Amazon is securing AI hardware earlier than ever before, but a shortage of power at data centers is leaving some GPU equipment stranded in warehouses.
According to reports, internal emails from August show that under a new early rack receipt process, AWS plans to accept roughly $5 billion worth of NVIDIA (NVDA) server racks at U.S. sites without immediately moving them into final deployment. One email noted that for the "foreseeable future," the company expects an additional $3 billion worth of racks to arrive each month.
People familiar with the matter said some AWS facilities still have space to place racks but lack available power to power them on, so finished equipment is temporarily stored in transit warehouses until power conditions at the facilities are ready. The emails show that at the time, at least one AWS team had no clear timeline for moving the incoming racks into data centers.
Rising power shortages and "dark GPUs"
Insufficient new power supply for "dark GPUs" is one of the bottlenecks constraining the pace of some AI data center construction. This has given rise to the "dark GPU" phenomenon: companies have already purchased AI chips and servers, but due to power supply or other problems, the equipment can only sit idle. Dark GPUs can also refer to another situation: hardware is powered on, but due to system design flaws, computing power cannot be fully utilized.
This is also part of the reason financial pressure on large data center projects continues to rise. Companies need to invest huge amounts of capital early to lock in NVIDIA (NVDA) GPUs and other high-value AI components; if the facility is not ready for deployment, this upfront investment ties up significant cash flow. Companies can rely on debt financing to fill the funding gap, but recently market willingness to lend to such projects has declined.
Trade-offs
Amazon's early rack receipt plan is one part of the cloud giant's large-scale AI infrastructure buildout. The report noted that AI expansion is putting new pressure on Amazon's cash flow. Amazon expects capital expenditure of about $220 billion this year, the highest among leading technology companies. Last quarter, the company's free cash flow turned negative for the first time since 2023 and could remain so for some time.
For AWS, this is a trade-off: secure scarce GPU racks before competitors, even if some facilities cannot yet be powered on. But recognizing receipt early means billions of dollars of equipment will tie up company funds before generating cloud revenue. Internal documents show that previously AWS only completed asset recognition when racks were delivered to data centers. Under the new plan, Amazon.com (AMZN) can pay and recognize some rack assets at an earlier stage in the supply chain.
An Amazon spokesperson responded that the company routinely stores finished hardware near deployment sites and installs it once resource capacity is available, calling this a "mature business practice." The spokesperson added that Amazon has a mature infrastructure asset tracking system and denied that it broadly lacks a path for moving racks into facilities.
Amazon CEO Andy Jassy previously said the company has "clear expectations" for strong returns on AI investment, calling AI a once-in-a-lifetime opportunity. In Amazon's annual shareholder letter, he wrote: "We experienced AWS's first high-speed growth cycle, and the results were quite good. We expect this new cycle to be the same."
Comments