Amazon is spending enormous sums on powerful Nvidia servers that run artificial intelligence. But instead of going to data centers, some of them are being sent to warehouses. Some centers do not have enough electricity to run them. Nevertheless, the company continues to buy them, paying billions of dollars more for them every month.
The information was reported by Business Insider, which obtained company emails from Amazon Web Services (AWS), Amazon's cloud division. According to them, Amazon was due to take delivery of Nvidia GB300 server racks worth roughly five billion dollars in August. These are complete systems packed with graphics chips used to train and run artificial intelligence models. The company expects to continue taking delivery of additional systems worth approximately three billion dollars every month for an extended period. Amazon did not disclose how long it would continue buying at this pace.
The catch is electricity. Some AWS data centers have enough space for the server racks, but cannot obtain as much power as they need to operate. The completed servers therefore wait in temporary warehouses until the centers are ready. According to the emails, one of the teams responsible for deliveries had no idea when the servers it was receiving would be installed and operational. The term dark GPUs has become established in the industry for this phenomenon. It refers to purchased chips that are not working because electricity or other infrastructure needed to operate them is unavailable.
Why Amazon Is Buying in Advance
The main reason is fierce competition for chips. Major technology companies want more high-performance processors from Nvidia than it can supply, and those that fail to secure them in time may miss out. Amazon has therefore introduced a new process called Early Rack Receipt. Previously, it only became the owner of the equipment once it arrived at the data center. Now it takes ownership while the equipment is still being delivered, pays for it, and records it as an asset, even if the servers have to wait in a warehouse. In practical terms, this means the chips belong to Amazon and are no longer available to its competitors.
The second reason is customer demand. The AWS division increased revenue by 37 percent to more than 42 billion dollars in the second quarter, while its backlog of orders awaiting fulfillment climbed to nearly half a trillion dollars. AWS Chief Financial Officer John Felton says that artificial intelligence servers pay for themselves in less than three years and that customer interest justifies such spending.
Amazon describes storing completed equipment near the location where it will be installed as standard practice. It also rejects the interpretation that the AWS division lacks an overall plan. According to the company, it has a system that tracks where the equipment is located and where it is supposed to go.
A Server Without Power Does Not Earn a Single Dollar
It has not yet been possible to determine how much equipment is sitting idle. What is certain is that a server without electricity generates no revenue. Every month spent waiting also reduces the time before new generations of chips push it to the sidelines. Meanwhile, Amazon has already paid for it.
And this is no small amount of money. Amazon plans to invest roughly 220 billion dollars this year. Its free cash flow—the cash left after the company pays for operations and investments—fell to nearly negative eight billion dollars in the twelve months ending in June. A year earlier, it was more than 18 billion dollars in positive territory. A company meeting in September showed how seriously the company views the situation. According to Business Insider, Felton joked about infrastructure chief Prasad Kalyanaraman, remarking that he had pushed the company into negative free cash flow and congratulating him ironically. Kalyanaraman himself acknowledged that the artificial intelligence business consumes a great deal of cash and expects cash flow to remain negative for several more years.
Amazon is not the only company being held back by electricity constraints. Morgan Stanley warns that U.S. data centers will face a shortfall of 32 gigawatts of electrical capacity by 2028. In September, AI hype critic Ed Zitron estimated that hardware worth 200 to 300 billion dollars could be sitting in warehouses or data centers without power. According to him, this is due to companies buying equipment faster than they can connect it. Amazon's case provides concrete evidence of what Zitron described.
Nervousness is also evident on the Polymarket prediction market. Traders now assign a ten percent chance that the artificial intelligence bubble will burst by the end of the year. Just a few days ago, the figure was six percent.
Sources: businessinsider.com, digitaltoday.co
