Amazon Locks Billions in Nvidia GPUs Idle Due to Power Shortages
Amazon has invested billions to secure Nvidia GPUs, but due to insufficient power, related warehouses remain idle, facing bottlenecks in AI computing power expansion. The key constraint in computing power competition is extending to electricity and infrastructure; securing chips does not equate to forming usable capacity.
According to Business Insider, based on Amazon's internal documents and employee meeting recordings, AWS is temporarily storing Nvidia GB300 racks in transit warehouses because the target data center "has space for more racks but lacks sufficient available power to turn them on." Internal emails indicate that the company expects to receive about $5 billion worth of GB300 racks in August alone, continuing to receive about $3 billion each month thereafter, maintaining this pace for the foreseeable future.
AWS previously adopted an "Early Rack Receipt" model to secure hardware supply in advance, even if some facilities are not yet powered. AWS CFO John Felton noted in an internal meeting that this investment has turned the company's free cash flow negative. Amazon's capital expenditure is expected to be around $220 billion this year, marking the first negative free cash flow since 2023.
The industry refers to these purchased but non-operational GPUs due to power shortages as "dark GPUs." Idle hardware does not contribute computing power, and the rental price for GPUs in the spot market has risen instead of falling; Cleanview Research data shows that Amazon has 44 GW of data center capacity under construction in 16 states in the U.S., more than three times its existing 13.9 GW operational scale. Power, rather than chips, has become the scarcer resource.
Funds and resources are flowing into grid expansion, substation construction, and distribution upgrades, with hardware arrival speeds exceeding the pace of power infrastructure implementation; beneficiaries are data center operators and power infrastructure providers that have secured grid quotas and completed power connections, while cloud vendors relying on idle hardware depreciation and unable to generate short-term revenue face pressure on capital expenditures.
Source: Public Information
ABAB AI Insight
Amazon AWS has long focused on large-scale data centers and cloud infrastructure, recently shifting its capital expenditure focus to AI computing power; the company locks in high-end racks like Nvidia GB300 through an early receipt model to prioritize quota acquisition during supply shortages, but the expansion of data center power requires grid approvals, substation construction, and high/low voltage distribution work, which takes longer than server delivery.
In terms of capital flow, the company has front-loaded hundreds of billions in capital expenditures for hardware procurement, even if some racks need to be temporarily stored while waiting for power; the motivation is to avoid losing quotas during chip supply shortages and to bet that future demand will cover depreciation and waiting costs, while tracking grid quotas and milestone risks through an internal "Secured Power at Risk" dashboard.
This is similar to the early cloud computing expansion model where servers arrive before power and cooling in the data center, and is akin to cases in semiconductor capacity expansion where wafer fab equipment arrives faster than utility support; currently, the transition from chip scarcity to "powered computing power" scarcity is underway, and hardware inventory no longer automatically translates into usable capacity.
Essentially, this represents a restructuring of the supply chain: shifting the constraints of computing power supply from chip procurement to electricity and grid infrastructure, decoupling "locking in chips" from "forming usable capacity"; the reason is that AI rack power density is much higher than traditional servers, and the grid expansion approval and construction cycle is longer than hardware delivery, leading to capital being locked in early while output is delayed by months or even years.
ABAB News · Cognitive Laws
Locking in chips does not equate to forming usable capacity
Electricity is the new bottleneck for computing power
From hardware scarcity to powered scarcity, is a misalignment in expansion pace