Amazon Web Services (AWS) is securing Nvidia AI servers on a large scale, but is unable to run the equipment due to power shortages at some data centres. As competition intensifies to invest in AI infrastructure, power supply has emerged as a key obstacle to expanding data centres.

On Oct. 20, Business Insider reported that AWS plans to take delivery of about $5 billion worth of Nvidia GB300 server racks through a new Early Rack Receipt process before they are finally deployed at U.S. facilities. An internal email from August obtained by the outlet also included an outlook that AWS would take delivery of an additional about $3 billion worth of GB300 racks each month for a considerable period ahead.

The problem is power supply. Some AWS data centres have secured space to install the server racks, but are said to be unable to receive enough electricity to run the equipment. Completed server racks are being stored in intermediate warehouses until the data centres finish preparations to start operating. The internal email also flagged that some AWS organisations had not set a clear schedule to place equipment they plan to secure at data centres. This phenomenon of purchased AI semiconductors being unable to operate due to factors such as power shortages is called a Dark GPU.

AWS previously took ownership when completed server racks arrived at data centres. Under the new process, it can take delivery earlier in the supply chain and record the equipment as assets.

It means AWS can secure high-performance AI semiconductors ahead of competitors, but large sums of money could be tied up before actual cloud revenue is generated.

Amazon expects capital expenditure to reach about $220 billion this year. In the previous quarter, free cash flow turned negative for the first time since 2023.

Amazon explained that storing completed equipment in advance near where it is slated to be deployed is a common operating practice. It also said it has a system to track the location of equipment and deployment plans, disputing interpretations that the company lacks deployment plans across the organisation.

Amazon Chief Executive Andy Jassy (앤디 재시) previously expressed confidence that it can earn high returns from AI investment.

Some have pointed out that for large-scale AI investment to translate into actual profit, it is important not only to secure semiconductors but also to align power supply for data centres with the timing of operations.