# Nvidia AI Server Prices Rise More Than 15% as Memory Swells to a Quarter of Rack Cost > Source: > Published: 2026-08-23 18:13:15+00:00 *The memory squeeze has reached the one company you would think could absorb it.* Nvidia AI server prices are going up by more than 15% in many configurations, and the company has already told its biggest customers why: memory. [Bloomberg](https://www.bloomberg.com/news/articles/2026-08-22/nvidia-customers-notified-about-ai-related-price-hikes-above-15) reported the warning on Saturday. The increases land on systems shipping early next year, and they cover both the current Grace Blackwell racks and the Vera Rubin generation replacing them. The notices took a roundabout path. Nvidia does not sell most of these racks directly, so the contract manufacturers that assemble them for Microsoft, Google and Oracle got the word first and passed it up the chain. How steep the increase is depends on which generation you are buying and how much memory you specify, which is the tell. This is not a broad margin grab. It is one line on the bill of materials dragging everything else along with it. ## Memory is now a quarter of the rack For two years the AI buildout has been framed as a fight over GPU allocation. That bottleneck has quietly moved. Server DRAM contract pricing roughly doubled across the first quarter of 2026, and memory now accounts for something like 25% of what a high-end rack costs to build. When a component that big doubles in price, there is no clever engineering that hides it. The rack-scale figures make the point better than percentages do. Blackwell-based GB200 NVL72 systems have been going out at roughly $2.8 to $3.4 million each. Pre-shipment inquiries on the Vera Rubin VR200 NVL72 racks that succeed them are reportedly landing between $5 and $7 million. Not all of that gap is DRAM and HBM, since Rubin is a bigger part in every dimension, but the memory bill alone is said to have climbed around 2.5x between the two generations, and it is the one component nobody can design out. What is striking is that customers appear to be swallowing it. Nobody in this segment delays a Rubin deployment over a 15% surcharge, because the alternative is not deploying at all, and the market punishes that far harder than a capex overrun. Which is precisely the problem for everyone else. ## Why this lands on your next RAM purchase Memory fabs allocate wafers where the money is, and right now the money is emphatically not in consumer DIMMs. Conventional DRAM contract prices reportedly jumped 90 to 95% quarter over quarter in Q1, with analysts forecasting another 58 to 63% for Q2. Anyone who has priced a 32GB DDR5 kit lately has already felt the edge of that. The usual relief valve here is demand destruction: prices spike, buyers balk, capacity frees up, prices settle. Nvidia’s note is evidence that the valve is not opening. Hyperscalers are absorbing double-digit increases without blinking, which tells the memory makers exactly where to point their capacity for the next several quarters. Consumer RAM, prebuilts, laptops and any graphics card with a generous VRAM buffer are all competing for what is left over. There is a version of this that ends well. New capacity comes online, HBM yields improve, the panic buying stops. None of it happens inside a quarter. If you are planning a build, the memory line on your spreadsheet is the one most likely to look worse in six months than it does today, and it is the rare component where buying early is genuinely the rational move. *Sources: Bloomberg, Tom’s Hardware, Fortune*