Stanford Project Publishes Seven Decades of Memory Price Data

Stanford's DAM project has published an interactive dataset that follows memory and storage prices across nearly seven decades, from 1957 to the present. The page, compiled and maintained by David Shim, collects historical lowest prices per gigabyte for DRAM, NAND flash and high-bandwidth memory (HBM), and lets visitors switch between nominal and inflation-adjusted dollars.

What the data shows

The main chart plots the lowest cost per gigabyte on a log scale over time, with one line each for DRAM, NAND flash and HBM. A second view breaks the DRAM line out by generation — pre-DDR SDRAM and core memory, DDR, DDR2, DDR3, DDR4 and DDR5 — and the project notes that generations were inferred from product descriptions, so older points are approximate. A separate chart covers HBM pricing by generation, from HBM2e through HBM3 and HBM3e to the projected HBM4, measured both per gigabyte and per terabyte per second of bandwidth. HBM4 is expected to launch in the third quarter of 2026.

Values can be read in nominal dollars or adjusted for inflation to constant 2024 dollars using annual US CPI-U averages. The full dataset, with a source noted for every point, is downloadable as a CSV file.

Where the numbers come from

Deep history from 1957 to 2024 comes from John C. McCallum's classic memory-price dataset, preserved through the Internet Archive. From mid-2024 onward, DRAM figures refresh monthly from Keepa, which tracks Amazon retail price history — specifically the cheapest new consumer DIMM listed each month. NAND pricing follows the cheapest consumer NVMe SSD each month from 2016 on, with four approximate pre-NVMe anchor points covering 2010 to 2016, since no equivalent flash dataset exists. SATA and enterprise drives are excluded.

HBM works differently. It is sold only to accelerator makers under confidential contracts, so there is no public spot market; the charts therefore rely on industry-analyst estimates from TrendForce and SemiAnalysis. Spend and component breakdowns — HBM, logic die, packaging and CoWoS, and auxiliary parts — come from Epoch AI, modeled as a production-volume-weighted average across Nvidia, AMD, Google's TPU and Amazon's Trainium.

Caveats and limits

The authors flag several limitations. Prices are the cheapest listed retail prices in nominal USD, not contract or average prices, and retail tends to lag contract pricing. The cheapest listing in a month often reflects an end-of-life generation being cleared out rather than the leading edge. The DRAM series splices two sources at mid-2024, so a small step there is expected. Obvious listing errors are filtered out: any month in which a drive appears more than 60 percent below its own typical price is dropped. HBM figures remain modeled estimates rather than measured transactions. DRAM and NAND refresh monthly; HBM updates quarterly.