RAM per GB now costs what it did in 2010 — a 16-year price regression
A Stanford chart plots the cheapest $/GB of memory from 1957 to 2026 on a log scale. The spike at the right end erases over a decade of progress: by per-GB price, DDR5 today sits roughly where DDR3 was in 2010.
The dataset is a revival: the long-defunct McCallum/jcmit price archive, pulled off the Internet Archive and extended past mid-2024 with Amazon pricing. Prices are nominal, not inflation-adjusted.
Commenters split on whether the chart even means anything: '$/GB is a totally useless unit' when 1960s machines shipped with kilobytes, says one. Another notes the log scale flatters the recent spike, and a third faults it for never marking the DRAM cartel years.
Sixteen years of progress, gone in one spike
Stanford's DAM project published an interactive chart of historical memory prices — the cheapest retail $/GB for DRAM, HBM and NAND flash, plotted on a log scale from 1957 to 2026. Maintained by David Shim, it is built on the McCallum memory-price dataset (1957–2024) and extended from mid-2024 using Keepa Amazon data. The numbers are nominal — "not contract, average, inflation-adjusted, or a confirmed sale price" — and HBM figures are modeled estimates, since HBM sells only to accelerator makers under confidential contracts. The story of the data is its right edge: after decades of near-monotonic decline, a recent spike pulls today's price-per-GB back to where it sat over a decade ago.
The thread's clearest reading of that spike, by generation:
| Comparison | What the curve shows |
|---|---|
| DDR5 today vs. DDR3 in 2010 | Roughly the same $/GB — a ~16-year regression |
| DDR3-to-DDR3 over time | Today's level last seen in 2018, '17, '15, '13, '11 |
| 1990s | The famous price crash — "good times" |
| 2010s onward | Curve flattens as Moore's law winds down |
The thread: a chart that flatters and omits
The sharpest pushback in the discussion is that $/GB is the wrong axis for most of the timeline.
"$/GB is a totally useless unit value because it's an arbitrary size… It's expensive relative to what you need to accomplish a useful task." — a commenter arguing the y-axis should be $/average-workstation-memory
Others attacked the presentation itself. One pointed out that pricing in GB before 1990 is anachronistic — "nobody thought in GB or purchased GB quantities" — recalling a 1973 IBM upgrade where a box of carefully packed circuit boards was, in the engineer's words, "a meg." A high-karma reader flagged what the log axis hides.
"It certainly doesn't look as bad as it really is when presented on a log scale chart." — a commenter with \~5.7k karma
And one reader faulted the project for omitting the industry's own history.
"This failed to annotate where the cartel years and when the cartel was 'broken up'. Not a bad assignment's work but clearly lacking the domain awareness necessary to report the complete story through graphs." — a commenter on the missing DRAM-cartel annotations
The recurring lament: software ate the gains
Several threads converged on the same human note — that cheaper memory never felt cheaper because software grew to fill it. The plainest version came from a reader recalling $50–$100 for a single 1 MB module as a kid, then adding that browsers and operating systems are now "oppressively hungry compared to in the past… This is no HIMEM.SYS." That is the quiet implication the numbers can't draw: a 16-year price regression bites hardest precisely because every modern app now assumes the abundance that the spike just took away.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
A router that scores every prompt for difficulty without calling a model
Wayfinder reads word count, code blocks and headings, hands each prompt a 0.0-1.0 score in under a millisecond, and sends anything under 0.5 to a local model. No inference, identical every time. On RouterBench the structural score alone cut costs 61%.