Memory inflation reaches the rack: Nvidia's 17% hike lands on Microsoft, Google and AWS

The pass-through is easier to read against the capital expenditure line than the price list. Bloomberg's projection, as carried by Seoul Economic Daily, has major US Big Tech capex going from $781.9 billion this year to $975.4 billion next year ^[11], a rise of 24.7%, or $193.5 billion of new money ^[18]. If every one of those dollars went into racks, and racks cost 17% more ^[1], the installed fleet would grow about 6.6% ^[19]. It will not be every dollar, because that budget also buys land and power distribution: LS Cable & System is planning its first busduct plant outside its Gumi base, in Poland or Northern Europe ^[13], on the back of long-term supply contracts worth roughly 5 trillion won at its US arm ^[14]. The shape of the problem survives the caveat. Most of the visible capex increase is being consumed by unit price rather than unit count.

The mechanism is unglamorous. Memory sits inside the rack bill of materials, and Nvidia has been absorbing rising memory chip costs on parts essential to its GPUs and systems ^[10]. What is being repriced is not a new product tier but announced generations already in customers' plans, with the increase applied to systems shipped next year ^[9]. Both reports date from August 2026 ^[22][23], which puts the affected deliveries in 2027 ^[21] and leaves roughly five months between notice and the early-year effective date ^[24]. Any 2027 capacity plan built on a 2026 quote is already carrying a stale number.

The squeeze is also wider than HBM allocation. Alongside the continuing high-bandwidth memory shortage, general-purpose chip scarcity pushed Amazon to raise the Echo Dot from $49.99 to $79.99 ^[6], a 60% increase ^[25] on a device with no accelerator in it. That is the same cost curve arriving at two very different price points, which argues against treating the rack increase as an artefact of AI demand alone.

For anyone modelling compute cost, the practical consequence is that rack price has stopped being an input and become an output. It is a function of memory contract pricing, sized by configuration ^[8]. A fixed dollar-per-rack assumption in a 2027 model is a DRAM forecast that nobody has written down.