Stanford dataset charts memory prices from 1960 to today — DRAM, NAND, and HBM

A Stanford DAM project led by David Shim has compiled an interactive history of memory and storage pricing stretching back to 1960, building on John McCallum’s well-known memory-price dataset. The core chart plots the cheapest dollar-per-gigabyte figures over time on a log scale for three technologies — DRAM, NAND flash, and high-bandwidth memory (HBM) — with the DRAM line further broken out by generation from pre-DDR SDRAM through DDR5. All raw data is downloadable as a CSV, and DRAM and NAND figures refresh monthly from Keepa pricing while HBM updates quarterly.

The project leans into the current AI-hardware moment with two additions beyond commodity memory. One models quarterly accelerator costs across the four biggest designers — Nvidia, AMD, Google’s TPU, and Amazon’s Trainium — stacked by component (HBM, logic die, CoWoS packaging, and auxiliary parts) using estimates from Epoch AI. The other tracks HBM pricing by generation, from HBM2e through a projected HBM4 launching in Q3 2026, including cost per unit of bandwidth ($/TBps).

The methodology is candid about its limits. The $/GB numbers are the lowest nominal retail listings, not contract or inflation-adjusted prices, and they often reflect end-of-life parts being cleared rather than the leading edge. Because HBM sells only through confidential contracts with no public spot market, those figures are sparse analyst estimates from TrendForce and SemiAnalysis rather than real transaction prices. The DRAM series also splices two sources at mid-2024, producing a small expected step in the curve.