Rack prices up "more than 15%", and Amazon's $220bn capex raise does not cover it

A single tidy percentage for next year's rack prices is a reading of a range, not a datapoint. What was reported is "more than 15%", varying by Nvidia chip generation and memory configuration ^[1][3]. Anyone quoting 16, 17 or 18 per cent is interpolating inside that band.

The pricing power sits with Samsung, SK Hynix and Micron, which between them make most of the world's DRAM and have gained what Bloomberg's sources call unprecedented leverage ^[5]. The awkward part for buyers is that memory is not a trim level: how much DRAM an Nvidia accelerator is paired with determines how much use you get out of it ^[6]. The usual answer to an expensive input, buy less of it, degrades the asset. Nvidia is not absorbing the difference either. Gross margin is 75%, and TSMC still cannot meet demand ^[7], so the increase travels through the contract builders, who have already notified Microsoft, Google and Oracle ^[4].

Then the arithmetic on Amazon's number. Andy Jassy told investors in late July that capital expenditure would be $220 billion this year rather than $200 billion, because of higher memory costs ^[9]. That is a 10% raise ^[2] against systems repricing by at least 15% ^[1]. At those prices, $220 billion buys roughly $191 billion of equipment in old money, about 4% less physical capacity than the original $200 billion plan implied ^[1]. The demand side has not slowed to match: AWS turned over $42.2 billion in the second quarter, up 37% ^[11].

For a sense of what uncushioned pass-through looks like, Amazon's own base Echo Dot went from $49.99 to $79.99 ^[12], a 60% rise ^[3], after the company said it had absorbed component increases as long as it could ^[13]. The gap between 60% and 15% is a bill-of-materials story: memory dominates the cost of a smart speaker in a way it cannot dominate a rack where a single accelerator carries tens of thousands of dollars of price on its own ^[7].