Amazon is committing billions of dollars to Nvidia hardware months before the sites meant to host it can deliver the electricity it needs, according to internal company messages describing finished server racks held in storage while construction and grid hookups trail behind. The company calls the arrangement ordinary procurement, but it has become the clearest illustration yet of a widening mismatch inside the artificial-intelligence boom: accelerators can be ordered and shipped on a schedule, while power cannot be summoned at the same speed.
- Key Point: Through an internal procedure known as Early Rack Receipt, AWS planned to receive about $5 billion worth of Nvidia GB300 systems during August alone, with roughly $3 billion forecast for each month that follows.
- Key Point: Several AWS facilities in the United States cannot yet install or energize the equipment, so the racks wait in intermediary warehouses until the destination data centers are completed.
- Key Point: Amazon’s trailing free cash flow slipped below zero in the second quarter, something that had not happened since 2023, while capital expenditure heads toward roughly $220 billion for the year.
Procurement that outruns construction
The purchasing cycle now begins long before a building exists to receive what is bought. Under Early Rack Receipt, AWS takes ownership of hardware early in the supply chain, and the numbers involved are enormous: around $5 billion of GB300 machines were slated for August, with approximately $3 billion per month expected to keep arriving for the foreseeable future. The company is locking in systems it describes as built for “the age of AI reasoning,” effectively paying ahead of deployment so that capacity is reserved by the time sites are ready.
That forward buying reverses a long-standing habit. In the pre-AI era, AWS did not take title to a rack until it had been fully assembled and handed over at a data center site. The new model ties up billions in cash up front, and internal discussion has flagged the absence of a firm schedule for installing everything that lands each month, a signal that the bottleneck has shifted from chip allocation to site readiness.
Idle iron sitting in warehouses
What results is a category of hardware that industry observers have started calling dark GPUs: accelerators that have been bought, built and paid for but produce nothing, because the building or the electricity is missing. Amazon is far from the only operator stockpiling equipment, yet its scale makes the pattern impossible to ignore. A company spokesperson characterized the stopover in third-party storage as “best business practice,” presenting warehouses as a buffer between a fast-moving supply chain and much slower construction.
The arrangement also carries a strategic edge for AWS. Holding racks early means competitors cannot claim the same manufacturing slots, and Amazon gets first call on Nvidia’s output at a moment when allocation remains tight across the industry. The trade-off is that capital is spent years before any of it generates revenue, and the hardware depreciates while it waits.
Cash leaving before revenue arrives
Amazon is outspending every other hyperscaler on AI infrastructure, and the pressure is visible in its finances. Free cash flow on a trailing basis went negative in the second quarter, the first such decline in roughly two years, and management expects the metric to stay depressed for a while as spending accelerates. Executives argue the outlay is warranted because AI represents a generational opportunity no large technology company can afford to skip, and they anticipate substantial returns once the buildout matures. Chief executive Andy Jassy has pointed to earlier infrastructure cycles at AWS as evidence the company knows how to absorb spending of this magnitude.
Even so, the sequencing matters. Servers bought today are billed today but only earn tomorrow, and every month of delay between delivery and energization widens the gap between what Amazon has spent and what its cloud division can sell.
The open question is no longer whether Amazon can obtain GB300 racks; it plainly can. What matters is whether substations, transmission queues and construction crews can keep up with a procurement operation running months ahead of them. Until they do, each shipment adds to an inventory that earns nothing, and the cost of arriving early compounds next to the cost of building slowly. The buildout narrative now rests on closing that gap, and the first credible sign will be a published installation timeline that matches the volume of hardware flowing in.