Amazon raises GPU rents 15% while shifting $8 billion of chips off its books

The reserved hour, repriced a third time

Book eight Nvidia H100s on AWS for a training run and the bill is $41.528 an hour in major US regions ¹. On 2 October Amazon said that rate, and the rest of its EC2 Capacity Blocks for machine learning, goes up about 15% next week, on chips from the older A100 to the B300 ².

At the announced rise, the eight-H100 p5.48xlarge hour steps up to about $47.76 on 7 October ³. Capacity Blocks book up to eight weeks ahead ³: reserve before then, or budget the raise.

Three raises in nine months

January took the same product up about 15%; 1 July added roughly 20% ³. Each rise lands on a higher base, so the three compound to about 59%.

The owner is shifting the chips off its books

Renters see shortage. The balance sheet sees something else. Amazon has approached outside investors to move about $8 billion of Grace Blackwell chips into a special-purpose vehicle funded by outside debt and lease them back, keeping up to 10% of the equity; the chips already run at more than a dozen sites across five US states ¹. One report ties the structure↗ to protecting Amazon's credit rating ⁴; Amazon and Nvidia did not respond to requests for comment ¹.

The backdrop is the ramp: 2026 capex is set at $200 billion against $131 billion last year, carried by a $17.5 billion term loan, a debut £4.25 billion sterling bond and nearly $100 billion of bond sales, the most among hyperscalers ¹. Q2 capex ran $53.1 billion against $45.4 billion of operating cash, with free cash flow negative in both quarters of 2026 ⁵.

The market charges less than the reservation

On demand, the same eight H100s cost less outside the reservation: Nebius lifted its H100 rate 17% to $4.50 a GPU-hour on 1 October, so eight run $36.00 against AWS's $41.53, about a 15% premium for the guaranteed hour ⁶. That premium is measured against today's AWS rate; after 7 October it widens to roughly a third. The glut case is not in this record, and on-demand rates are rising too. What moves is residual value: if the chips age badly, the loss lands on the SPV's debt holders, not on Amazon.

A billion for the towns, an open investigation

The build-out keeps accelerating: AWS said in March it would deploy more than a million Nvidia GPUs within a year ³, and TrendForce's tracker↗ shows AWS adding more megawatts than any hyperscaler this month, nearly half at Susquehanna ⁷. The other side is local. At least 75 projects worth about $130 billion were blocked or delayed in the first quarter alone ⁹, and AWS chief Garman warns that over 100 moratoriums now under consideration could cost the US the AI lead: "the U.S. could be writing its own losing ticket to this race" ⁸.

The $1 billion, five-year Built Together pledge works out to about $200 million a year, roughly a tenth of a percent of the $200 billion 2026 capex budget ¹⁹. Amazon also ended its NDAs with government agencies on new data center projects ⁸, the same week Representative Raskin opened an investigation into secrecy deals↗ across Amazon, Google, Meta and Oracle ¹⁰.

The tell is the spread

Watch the SPV: the spread on $8 billion of chip-backed debt is the market's first posted price for a used Blackwell. The rent is set by the landlord. The residual is somebody else's problem.