Reported AWS Power Constraints Delay Deployment of NVIDIA GB300 Racks
Редакция ЦОД:Главное
Amazon is reportedly taking delivery of NVIDIA GB300 racks before AWS data-center installations are complete, while power shortages prevent a significant share of the equipment from being deployed on schedule. The racks are said to be held in interim warehouses, extending the gap between hardware procurement and available AI compute capacity.
- AWS is reportedly receiving NVIDIA GB300 racks before data-center installation is complete.
- Available electrical capacity, rather than rack supply alone, is reportedly delaying deployment of some equipment.
- Amazon reportedly imported about $5 billion of equipment in August under its Early Rack Receipt process.
- Further GB300 purchases of about $3 billion per month were described as expected, not completed.
- The cited B200 on-demand rental-price index reached $6.01 per GPU-hour, indicating continued pressure on deployable GPU supply.
Amazon reportedly began an Early Rack Receipt process in August, receiving NVIDIA racks before their installation in AWS data centers is complete. According to the report, Amazon brought in approximately $5 billion of equipment during that month.
A significant portion of the received equipment reportedly cannot be installed on schedule because AWS data centers lack sufficient available power. The racks are said to be held in interim warehouses, and the report indicates that AWS supply-chain teams do not have a clear deployment path for some of the equipment.
The report described further GB300 rack purchases of approximately $3 billion per month as expected rather than completed. It did not disclose the quantity of racks held in storage or the magnitude of the unavailable electrical capacity in MW or GW.
For the data-center sector, the case illustrates the distinction between hardware delivery and deployable compute supply. GPUs and integrated racks cannot generate capacity until the supporting data-center infrastructure and grid-connected power are available. Early equipment receipt may therefore create a period in which capital is committed but the associated compute capacity cannot yet be monetized.
The supply constraint may also be contributing to elevated GPU rental pricing. Goldman Sachs data cited in the report put the on-demand B200 rental-price index at $6.01 per GPU-hour, up from $4.29 per GPU-hour in January 2026 and close to a reported historical high of $6.11 per GPU-hour in March. The report further projected that demand would still exceed capacity available for deployment in 2027; this is a forecast, not an operating result.
The reported conditions could strengthen the relative value of data-center projects with secured power supply and completed grid interconnection. They could also prolong the shortage of rentable GPU capacity and postpone the economic return on hardware procurement until power-constrained sites can support installation and operation.
The reported situation underscores that the binding constraint on AI infrastructure deployment can be timely, interconnected electrical capacity rather than GPU and rack supply. For operators and investors, it could increase the value of projects with secured grid interconnection while deferring revenue generation from already-procured compute hardware.
- TechNews 科技新報 / КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.