Large numbers of Amazon NVIDIA GB300 servers temporarily idle due to insufficient power supply

According to reports, Amazon's cloud computing division AWS is purchasing Nvidia's latest AI server systems at an unprecedented pace. According to internal planning, AWS at one point expected to receive Nvidia GB300 server racks worth about $5 billion, and plans to continue purchasing related equipment at a scale of about $3 billion per month in the future. However, some of the equipment was not put into use immediately after delivery, but was temporarily stored in warehouses or transit facilities.

The reason for this phenomenon is not insufficient demand, but that although some data centers already have installation space, they lack sufficient power supply capacity to drive these high-power AI systems. The industry calls this type of equipment that has been purchased but cannot be operated temporarily "Dark GPU." This concept can refer both to hardware left idle due to power or infrastructure constraints and to AI computing resources that have been deployed but are not fully utilized.

Over the past few years, the main challenge facing the AI industry has been tight supply of advanced chips. Major tech companies have competed for Nvidia GPU quotas, and chips are often in short supply. But as companies lock in supply chain resources in advance, the new limiting factor is shifting to power grids and energy infrastructure. For large data center operators, obtaining a stable and sufficient power supply is becoming as important as obtaining chips.

To ensure it does not fall behind in the future chip competition, Amazon has adjusted its procurement model. Under the previous process, the company usually formally accepted and paid for server racks only after they arrived at a data center ready for operation. Under the new "early rack acceptance" mechanism, Amazon completes the acceptance and payment process earlier, even if the equipment has not yet been installed at its final operating location.

This approach helps lock in scarce hardware resources in advance, but it also means companies need to invest large amounts of capital upfront, and this equipment cannot generate revenue before it officially goes online and provides cloud computing services to customers. As a result, the contradiction between huge capital expenditure and delayed returns is intensifying.

Amazon plans to invest about $220 billion in capital expenditure this year for AI and cloud computing infrastructure, a scale exceeding that of many tech peers. At the same time, the company's free cash flow is already under obvious pressure and has turned negative for the first time in recent years. For AWS, taking possession of key hardware resources in advance is regarded as a strategic choice, even if some equipment cannot operate temporarily.

Analysts point out that the focus of the current AI industry's development is changing. Two years ago, the industry focused on who could obtain more Nvidia GPUs; now more and more companies are finding that the truly scarce resource may no longer be the chips themselves, but the energy and power networks needed to support the operation of these chips.

As AI models continue to expand in scale and computing power demand continues to grow, power supply, transmission capacity, and data center construction cycles are gradually becoming key factors affecting industry development. The large number of temporarily idle GPUs also reflects a reality from another angle: the AI race is no longer just a contest over chips, but also a contest over energy and infrastructure.