Nvidia has told some of its biggest customers that the servers built around its AI chips are about to cost more than 15% extra in many configurations, according to a Bloomberg report picked up across the trade press over the weekend. The increases hit systems shipping early next year, including the flagship Vera Rubin and Grace Blackwell racks, and the contract manufacturers that assemble those machines for Microsoft, Google and Oracle have already begun warning their clients. What makes this more than a routine price bump is the reason behind it. Nvidia's chips did not get 15% harder to make. The memory around them did.
Memory has quietly become one of the most expensive ingredients in an AI server. High-bandwidth memory and server DRAM now account for roughly a quarter of the bill of materials on a high-end rack, and prices have gone vertical. Analysts pegged conventional DRAM contract prices rising 90% to 95% in the first quarter of this year and another 58% to 63% in the second, a run the industry has taken to calling "RAMageddon." A single Nvidia rack can carry more than 20 terabytes of memory before you count the CPUs, so even a modest per-chip increase compounds into hundreds of thousands of dollars per rack.
The irony is that Nvidia helped build the fire it is now paying to escape. Samsung, SK Hynix and Micron, the three companies that make most of the world's DRAM, spent the past two years shifting their most advanced production toward the high-bandwidth memory that AI accelerators crave. That starved the ordinary memory market, and the shortage is not a quick fix. Gartner expects the crunch to last into 2027, while Deloitte thinks meaningful new capacity will not arrive until 2029 or 2030. With a 75% gross margin, Nvidia can comfortably pass the cost through rather than absorb it, and that is exactly what it is doing.
Here is where it stops being a data-centre story. The same memory squeeze has already escaped into the shops. Apple raised prices on Macs, iPads and other hardware in June by as much as 20%, with Tim Cook blaming soaring memory and storage costs driven by AI construction. Amazon has followed: the Echo Dot jumped 60%, from $49.99 to $79.99, and the base Kindle rose 37%, from $109.99 to $149.99, with the company citing "significant increases" in component costs. Consumer DDR5 memory has more than doubled in a year. A budget gadget in a checkout basket is now carrying a sliver of the cost of the AI boom.
The pattern matters because it reframes what the AI build-out actually consumes. It is not just electricity and GPUs. It is bidding up the price of the physical components that every other device shares, and turning a compute race into broad hardware inflation. The clearest winners are the memory makers, who suddenly hold rare pricing power over the most valuable company in tech.
There is a longer-term risk buried in it for Nvidia too. Every 15% adds to the capital needed to deploy the same amount of computing, and the more expensive its platform becomes, the more incentive hyperscalers have to lean on their own custom silicon. That shift takes years, though, and all the alternatives draw memory from the same three constrained suppliers. The shortage is happening now. The escape hatch is not.