The "Dark GPUs" phenomenon reveals the limits of massive AI infrastructure expansion
The global AI boom is facing a hidden crisis represented by the severe shortage of electrical power needed to operate advanced hardware. In this context, reports and leaks from internal correspondence have revealed that Amazon is purchasing Nvidia AI server GPUs worth billions of dollars monthly, at a time when its sites and data centers do not have enough power to operate these processors upon arrival.
Internal messages point to a process dubbed "Early Rack Receipt," where Amazon Web Services (AWS) expected to receive Nvidia GB300 server racks worth approximately $5 billion in a single month, followed by about $3 billion more each month. Given that some of the company's facilities in the United States are not ready, this expensive hardware is being stored in intermediate warehouses until the final data centers are completed.
What does the concept of "Dark GPUs" represent?
This trend has led to the emergence of a phenomenon known as "Dark GPUs," where major cloud computing companies hoard as much hardware as possible, while a large portion of it remains in storage awaiting the availability of power infrastructure. This situation has cast a financial shadow over the company, as Amazon's free cash flow turned negative during the second quarter for the first time since 2023, with capital expenditures expected to reach about $220 billion this year.
Why is this trend changing traditional procurement policies?
Before the massive AI wave, AWS followed a policy of not owning servers until they were completed and actually delivered to ready data centers. Today, the company is willing to pay billions of dollars in advance to ensure it secures and reserves its share of the global AI supply chain, bypassing standard operational practices to ensure it does not fall behind.
What does this mean for the sector and users?
Amazon officials defend these practices as a sound business strategy to secure a historic, unmissable opportunity in the AI sector, noting that the company has overcome similar challenges in previous growth phases. However, the real challenge lies in the ability of the energy and infrastructure sector to keep pace with the rate of chip manufacturing and procurement, which places practical constraints on the speed of deploying actual AI capabilities in the market.