AWS Reportedly Stockpiles $5 Billion in Nvidia GB300 Server Racks in Warehouses While Waiting for Data Center Power
Nvidia is accused of worsening the gaming GPU shortage by prioritising AI hardware for Big Tech. The claim is that Nvidia is sending huge amounts of GPUs to companies building AI data centres, even when some of that hardware cannot be used yet because the data centres don’t have enough power. Companies like Amazon receive billions in Nvidia AI hardware before data centres have enough power. About $5 billion in GB300 systems were delivered to AWS facilities, with another $3 billion monthly expected, leaving some unused in warehouses.
Context
Billions of dollars in frontier AI server racks sitting unpowered in intermediary warehouses show that utility electricity capacity, rather than chip manufacturing alone, is now the binding constraint on hyperscale AI deployment.
Hyperscale cloud providers are racing to secure next generation Nvidia Blackwell Ultra GB300 server racks even as US utility grid interconnections face multi year delays. Historically, Amazon Web Services paid for server racks only once they were delivered and installed inside active data centers. In August 2026, internal AWS emails reviewed by Business Insider revealed a policy shift called Early Rack Receipt to lock in Nvidia hardware allocations ahead of competitors.
Deep dive
Pirat_Nation highlighted reporting that Amazon Web Services has taken delivery of billions of dollars in Nvidia AI server hardware before its data centers have sufficient electricity to turn the systems on, fueling criticism from gaming hardware analysts. The underlying facts come from an October 2026 Business Insider investigation citing internal Amazon emails from August 2026. According to those emails, AWS created an Early Rack Receipt program under which it expected to receive roughly $5 billion worth of Nvidia GB300 server racks in August alone, followed by approximately $3 billion per month for the foreseeable future.
Because several US AWS facilities had physical floor space but lacked utility power capacity to energize the racks, finished GB300 systems were staged in intermediary warehouses, a phenomenon Business Insider described as dark GPUs. One internal email noted that at least one AWS team lacked a clear forecasted path for moving the incoming racks into powered data centers. An Amazon spokesperson told Business Insider that staging hardware early is a routine best business practice so racks can be deployed immediately once power comes online.
Separately, tech commentator Ed Zitron and gaming publications linked hyperscale stockpiling to severe retail shortages and price spikes for consumer GeForce RTX 5090 graphics cards, arguing that Nvidia's allocation of TSMC wafer capacity to data center orders starves the gaming market. That market manipulation claim represents analyst commentary rather than proven misconduct, and Nvidia has not publicly responded to the accusation.
What to expect next
Cloud investors will watch whether unpowered GPU inventory continues to weigh on hyperscaler free cash flow in upcoming quarterly disclosures, and how quickly utility power agreements and on site power generation can bring warehoused GB300 clusters online. PC gamers and retail buyers will also monitor whether data center wafer allocations continue to constrain consumer GeForce RTX 5090 supply and pricing through the holiday quarter.