Power shortages and debt burden weigh on data center expansion
Oracle faces financial hurdles with its Stargate project, while power availability becomes the primary constraint for data center contracts. CoreWeave carries significant debt to acquire GPUs amid rising insurance costs.
Oracle is currently in negative free cash flow and holds $288 billion in leases that are not currently showing on its balance sheet. Half of its $630 billion in contracted future revenue is attributed to OpenAI. On September 24, Oracle notified its landlord of the Stargate campus in New Mexico that it may delay full rent payments by up to three years. The campus lacks sufficient power generation, relying on a natural gas pipeline from Texas that is blocked by a half-mile stretch of state-owned land in New Mexico where access has been denied. To achieve a 12% return on its $250 billion in invested capital, OpenAI would need to generate $30 billion annually in profit, not revenue. The discussion argues that Oracle is misallocating resources from its database and ERP businesses to fund the OpenAI contract, a move that is financially risky and operationally constrained by power shortages. Microsoft canceled the InScale UK data center site due to insufficient power availability. Power has become the primary constraint for data center contracts, replacing floor space as the main driver since 2020. A 42U or 48U rack can now hold only four fully populated AI servers before power limits are reached. New hardware deliveries from major networking providers are 60 weeks out. CoreWeave has taken on $35 billion in debt to acquire GPUs, with insurance costs estimated at 11% per year, a level described as unsustainable. Cloud providers are prioritizing capacity for major AI labs like OpenAI and Anthropic, leaving little for enterprise customers, and are implementing protections against model distillation at the inference layer. The situation is worse than previous hardware shortages because power is the limiting factor, and enterprise customers are competing for leftover capacity from major AI labs.