AWS Is Paying Billions for Nvidia GB300 Racks It Cannot Switch On Yet

The AI infrastructure story just flipped. For two years the bottleneck was getting the chips: allocation, lead times, Jensen Huang's leather jacket deciding who ate first. Now, according to internal AWS emails reported by Business Insider this week, the bottleneck is plugging them in. AWS expected to receive roughly $5 billion worth of Nvidia GB300 server racks at US facilities before final deployment, with about $3 billion more arriving per month "for the foreseeable future," under a new process called "Early Rack Receipt." The company is buying finished AI hardware faster than it can turn the hardware on.

Read that again. These are not forecasts of demand. These are forecasts of delivery: racks physically arriving that cannot yet be installed or powered. Some AWS sites reportedly had floor space for more racks but not enough electricity to run them, so finished equipment sat staged at intermediary warehouses. "Dark GPUs": depreciating assets earning zero revenue, waiting on power.

The GB300 racks are arriving before the power

The mechanics matter. AWS used to take ownership of a finished rack when it was delivered to the data center. Under Early Rack Receipt, it pays for racks before installation or electricity is available. That is a quiet but significant shift of risk: inventory and delay risk moves from the supplier onto Amazon's balance sheet. When the power shows up late, Amazon eats the carrying cost, not Nvidia.

The emails describe a planning apparatus straining at the seams. At least one AWS team reportedly lacked "a clear forecasted path" for getting incoming racks into data centers. Power, permitting, and grid upgrades are now as rate-limiting as chip supply ever was. Amazon's spokesperson pushed back, calling staging finished hardware near deployment sites "best business practice" and disputing that the company lacked a deployment path. Both things can be true: staging near the site is sensible, and still an admission that the last mile of the AI buildout is the electrical grid.

Why power became the bottleneck

This has been coming. Data center demand doubled in six months earlier this year, and the industry's own analysis named permits as the new bottleneck. DayOne's billion-dollar bet on Johor was explicitly a power play: go where the megawatts are. The GB300 racks are the most power-hungry systems Nvidia has ever shipped, and they are landing in a country where grid interconnection queues stretch for years.

It is a weird inversion of the last decade of cloud economics. The industry spent years optimizing for compute efficiency, assuming power was someone else's problem. Now power is the problem, and it is the least flexible part of the stack. You can redesign a chip in eighteen months. You cannot redesign a substation on that timeline. Every hyperscaler is learning this simultaneously, which is why the land grab has moved from silicon to sites with grid capacity already secured.

The money behind it

The scale of the spend is worth stating plainly. Amazon expected around $220 billion in capital expenditure in 2026. Trailing-twelve-month free cash flow was negative $7.6 billion at midyear, versus positive $18.2 billion a year earlier. AWS sales grew 37 percent to $42.2 billion in Q2, and the company sits on roughly $496 billion in unrecognized commitments, mostly AWS. This is a company spending like the demand is infinite while the cash flow says the spending has a cost.

None of this means the AI buildout is stopping. It means the narrative is maturing from "who can buy the most GPUs" to "who can energize the most GPUs," which is a much harder game and a much slower one. The winners of the next phase are not the companies with the biggest purchase orders. They are the companies with the substations, the permits, and the grid interconnections already in hand. Chips are fungible. Megawatts are not.

For everyone watching cloud prices, there is a direct line here. The UK just forced cloud switching to be free for 180 days, but nobody is switching, because moving is operationally painful. Now add this: your cloud provider is paying billions for hardware it cannot switch on, and that carrying cost lives somewhere in the price of your instances. The GB300 racks will get their power eventually. The question is what your bill looks like when they do.