Explained: What are dark GPUs and how are they holding back the AI boom?

Power shortages are leaving billions of dollars' worth of AI chips unused at data centres. Here's what dark GPUs are, why they matter and how they could slow the AI boom

The artificial intelligence (AI) boom has run into an unexpected problem: some of the world's most powerful computer chips are sitting idle because data centres do not have enough electricity to run them, Business Insider reported. These unused chips, known as 'dark GPUs', highlight a growing gap between the pace of investment in AI infrastructure and the availability of power to support it.

 

Business Insider reported last week that Amazon was buying billions of dollars' worth of Nvidia AI systems even as some of its data centres lacked the electricity needed to operate them. Internal company emails reviewed by Business Insider showed that Amazon Web Services (AWS) expected to receive about $5 billion worth of Nvidia GB300 server racks in August this year, followed by around $3 billion worth every month.

 

Some of the equipment was being stored in warehouses until the required power infrastructure became available.

 

What are dark GPUs?

 

A graphics processing unit (GPU) is a specialised computer chip designed to handle several calculations at the same time. Unlike a central processing unit (CPU), which handles a wide range of general computing tasks, GPUs can process large amounts of data simultaneously.

 

This makes them particularly useful for AI applications, including training large language models, generating images and running AI-powered services.

 

Companies such as Amazon, Microsoft and Google buy thousands of these chips and install them in data centres, where they work together in server racks to provide computing capacity.

 

Dark GPUs are not a separate category of chips. The term refers to GPUs that have been purchased but remain unused because they cannot be switched on or fully utilised. This could be due to inadequate electricity supply, incomplete data centre infrastructure or problems with system design that prevent the hardware from being used efficiently.

 

Why are GPUs sitting idle?

 

The biggest reason is the shortage of electricity and the time required to build supporting infrastructure.

 

AI data centres consume large amounts of electricity because they house thousands of powerful chips that need to operate continuously. They also require cooling systems to prevent the equipment from overheating, along with networking infrastructure to connect the servers. However, building a data centre and securing the required electricity supply can take years. Grid connections, transmission infrastructure and approvals can delay projects even after companies have purchased the hardware.

 

As a result, chips may arrive before the facilities are ready to accommodate them.

 

Research organisation Epoch AI estimated in December 2025 that GPUs account for about 40 per cent of the electricity used by frontier AI data centres at peak operation. Cooling, networking and other infrastructure account for much of the remaining consumption.

 

What problems do dark GPUs create?

 

• Billions of dollars tied up in idle equipment

GPUs are expensive, and companies often spend billions of dollars securing supplies. When these chips remain unused, the investment does not immediately generate revenue.

 

In Amazon's case, its Early Rack Receipt programme allows it to take ownership of server racks earlier in the supply chain. While this helps the company secure scarce Nvidia hardware, it also means money is tied up before the equipment starts generating revenue.

 

• Higher costs and financial pressure

Companies must also bear costs associated with storing, maintaining and eventually installing the equipment. They may need additional financing while waiting for data centres to become operational.

 

• Delays in AI services

When computing infrastructure is not ready, companies cannot deploy the full capacity they have purchased. This can delay the expansion of cloud computing services and the availability of computing resources for AI developers and businesses.

 

• Pressure on the wider supply chain

The problem could extend beyond data centre operators. A Reuters report on October 5, citing Morgan Stanley, said US data centre developers faced an estimated net power shortfall of 34 per cent, or 32 gigawatts, through 2028. The brokerage warned that delays in deploying AI infrastructure could also affect demand for memory, optical and other chip components.

 

Is the problem getting bigger?

 

Recent developments at Amazon suggest that the infrastructure challenge is continuing even as technology companies increase their AI investments.

 

On Thursday, Amazon reiterated its plan to spend $220 billion on capital expenditure in 2026, with Chief Executive Officer (CEO) Andy Jassy saying the company expected to spend even more in 2027.

 

The company's results for the third quarter showed that its free cash flow had turned negative by $42.4 billion over the trailing 12 months. Jassy, however, said Amazon had strong visibility into customer demand for AI computing capacity through 2027, 2028 and 2029.

 

Amazon's third-quarter results showed its free cash flow turned negative, with an outflow of $42.4 billion over the trailing 12 months. The company attributed the decline primarily to higher spending on property and equipment, driven by AI investments.