Amazon is warehousing $5B of Nvidia racks it cannot power — and paying to develop 960 MW of reactors that do not exist yet

Business Insider reported, citing internal August emails, that Amazon set up an "Early Rack Receipt" program to take delivery of roughly $5 billion of Nvidia GB300 server racks at US facilities ahead of deployment, with about $3 billion more expected each month. The finished equipment is being staged in intermediate warehouses because several AWS sites lack the utility capacity to energise it. An Amazon spokesperson called the staging a routine "best business practice" to allow rapid installation once power arrives and disputed indications that it lacked a clear path to bringing the sites live. Market commentators noted that racks delivered into staging count as revenue for Nvidia at delivery.

On 19 October American Electric Power, X-energy and Amazon agreed to evaluate a 960 MW nuclear plant at the retired Breed coal plant in Sullivan County, Indiana: twelve 80 MW Xe-100 reactors in three groups of four. Amazon will fund development in tranches and be the first buyer of the power, covering its own project costs rather than passing them to other customers; AEP would operate the plant and the output would go onto the public grid. No construction date, final investment decision or budget has been disclosed. X-energy has not yet built a reactor; its first project, at Dow's Seadrift site in Texas, has not started construction.

Why it matters

These are the same constraint at two timescales. In the short run, the bottleneck in AI infrastructure is no longer chips: Amazon has the hardware and is storing it, because the grid connection is not there. Two weeks ago Oracle was weighing trucked gas to run a campus whose pipeline was blocked; here the workaround is a warehouse. In the long run, the same company is paying to develop generating capacity a decade out, on a design that has not yet been built anywhere, because it cannot otherwise count on the power.

Read the Indiana deal for what it is. "Amazon to fund 12 small nuclear reactors" was the headline; the agreement is a development and feasibility commitment, with Amazon paying in tranches as milestones are met. It is meaningful — a hyperscaler underwriting the early, riskiest phase of new nuclear is how first-of-a-kind projects get financed — but it is not twelve reactors, and the tracker records nothing until there is a permit application or a construction start. The Calvert Cliffs agreement earlier this month, for an uprate at an operating plant, will deliver power years sooner.

The Nvidia detail matters for anyone reading AI demand from chip revenue. Revenue booked at delivery into a warehouse is demand that has been paid for but not yet turned on. That is not the same as demand that is running, and the gap between the two is a measure of how far power lags compute.

What to watch

How much of Amazon's staged hardware is energised in the next two quarters, AEP's filings for the Indiana site and whether an NRC application follows, and progress on X-energy's first unit at Seadrift.