Three findings from the past two weeks of daily briefs converge on the same conclusion: CoreWeave's reported pricing is increasingly difficult to reconcile with the marginal power costs it will face on new capacity.
The pricing arbitrage. CoreWeave is currently pricing H100 in Dublin and Oregon at $5.77–$5.96/hr while AWS holds H100 on-demand in us-virginia at a flat $8.69/hr ceiling — a sustained 32–35% discount. On hardware margin alone, this is a defensible position. The open question is whether the discount can hold once power costs feed through to the catalog.
The Iowa pricing anomaly. The H100 on-demand floor in Iowa collapsed earlier in May under what appears to be single-provider monopoly pricing at ~$0.45/hr — an 84% discount to the broader US market. The discount provider is consistent with an aggressive capacity-fill strategy by a leveraged operator with concentrated customer commitments. If that actor is CoreWeave (one of several candidates including Lambda and Crusoe), the gap between catalog price and marginal cost on that capacity is uncomfortably thin.
The structural exposure. CoreWeave's Q1 2026 10-Q (filed May 8, 2026) discloses "significant customer concentration risk" with Customer A and Customer B representing material portions of revenue across both Q1 2025 and Q1 2026. These customers are widely believed to be Microsoft and Meta — both of which are simultaneously building custom ASIC capacity (Azure Maia, Meta MTIA via Broadcom). Every dollar spent on in-house chips is a dollar not spent renting H100 capacity from CoreWeave. The company carries no disclosed long-term power purchase agreement or fixed-rate hedging in regulatory filings. Senator Warren's December 2025 power investigation named CoreWeave alongside the hyperscalers as exposed to data-centre power-cost escalation.
The composite picture: a capacity-leveraged business with concentrated demand, undisclosed power hedging, and aggressive discount pricing that may be subsidising margin to fill racks ahead of demand normalisation.