Amazon has begun taking ownership of Nvidia GB300 server racks before some AWS data centers have enough electricity to operate them, leaving billions of dollars’ worth of hardware idle as the company accelerates its artificial intelligence capabilities. The episode highlights how electricity access has become the gating factor in AI infrastructure: owning the hardware no longer means being able to run it.

The racks in question are GB300 systems, Nvidia’s Blackwell Ultra-generation rack-scale platforms built for large AI training and inference workloads. According to Business Insider, internal emails from August projected initial rack receipts of about $5 billion, followed by approximately $3 billion per month. The figures are forecasts, not confirmed purchases.

Under AWS’s Early Rack Receipt policy, ownership transfers before the equipment is physically received. As a result, some racks are being stored in warehouses. An Amazon representative described the practice as a “best business practice,” while the company denied allegations that it lacks a deployment plan.

Racks arrive before the power

AWS previously took ownership of racks when they reached their intended locations. It is now paying for some equipment before installation or electricity is available.

The resulting “unpowered” systems, sometimes called “dark GPUs,” generate no revenue while they wait to be activated. They also lose value over time, adding pressure to Amazon’s cash flow.

Capital spending adds to the pressure

Business Insider reported that Amazon is projected to spend $220 billion on capital investment in 2026. According to Amazon’s filing with the SEC, the company recorded $96.3 billion in cash investment and $71.4 billion in operating cash flow during the first half of the year. The capital spending included investments in fulfillment infrastructure.

Amazon separately spent $13.7 billion on OpenAI and $10 billion on Anthropic in the second quarter. Those investments are not classified as capital expenditure.

At a September staff meeting, AWS infrastructure executive Prasad Kalyanaraman described the AI business as “hugely free cash flow negative,” according to Business Insider. He warned that the strain could persist for several years.

Chief Executive Officer Andy Jassy, however, has called AI a “once in a lifetime opportunity.” AWS earnings increased 37% in the second quarter, while its contracted backlog rose to $496 billion.

Amazon expands its search for power

Cleanview estimates that Amazon is developing 44 gigawatts (GW) of data-center capacity in the United States, including 27 GW across 13 newer markets, according to its analysis. The analysis also identifies 17.5 GW of planned gas generation and 25.6 GW of contracted carbon-free electricity.

These figures represent separate commitments and should not be added together as if they were all operational capacity.

The constraints extend beyond Amazon. Bain has cited shortages of electricity, equipment and skilled workers, along with delays in obtaining permits. The International Energy Agency has also highlighted increasing stress on the power grid in its report. Local opposition is making further expansion more difficult. Amazon has pledged $1 billion to support communities hosting its data centers, according to Cryptopolitan.

Operational capacity may affect AI pricing

The Boston Consulting Group expects AI-compute pricing to become more usage-based and potentially dynamic, according to its analysis. Power shortages could affect where AI facilities are built, which customers receive computing capacity and what users pay worldwide.

For Amazon, purchasing GPUs is only part of the challenge. The larger test is putting those systems into operation and converting the investment into revenue. Signs to monitor include how quickly warehoused racks are energized, how the 44 GW pipeline progresses toward operation, and how much of the $496 billion backlog converts into delivered capacity.