There is a kind of problem only very rich companies get to have. You order billions of dollars of the most sought-after hardware on earth, it arrives on schedule, and then it sits in a warehouse because nobody can plug it in.
That is the picture painted in Business Insider reporting published this week and summarized by Benzinga on Yahoo Finance on October 21. According to that reporting, Amazon has been buying Nvidia AI server racks faster than some of its data center sites can power them. Amazon disputes the framing. But the story lands on a question every investor in the AI trade should be asking right now. When the spending outruns the electrical grid, who pays for the gap?
What the reporting actually says
Start with the facts, because they matter more than the headline. Business Insider cited internal emails showing Amazon expected about $5 billion of Nvidia AI server systems to arrive in August, then roughly $3 billion more each month after that. One email said the monthly deliveries would continue “for the foreseeable future,” without giving an end date.
The company also runs what the report calls an Early Rack Receipt program. Under it, Amazon takes payment and ownership of equipment before it reaches a data center. Previously, it waited until the hardware showed up on site. Some sites, the report says, have the physical room for the racks but not enough power to run them, and one supply-chain team reportedly did not know when the servers it was receiving would be installed and ready for use.
Amazon’s response was that staging hardware near deployment sites is “best business practice,” and it rejected the suggestion that AWS lacks a broader deployment plan. That is a fair point. Large builders do stage equipment, and a rack in a warehouse for a few weeks is not a scandal. The question is how many racks, for how long, and what it costs to hold them.
One outside estimate puts a number on the worry. AI skeptic Ed Zitron estimated in September that $200 billion to $300 billion of AI chips across the industry could be sitting in warehouses or in data centers without power. That is an estimate from a critic, not a measured figure, and it should be treated that way. Still, it shows how the debate has shifted. A year or two ago the argument was whether anyone would buy the chips. Now it is whether anyone can turn them on.
Why idle hardware is a cash flow story
Here is where the numbers get uncomfortable. Amazon plans about $220 billion in company-wide capital spending this year. Its free cash flow was negative $7.6 billion for the 12 months ended June 30, compared with positive $18.2 billion a year earlier. That swing is the arithmetic of a company spending cash on equipment faster than its operations generate it.
The most striking detail is a quote. At an internal meeting in September, AWS CFO John Felton reportedly joked to infrastructure chief Prasad Kalyanaraman, “You’ve taken us free cash flow negative as a company, so congratulations.” Kalyanaraman described Amazon’s AI business as “hugely free cash flow negative” and, per the report, expected that to last “the next couple of years or so.”
Read that again. The executive running the buildout is saying, in the company’s own meeting, that this business will consume more cash than it produces for roughly another two years. That is not a secret to anyone who follows the numbers, but it is rare to hear it said so plainly.
Why does an unpowered rack matter on top of that? Because hardware earns nothing until it runs. A server that has been bought, paid for and owned, but cannot be switched on, is capital that is tied up with no revenue attached. Every month of delay stretches the time before that spending turns into customer billings. Chips also age. The newest generation arrives on a steady cadence, so a rack that waits is a rack losing some of its edge before its first day of work. We can debate how much, but the direction is not in doubt.
The bull case is not silly
It would be lazy to stop there, because the demand side of this story is real. AWS revenue rose 37% to $42.2 billion in the second quarter, and its backlog reached $496 billion, per the reporting. A backlog that size means customers have signed up to spend. Felton also told Axios that AWS is “seeing paybacks of under three years” on its AI servers, a figure that covers servers and networking gear rather than the data centers and power behind them.
If you believe those claims, the logic is simple. Customers are waiting, the hardware pays for itself, and the only thing standing in the way is concrete, copper and transformers. Bottlenecks like that tend to be temporary. Power gets built eventually, and a company with a backlog can afford some patience.
But notice what that argument depends on. It depends on the payback period holding up, which is a claim by the company selling the service. It depends on the backlog converting into actual revenue on schedule. And it depends on power arriving in time. Morgan Stanley has warned of a 32-gigawatt power shortfall for U.S. data centers through 2028. If that warning is even close to right, the constraint is not a quarter-long hiccup. It is a multi-year feature of the business.
What this means for the wider AI trade
Amazon is one company, but it is a useful window onto everyone else. Nvidia sells the racks, so its orders can look healthy while customers struggle to deploy what they already own. That is a timing issue, not necessarily a demand issue, but timing is exactly what markets price. The suppliers of grid equipment, the utilities and the power producers sit on the other side of the same trade. Their problem is the mirror image: too many customers, not enough capacity.
Investors should also keep the sentiment layer in perspective. A Polymarket contract on whether an AI bubble bursts by the end of the year moved up to 10% from 6% in a matter of days, according to the Benzinga write-up. Odds on a version running to June 30, 2027 sit near 22%, though only about $35,000 has traded on that contract. A “Yes” requires three specific industry stress events within 90 days, so unpowered chips alone would not trigger it. It is a mood gauge with thin volume, not a forecast. The more useful signals are the ones in company filings.
So what should you actually watch? First, capital spending guidance against free cash flow, because the gap between them is the real cost of the buildout. Second, any change in how companies describe the lag between receiving hardware and putting it to work. Third, whether cloud backlogs keep converting into reported revenue at the pace executives have described. And fourth, whether the power constraint shows up as a line in earnings calls rather than a leak in a news story.
None of this means the AI buildout is a mistake. Plenty of great businesses spent heavily before the payoff showed up. But the market has spent two years pricing the demand for chips. The next chapter is about the plumbing, and plumbing does not care about narratives. It either delivers power on time or it does not.
The companies that have the cash, the sites and the grid connections will look very different from those that only have the orders. That difference will not show up in a press release. It will show up in the cash flow statement, one quarter at a time.
This article is written for educational and informational purposes only and does not constitute financial or legal advice. The views and analytical frameworks presented draw on publicly available information and reported commentary from industry participants. Readers are encouraged to consult primary sources and form their own informed views on these complex topics.
