The Stanford curators document a ten-order-of-magnitude compression in memory cost from $5/bit in 1960 to $0.000000003/bit today, presenting it as a singular industrial phenomenon. Their dataset frames this as unmatched by any other technology — not transistors, solar, or LEDs — establishing memory's price descent as the central enabling fact of the computer age.
By submitting the Stanford page to HN where it reached 239 points, vga1 surfaces the dataset as a piece of foundational industry history worth broad attention. The submission framing treats the long-run price curve as the story, not merely the recent reversal.
The curators explicitly flag the inflection in their page notes, attributing it to 'sustained allocation pressure from AI accelerator demand' with DDR5 capacity being re-routed to HBM stacking. They emphasize this is the first multi-year wrong-way bend since the mid-1970s shakeout, signaling something qualitatively different from prior cyclical dips.
The editorial argues that garbage collectors, JVM heap defaults, V8 sizing, Python's reference-counted bloat, and Electron's per-window baselines are all downstream consequences of assuming memory would keep getting cheaper forever. With the curve bending the wrong way, these accumulated profligacies become liabilities rather than rational trade-offs.
Stanford's Digital Assets Management group quietly refreshed its long-running memory-price dataset (`dam.stanford.edu/memory-prices.html`) and shipped it to the front page of Hacker News, where it hit 239 points. The page itself is unglamorous — a single long table, a log-scale chart, and a list of citations stretching back to *Computers and Automation* issues from 1957. But the data is the most complete public record of what a bit of working memory has cost, in nominal and inflation-adjusted dollars, for the entire history of the commercial computer industry.
The headline number: a bit of DRAM cost roughly $5 in 1960 and costs roughly $0.000000003 today, a compression of about ten orders of magnitude. Put differently, one dollar bought you about 0.2 bits when Eisenhower was president and buys you about 330 million bits now. The same dollar, in 1960 money and accounting for inflation, would have purchased a single bit and left change. There is nothing else in the industrial economy — not transistors, not solar panels, not LED lumens — that has fallen this far this fast.
The new tail of the dataset is what makes the 2026 update interesting. For the first time since the mid-1970s shakeout, the curve has multiple consecutive years that bend the wrong way. Spot DRAM, contract DRAM, and especially HBM3e all trended up through 2024 and 2025. The Stanford curators flag the inflection explicitly in the page notes: "2024-2026 reflects sustained allocation pressure from AI accelerator demand; conventional DDR5 capacity has been re-routed to HBM stacking."
It is hard to overstate how much of modern software practice was built on the assumption that the descent in this chart would continue forever. Garbage collectors are tuned for a world where you can over-allocate by 2-3x and pay for it in cents. JVM heap sizes, Node's V8 defaults, Python's reference-counted bloat, Electron's per-window 150MB baseline — all of these are downstream of a price curve that fell about 30% per year, compounded, for six decades. The free-lunch era of "throw RAM at it" is the longest-running implicit subsidy in our industry, and the Stanford data is the receipt.
Compare the memory curve to the two adjacent curves it is usually grouped with. Compute-per-dollar (Moore's law in its loose modern form) has slowed but not reversed; TSMC N3 is more expensive per transistor than N5, but you still get more transistors per dollar at the node frontier. Storage-per-dollar (Kryder's law) actually decoupled from this trend years ago — HDD $/TB has been flat since roughly 2018, and the industry quietly pretends NAND is the heir apparent even though NAND $/TB also flattened in 2023. Memory was the last of the three to hold its exponential. It is no longer holding it.
The community reaction on the HN thread is split in a revealing way. One camp reads the chart as vindication of "memory is the new bottleneck" — the same argument Jim Keller, Sophie Wilson, and a dozen Chips and Cheese articles have been making for three years. The other camp points out, correctly, that nominal price isn't the right axis: the unit of compute that matters now is the token, not the byte, and DRAM-per-token is still falling because models are getting more efficient at the same wall-clock memory budget. Both can be true. The chart measures dollars per bit. It does not measure dollars per useful unit of work, and the gap between those two metrics is widening fast.
A third reaction worth flagging: several commenters pointed out that the chart's worst inflections — 1974, 1988, 1996, 2017 — all preceded major shifts in how application software was written. After the 1988 spike, embedded developers spent a decade obsessing over byte-level layout. After the 2017 spike (DRAM contract prices roughly doubled in 18 months), nobody changed anything, because the curve resumed its descent within two years. The 2024-2026 spike is structurally different: HBM demand is not cyclical. It is the new floor under which DDR5 supply is allocated. If you are building a backend today, you should plan as if 32GB-per-container is the new 8GB-per-container — a budget, not an abundance.
Three concrete shifts are worth thinking about. First, memory-aware runtimes are about to matter again. Languages that let you reason about allocation — Rust, Zig, modern C++, even Go with its arena proposal — are going to look better in cost models than they did during the cheap-RAM decade. This isn't a hot take about "the death of GC"; it's a budgeting observation. When DRAM was free, the productivity premium of managed runtimes was worth their memory overhead by an order of magnitude. When DRAM doubles in two years, the overhead column on the spreadsheet starts to bite.
Second, the AI inference cost stack is going to keep diverging from the AI training cost stack in a way most teams haven't priced in. Training cost is dominated by HBM and interconnect; inference cost, especially for the long-context regime that everyone is moving to, is dominated by KV-cache memory. If you're paying for tokens via API, your provider's KV-cache-per-request cost is rising in real dollars even as model perplexity-per-flop keeps falling. That's why context-window pricing is getting weirder — see Anthropic's recent prompt-caching introduction and OpenAI's batch tier — and it will keep getting weirder until either HBM yields recover or the industry standardizes on sparse-attention variants that don't blow up the cache.
Third, the cloud pricing arbitrage is shifting. AWS and GCP have not raised on-demand memory-optimized instance prices in step with spot DRAM, because they bought their inventory on long-term contracts in 2022-2023. That hedge expires somewhere in the next 12-24 months. If your reserved-instance plan ends in 2026, get ahead of it now.
The Stanford dataset is the kind of artifact that gets cited for decades because it makes a long-tail truth legible at a glance. The truth here is straightforward and uncomfortable: a 65-year exponential has paused, and we built our entire software culture during the descent. The next decade of systems work is going to be partly about rediscovering the discipline of treating memory as a finite resource — not because it's morally superior, but because the chart says so. Whether the pause is a brief detour or a permanent step-change depends on HBM yield curves, on whether 3D DRAM stacking pans out at Samsung and Micron, and on whether the next generation of accelerators stops eating quite so much bandwidth. None of those are settled. The chart is the only honest map of the territory we have, and it just changed direction for the first time in most working engineers' careers.
If my memory serves me correct (no pun intended), when I was a kid I remember bugging my mom to buy me like 2 or 4 1 MB modules, it was at least 50 bucks or 100 bucks each.Now everyone's going to talk about how cheap everything is by comparison - but someone needs to talk about how oppressively
The log scale is nice to compare decades. Wether it's inflation-adjusted or not isn't too important but it's still a factor of 10, which would show in a linear recent graph. The fact that we're comparing GBs instead of the average RAM stick shows how much the price has decreased
Look at it this way: while the upfront cost to scale up production is huge, prices are now high enough to justify it even if demand is expected to drop abruptly later on. So if you can wait 5 years for your next PC, 1TB RAM might go for what 64GB would have cost without the AI demand spike.Granted,
In the first graph, if you hover over the DRAM line you'll notice that the most recent data points are for DDR3. One of the data points from 2025 is a 2 GB stick. This paints a more rosy picture than the situation deserves.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Says, not inflation-adjusted. With reason; adjusting those 1960-1980 prices for inflation would make the graph a lot taller.Pricing "per GB" before 1990 is unrealistic, though; nobody thought in GB or purchased GB quantities, or conceived of GB systems. I remember a moment circa 1973 when