AWS Reportedly Holds Some Delivered NVIDIA AI Systems Pending Data-Center Power Availability
DataNews Editorial Team
AWS has reportedly accelerated procurement of NVIDIA GB300 AI server racks, while some delivered systems are temporarily held in warehouses or interim facilities because certain data-center sites lack sufficient available power. The reported constraint is not rack space but the electrical capacity required to operate high-performance AI infrastructure.
- AWS reportedly accepted and paid for some NVIDIA AI server racks before final-site installation to secure equipment supply.
- Some delivered systems are reportedly awaiting sufficient data-center power rather than physical installation space.
- AWS had previously expected approximately $5 billion in NVIDIA GB300 racks, according to S1.
- S1 reports planned further equipment purchases of about $3 billion per month.
- Amazon reportedly plans about $220 billion in AI and cloud infrastructure CAPEX in the current year, although the allocation is not disclosed.
According to S1, AWS is using early acceptance and payment for NVIDIA server racks to secure scarce equipment before it can be installed at its final operating location. Some of the delivered AI systems have not been commissioned immediately, with available installation space at certain data centers reportedly outpacing the power available for high-performance AI loads.
S1 describes this purchased but temporarily non-operational equipment as “Dark GPU.” The source does not disclose the number of affected GPUs or racks, the number or location of impacted data centers, the available or missing electrical capacity, or the expected timeframe for resolving the constraints.
AWS had previously expected to receive NVIDIA GB300 server racks valued at about $5 billion, according to S1. This was an expected receipt value rather than confirmation that the full volume has been delivered. The source also reports plans for further purchases of related equipment at approximately $3 billion per month; these are forward plans and may change.
Separately, S1 says Amazon plans approximately $220 billion of capital expenditure for AI and cloud infrastructure in the current year. The source does not specify what portion of that planned spend relates to AWS, server equipment or energy infrastructure.
For data-center operators and hyperscalers, the reported situation suggests that the commercial value of AI server deployments may increasingly depend on synchronizing hardware procurement with grid interconnection, power-delivery infrastructure and campus readiness. If power availability and site construction lag GPU procurement, more capital could remain temporarily committed to equipment that cannot yet be placed into service.
The report illustrates that GPU delivery alone does not translate into deployable AI compute capacity. Grid interconnection, power-delivery infrastructure and site readiness can determine both deployment timing and the period for which capital remains tied up before cloud infrastructure can generate revenue.
- — / October 24, 2026 · ChineseSource →
This article was prepared by the DataNews editorial team based on the sources listed above.