Here is the error: Bernstein’s report claims Core Scientific’s AI hosting returns are distorted by an accounting artifact—a capital injection from its client, CoreWeave, that inflates the apparent yield. But the real problem is not the number; it’s the contract structure. I’ve spent 13 years auditing smart contracts, and I know a state transition when I see one. When a revenue stream depends on a counterparty’s financing round, you are no longer measuring operational efficiency—you are measuring financial engineering. And in blockchain terms, that is a reentrancy vulnerability waiting to execute.
Context: The Mechanics of the Hybrid Hosting Model
Bitcoin miners like Core Scientific operate large-scale data centers with high-capacity power and cooling infrastructure. In the post-halving era, many are pivoting to colocation services for AI workloads—renting out their facilities to companies like CoreWeave, which need massive GPU clusters. The business model is straightforward: charge a monthly fee for space, power, and connectivity. But Core Scientific’s contracts with CoreWeave include a twist: CoreWeave provided capital upfront to fund facility upgrades, and the hosting fees are structured in a way that effectively passes that capital back to Core Scientific as "revenue."
Bernstein’s analysis reveals that this arrangement inflates Core Scientific’s reported returns by approximately 40% compared to peer miners without such capital infusions. The market has priced this premium as sustainable, but the underlying financial geometry is fragile. Think of it as a token distribution that allocates voting power to a single whale—here, the whale is CoreWeave’s ability to raise more capital.
Core: A Forensic Analysis of the Capital Loop
Let’s deconstruct the cash flow. In a standard colocation contract, the miner incurs operating expenses (power, cooling, labor) and receives a fixed monthly fee. The profit margin is the difference. In Core Scientific’s case, the contract includes a variable component tied to CoreWeave’s financing milestones. I have modeled this as a pseudo-code state machine:
