Memory prices fell 500 billion-fold since 1960

Key Takeaways

• DRAM costs dropped from $0.50 per KB in 1970 to fractions of a cent per GB today
• HBM now represents a major share of AI accelerator costs, with no public spot market for pricing
• Stanford's dataset extends John McCallum's legendary memory price research through 2026 projections

Stanford's DAM project has published an interactive dataset tracking memory prices from 1960 through 2026, extending John C. McCallum's legendary memory price research that semiconductor professionals have relied on for decades. The data shows DRAM prices fell by roughly 500 billion-fold over six decades, a decline so steep it reshaped every industry that touches computing.

More pressing for anyone building or buying AI infrastructure: the dataset now includes HBM pricing estimates, the high-bandwidth memory that dominates accelerator bills of materials. HBM has no public spot market. Nvidia, AMD, Google, and Amazon buy it on confidential contracts. This Stanford compilation pulls together the best available analyst estimates from TrendForce and SemiAnalysis, giving outsiders their clearest view yet of what AI hardware actually costs at the component level.

What does the historical price data actually show?

The headline number is staggering but familiar to anyone who lived through the PC era: in 1990, a megabyte of DRAM cost around $1,000. Today, a terabyte of NAND flash runs about $3 at retail. That's a 300-million-fold improvement in cost per unit of storage. DRAM followed a similar trajectory, declining at roughly 40-45% annually for most of its history, a pace semiconductor economists call Moore's Law pricing.

The dataset breaks DRAM into generations: pre-DDR, DDR through DDR5. Each generation appears as a separate line, revealing a pattern that procurement teams know well. The cheapest price at any moment usually tracks end-of-life inventory being cleared out, not the current generation. If you're chasing the lowest possible $/GB today, you're probably buying DDR4, not DDR5.

NAND flash data starts in 2016 for NVMe drives, with approximate anchor points back to 2010. The methodology note is worth reading: these are cheapest listed retail prices, not confirmed sales. Obvious posting errors, where a $130 SSD shows up at $4, get filtered out. The data captures the market's floor, not its center.

Why HBM pricing matters for AI economics

The dataset's most valuable new contribution is its HBM section. High-bandwidth memory costs 3-5x more per gigabyte than standard DRAM, and it's the bottleneck component in AI accelerators. Epoch AI, whose modeled estimates appear in the dataset, tracks quarterly HBM spend across Nvidia, AMD, Google TPU, and Amazon Trainium. Their data shows HBM now represents a substantial fraction of the total accelerator bill of materials.

The generational breakdown runs from HBM2e through HBM3, HBM3e, and projected HBM4 (launching Q3 2026). A useful metric here is $/TBps: cost per unit of memory bandwidth. That number captures what AI workloads actually need from HBM. Raw capacity matters less than how fast data moves.

Because HBM sells only through confidential contracts, the figures here are analyst estimates, not transaction prices. TrendForce and SemiAnalysis provide the underlying data. It's the best public information available, but anyone citing these numbers should note the uncertainty.

How the dataset is built and updated

David Shim at Stanford compiled the project, splicing together multiple sources. The DRAM backbone comes from McCallum's original dataset, maintained from 1957 through mid-2024. From mid-2024 onward, the data pulls from Keepa's Amazon price history, tracking the cheapest new consumer DIMM each month. There's a small step at the splice point, since Amazon clearance pricing can dip below McCallum's representative lows.

NAND data uses Keepa exclusively from 2016 forward, filtering for consumer NVMe drives only. No SATA, no enterprise or datacenter SKUs. The HBM cost breakdown updates quarterly from Epoch AI. DRAM and NAND refresh monthly.

The raw data is downloadable as CSV, and the interactive charts let you toggle series, zoom with a slider, and export images. For researchers and analysts, that export function matters. You can pull exactly the chart you need for a slide deck or report.

What the accelerator cost breakdown reveals

Epoch AI's component breakdown stacks HBM, logic die, packaging/CoWoS, and auxiliary costs into a quarterly total for the four major accelerator designers. You can view it as absolute spend ($B/quarter) or as percentage share. The packaging line, representing advanced interconnect technologies like TSMC's CoWoS, has grown faster than most observers expected. HBM and packaging together now drive accelerator costs more than the logic die itself.

This has implications for AI training economics. Memory bandwidth, not compute, increasingly determines what you can train and how fast. The companies that lock up HBM supply, or develop alternatives, gain structural advantages. SK Hynix and Samsung dominate HBM production. Their capacity constraints ripple through every major AI project.

Caveats worth knowing before you cite this

The methodology section deserves careful reading. These are nominal USD prices, not inflation-adjusted. They're retail prices, not contract prices. The cheapest listing often reflects end-of-life clearance, not leading-edge product. For enterprise procurement, actual costs will differ.

HBM figures are modeled estimates, not measured transactions. The dataset is transparent about this, but anyone using the numbers should be equally transparent. Academic papers and industry reports should cite the underlying sources, TrendForce, SemiAnalysis, Epoch AI, not just the Stanford compilation.

Frequently Asked Questions

Where can I download the historical memory price data?

The Stanford DAM project hosts the interactive charts and CSV download at dam.stanford.edu/memory-prices.html. The CSV includes every data point with its source.

How often is the memory price dataset updated?

DRAM and NAND prices refresh monthly from Keepa's Amazon price history. HBM cost estimates update quarterly from Epoch AI. The historical McCallum backbone is fixed.

Why is HBM so much more expensive than standard DRAM?

HBM stacks multiple DRAM dies with through-silicon vias and requires advanced packaging like CoWoS. Manufacturing is complex, yields are lower, and only two companies produce it at scale. Supply constraints keep prices 3-5x higher than commodity DRAM.

Does the dataset include enterprise or datacenter pricing?

No. The NAND data tracks consumer NVMe drives only. DRAM tracks consumer DIMMs. Enterprise and datacenter pricing follows different curves and is harder to track publicly.

Source: Hacker News: Best