Stanford's Memory Price Dataset Maps 66 Years of DRAM, NAND, and HBM Costs

Sixty-six years of memory pricing data, from SDRAM through HBM4, just became searchable in a single interactive chart. David Shim's DAM project at Stanford publishes the dataset, merging John C. McCallum's classic memory-price backbone (1957–2024) with live Keepa Amazon retail scrapes from mid-2024 onward and Epoch AI's quarterly accelerator cost models.

From SDRAM to HBM4: What the Dataset Tracks

The chart plots the lowest advertised $/GB on a log scale for three memory types: DRAM, NAND flash, and HBM. DRAM gets a per-generation breakdown—Pre-DDR, DDR, DDR2, DDR3, DDR4, DDR5—with generation inferred from product descriptions (older points are approximate). NAND covers the cheapest consumer NVMe SSD each month from 2016 (with four pre-2016 anchors). HBM is broken out by generation: HBM2e, HBM3, HBM3e, and the projected HBM4 (launching Q3 2026), plus a cost-per-TBps metric.

The HBM data is the real catch. Because HBM is sold only to accelerator makers under confidential contracts, there's no public spot market. The dataset relies on sparse industry-analyst estimates—TrendForce and SemiAnalysis—and bandwidth specs from JEDEC and Rambus. For the accelerator cost breakdown, Epoch AI (CC-BY) provides a quarterly production-volume-weighted average across the four largest designers: Nvidia, AMD, Google (TPU), and Amazon (Trainium), stacked by HBM, logic die, packaging/CoWoS, and auxiliary components.

Why HBM Pricing Matters for AI Hardware

HBM now dominates the bill of materials for AI accelerators. The dataset's component breakdown (absolute $B/quarter and share) shows exactly how much each generation costs relative to logic and packaging. As models scale, HBM bandwidth and capacity become the bottleneck—and the cost per TBps is the key metric. For the first time, you can trace that cost from HBM2e through HBM4 and see the projected decline. The raw CSV is downloadable, so you can build your own models.

Retail vs. Contract: Reading the Fine Print

Every number comes with honest caveats. $/GB is the cheapest listed retail price in nominal USD—not contract, average, or inflation-adjusted. Retail lags contract, and the cheapest listing often tracks an end-of-life generation being cleared out, not the leading edge. Keepa scrapes are not confirmed sales; obvious posting errors (a $130 SSD listed at $4) are dropped. The DRAM line splices two sources at mid-2024, so a small step is expected. HBM figures are modeled estimates, not measured prices.

Updates are monthly for DRAM and NAND, quarterly for HBM (via Epoch AI). The McCallum backbone and HBM estimates are fixed. Shim provides a downloadable CSV with every point and its source.

As HBM4 enters production in Q3 2026, this dataset becomes the definitive reference for tracking whether cost-per-bandwidth continues its steep decline—and what that means for the next generation of AI hardware.

Source: Historical memory prices 1960-2026
Domain: dam.stanford.edu