Amazon is accelerating the early locking of AI hardware, even though power shortages in data centers mean that some GPU systems will still have to wait in warehouses.
According to an internal email from August viewed by Business Insider, under a new “Early Rack Receipt” process, AWS is expected to receive approximately $5 billion worth of NVIDIA GB300 server racks before the final deployment in its US facilities. The email also states that the company anticipates receiving another $3 billion worth of GB300 racks each month for the “foreseeable future.”
People familiar with this arrangement stated that although some sites of AWS still have space to accommodate more racks, there is a shortage of available power, which prevents them from being put into use. As a result, the completed equipment will be temporarily stored in a transit warehouse until the facilities are ready to accommodate it.
One of the emails stated that at least one AWS team did not have a “clear, predictable path” to deliver the newly arrived racks into the data center at that time.
Dark GPU
Insufficient additional power capacity is one of the multiple bottlenecks that are slowing down the construction of some AI data centers. This has given rise to what is known as the "dark GPU" phenomenon, where AI chips and servers have been purchased but cannot be put into use due to power supply issues or other challenges. ("Dark GPU" can also refer to system design flaws that prevent these devices from being fully utilized even when powered on.)
This is also one of the reasons why large data center projects face financial pressures. Enterprises must invest funds in the early stages to secure the supply of NVIDIA GPU and other high-value AI components. However, if some data centers are not yet ready to operate, these early investments will tie up a large amount of cash. Debt financing can fill the gap, but recently, the market's willingness to lend against these projects as collateral has weakened.
Trade-offs
Amazon's early rack projects were part of this cloud computing giant's broader AI construction plan, which is now putting new pressure on its cash flow, as Business Insider reported exclusively this week. Amazon expects its capital expenditures this year to be around $220 billion, exceeding any of its large tech peers. Its free cash flow turned negative for the first time since 2023 in the previous quarter, and this situation may continue for some time to come.
For AWS, the trade-off is as follows: to secure the scarce GPU racks before competitors can, even though some data centers are not yet ready to put them into use. However, taking control earlier also means that billions of dollars will be invested in hardware before these devices start generating cloud revenue.
Previously, according to internal documents, AWS only took ownership of the completed racks once they were delivered to the data center. Under the new approach, the company can make payments for certain racks and record them in the accounts at an earlier stage of the supply chain.
An Amazon spokesperson stated that the company typically keeps the completed hardware near the deployment sites so that it can be installed immediately when production capacity becomes available, calling this "best business practice." The spokesperson also added that Amazon has had a well-established infrastructure asset tracking system for a long time and denied that the company generally lacks pathways to transport racks into data centers.
Amazon CEO Andy Jassy previously stated that the company "sees very clearly" the high returns on investment in AI, calling this technology a "once-in-a-lifetime opportunity."
