Nvidia chips pile up in Amazon warehouses as power shortages stall AI buildout
AWS bought Nvidia GB300 racks faster than its data centers can plug them in, leaving some hardware sitting dark in interim storage
Amazon Web Services has a problem most companies would envy: it bought so much Nvidia hardware that it has nowhere to plug it in.
Through what it calls the “Early Rack Receipt” program, AWS is set to take in roughly $5 billion worth of Nvidia GB300 server racks in August 2026. Some of those racks are reportedly landing in intermediary warehouses instead of humming away inside data centers, because the sites waiting for them lack the power to run them.
The industry has a name for this: “dark GPUs.” These are chips that have been paid for, shipped, and stored, but never switched on.
Buying first, plugging in later
The scale matters here. After that initial batch of roughly $5 billion, AWS expects approximately $3 billion in monthly deliveries of these AI server racks in the months that follow.
The strategy behind it appears deliberate. Internal emails indicate the procurement push was designed to lock down scarce AI hardware before rivals could get their hands on it.
The bottleneck is not floor space. Some AWS facilities reportedly have room for the racks but lack enough electrical capacity to power them, which forces the company to park hardware in interim storage until utility infrastructure catches up.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The bill is coming due
This hardware binge carries a hefty price tag, and it is showing up in Amazon’s numbers. The company’s capital expenditures for 2026 are projected at around $220 billion, nearly double the 2025 level.
AWS reported negative free cash flow of $7.6 billion for its most recent quarter, the first time it has slipped into the red on that measure since 2023.
The demand side, at least, appears to justify the ambition. AWS revenue grew 36.7% year over year in the second quarter.
AWS plans to deploy an additional 2 million Nvidia GPUs across its infrastructure between 2027 and 2028, including Blackwell Ultra and Rubin models. That comes on top of prior commitments covering more than 1 million GPUs.
What this means for the AI buildout
The most immediate question for anyone watching the AI infrastructure trade is how much of this hardware sits unused, and for how long. Amazon has not released details on whether any of the stockpiled racks could end up on the secondary market, or on any timeline for that.
For Nvidia, the picture is straightforwardly positive in the near term. A customer committing to roughly $3 billion in monthly rack deliveries, plus millions more GPUs in 2027 and 2028, is about as strong a demand signal as a chipmaker can ask for.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.

