapid payback assuming rapid obsolescence. Building managed inference services that abstract away the underlying chip extends monetization well beyond hardware refresh cycles and reframes GPUs as long-duration assets rather than depreciating liabilities.
• •Energy Prices Fall, Not Rise, Near Data Centers: Contrary to the dominant public narrative, markets where large-scale data centers have been built show energy price decreases for local communities. More generation capacity gets added, amortized across existing transmission infrastructure, lowering per-unit costs. Crusoe's 140-megawatt Abilene buildings also use roughly the same annual water as 10 single-family homes, using closed-loop liquid cooling — making both the energy and water scarcity arguments factually incorrect.
• •AI Infrastructure Will Distribute, Not Centralize: Traditional web infrastructure concentrated in hubs like Northern Virginia because network latency dominated. AI inference workloads are compute-bound inside the data center, not latency-bound to the user. This opens geographies where low-cost, abundant energy exists. Crusoe's strategy targets these distributed energy-rich locations rather than competing for constrained power in established data center markets — a structural advantage as power and skilled labor remain the two primary bottlenecks.
• •Most Moats Are Illusions: Lochmiller revised his view on durable competitive advantages over the past 12 months. During periods of accelerating model capability and rapid infrastructure change, most moats are ephemeral. The actual competitive advantage is organizational speed and adaptability — the ability to re-plan quickly when conditions shift, analogous to mountaineering contingency planning (Plan A through Plan D). Execution velocity and supply chain control matter more than any static structural advantage.