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 I saw an IBM CE about to do an upgrade on a 370 system at Cal Berkeley. He had a box with several carefully-packed, large circuit boards. "So, is that a megabyte?" I asked. "Yup, that's a meg."
The author clearly wasn't implying that these were 1GB chips. They just wanted to show a graph scaled per unit of memory. It could just as well have been per byte, and the graph would've been identical but the values on the left would be changed by a factor of a billion.
You could argue that you'd rather see a "price per typical-sized RAM chip as sold at the time". That would also be a perfectly valid thing to graph (though a bit more subjective), but it doesn't invalidate this one. Since per byte (or GB or whatever you want to say) has continued downward all this time, it makes the recent spike all the more notable.
(I'm not sure it's right to label vacuum tubes and core memory as "DRAM" though.)
You got it backwards; ending the gold standard was very much a unilateral decision by the United States because Nixon couldn't handle making politically unpopular decisions to cut spending and/or end the Vietnam war. Many countries that had their gold reserves held via US dollars were livid.
What? Gold underwent a massive revaluation with the end of Bretton Woods in 1971. It was prior to then that government were actively involved in making the price of gold artificially low.
Fun fact: under 31 U.S. Code § 5117 a troy ounce of gold is still valued at 42 and 2/9ths dollars.
I wouldn't go so far as to say "nobody". Electric Boat had 2 GB memory in one of its systems at that time, with the hardware capacity to increase to 4 GB. It sounded insane at the time, but it absolutely existed, and thereby seems reasonable to include it in any research of historical pricing.
Yes, you really need "dollars per amount of RAM you need for standard computing tasks." Windows 11 requires a bare minimum of 4 GB of RAM, Window 10 only needed 1 GB.
If what you're interested in is fluctuations in production versus demand then you absolutely do not want a subjective metric. Measures of the form dollars per unit, units per watt, units per flop, etc are what you're interested in.
Discussion of memory in terms of words and their bit length, time to complete a task is more meaningful to intent on use and compaction, see greycode technique. Dollars of slop a unit sacrifice skills in the industrial base for the gain of paper profits at the repeating business meetings.
Have you ever actually tried using Windows 10 with 1GB of RAM? I wouldn't consider it suitable for "standard computing tasks."
And that's the hangup, what do you consider a "standard computing task?" On what OS? Running what software? How well? Plenty of people were still using XP in 2009, so is 256 MB of RAM okay for "standard computing tasks" in 2009?
> Have you ever actually tried using Windows 10 with 1GB of RAM? I wouldn't consider it suitable for "standard computing tasks."
I have [0], and it's actually not quite as bad as you would expect. It certainly wasn't fast, but I had no problem using it for basic web browsing and document editing. The painfully slow hard drive and processor speeds on that computer actually caused more issues than the lack of RAM.
Well it's complicated. Y2K was a combination of logic issues and the consequences of certain inefficient ways to store dates, like text and BCD. Migrating to binary could fit plenty of dates into the same space or even less.
In particular, 16 bits is enough to store the entire date, year month day, from 1900 into mid 2079. Any date format that couldn't go past 1999 was probably using 24-48 bits.
Abstractions on abstractions on abstractions; background tasks and their abstraction stacks; increased cache and buffer sizes to take advantage of increased typical memory capacity. For an example of the latter, handling TCP on a Commodore 64 is a problem because the memory can only fit about 45 packets with nothing left over, but now you can just allocate a megabyte receive buffer per connection.
Neither do the developers, because until recently, RAM was so cheap it didn’t matter, and we were in a situation where almost no one ever needed to consider “how much RAM will this take?” when writing code.
I once held in my hand the main part of a ferrite core memory module from the early 70s. It was kilobytes at best.
I also recall looking at recommended requirements for Dungeon Keeper 2 - 266MHz CPU, 64MB RAM and thinking "that's absurd - no such device exists!". I was a kid back then, so what did I know?
Later on in college a friend showed us his absolute monster of a laptop with a whopping 8GB of RAM - he could spin up several VMs on one device! Groundbreaking on a (nominally) portable device.
So yeah, safe to say the notion of gigabytes of RAM anywhere close to a regular person belongs firmly to the 21st century.
So total system cost per unit of memory is going up.. 2GB costs in 1985 was $2 million (from the graph), a cray-2 was $16 million (from wikipedia). A GPU server with 8xB200 today can be had for ~$500k (estimate), 1.5TB memory is $25k (from the graph).
The natural unit of measure for integrated circuits is a power of 2 since that's what the systems operate in. It's so natural that early 9 and 36 bit architectures were squeezed into 8 and 32 bits as it just works so much more efficiently.
Long term storage and communications? Those start to introduce things like human division of timings, frequencies, and other analog systems like rotating disks. It still generally makes sense fab actual flash chips in various powers of 2 though. The discrepancy there tends to be various forms of 'overhead' for the translation table / wear level indirection, over-provisioning, and even variations in density caused by different levels of physical cell utilization.
Still, most network stuff ships around packets of 'up to' 1500 bytes ( https://en.wikipedia.org/wiki/Ethernet_frame and lets just exclude jumbo frames ) so arguably it'd be better to talk about all computer measures in binary powers of two, exclude the marketing huckster trying to make things more impressive by shoehorning SI engineering units into a realm that uses binary math.
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 per GB rather than per unit (much smaller decrease).
But a linear graph that represents only the last decade and where the bottom is 0 (not the min value) would tell a different story, but I guess we already know that story because we're living it.
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 hungry browsers and OSes are compared to in the past. This is no HIMEM.SYS
There’s been a sharp divergence in memory requirements. Talk to developers and they think that 32GB is the bare minimum these days, with 64GB or more preferred. They’ll point to Electron and Chrome tabs and everything else.
Then you sit down with an average computer user on their 8GB RAM MacBook Neo and they’re in love with how fast and smooth it is, even with their chrome tabs and the company Slack up and Spotify in the background.
I still have an older 8GB machine to kick around with on the go when I don’t want to haul the expensive laptop. It’s fine, even for a lot of development.
Because they use one or two apps at at time, the ones they must spend all their time to perform their job. E.g. Excel and a web app to work on invoices and a stack of paper documents. I see 8 GB on Windows PCs too.
I think it would be better if one has the discipline to just use older machines and play older games and only visit certain websites that can be visited on older version of browsers. A second-hand 16GB laptop can go a long way.
But yeah that probably sucks from time to time, especially for young people.
My laptop has 8 GB. I write blog posts, I have a dozen-ish tabs open, I do KiCAD things (including 3D renders!). Works great. I was doing Verilog synthesis on a similar machine in college in 2020.
The truth is that, if you do the same things you were doing with your computer 10 years ago, well then you don't need a new computer!
If all you do is write books, a Pentium III will do the job just as well as a brand new PC.
Of course, the web throws a wrench in this. Word 2003 is still far more capable than Google Docs, yet tons of people opt for the cloud slop because it's convenient and free-as-in-beer. And, Google Docs will continue to become less efficient with time.
You can do a lot on old machines but developers also need to optimize a bit. Youtube almost plays on a 20-year-old machine, which means with some effort it'll play just fine. Most the other sites work just fine.
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, if you need a new system before then, you're SOL.
One thing to look out for is supply capacity curiously going offline in 2030 or whatever. That would hint at market power or collusion.
It’s possible we’ll see a huge price drop on the near term but SSD + Cache + GPU’s seems to have changed the equation where RAM speed is considered more important than size. And from a pure architecture standpoint it makes sense.
Lowest 2012 price listed is 3.7 (2012-10-30) vs highest listed in 2026 is 5.375 (2026-2-1), which overlaps based on the margin for error involved. https://www.usinflationcalculator.com/
> 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.
Given the nature of the industry and how critical the product is I think it would make more sense for governments to bankroll fab construction in a way that the public takes on the risk of consumer prices falling below a certain level within some limited timeframe. Mildly subsidized chip production seems like a much better downside than the current sky high prices.
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.
TIL someone took over the now defunct jcmit dataset[1] (archive[2]). I expected his dataset to die off when his website did, but I guess someone found the data dump on archive.org and revived it. Which raises a question: how will this dataset fare five years from now?
This graph is the touchstone one should rub all the "RAM and storage are no longer a commodity" bs that micron, sk hynix, samsung, western digital, seagate and others are peddling as of late while the valuation of their companies have gone from "supplier of widely available fungible goods" to "state-of-the-art moat AI backbone tecnology".
One could also blame crypto and AI (they're clearly responsible for some of the volatility in the graph), but I can see the curve flatten in the 2010s, just as Moore's law ended.
1979 to 2009 in the OP graph has a pretty steady drop from 10^7 to 10^1 USD/GB: 6 OOMs in 30 years. Then till before the recent spike it was around 1 OOM in 15 years: 1/3 the rate of progress on a log scale.
When it comes to CPU progress we blame the end of Dennard scaling several years before the knee in this memory curve. I'd guess the story of memory is similar in also hitting technical difficulties, but I don't know.
Moore's law didn't end in any broad sense and certainly not that far back. It's a tiring piece of misinformation that just won't die.
Progress has consistently become more difficult (ie more expensive) but has generally kept up. The scaling of a couple specific technologies noticably slowed down a few years back but that's not the general case.
The node names aren't representative of the reality.
Arguably they already did with the "cloud native" systems. There were plenty of examples personally known to me in the mid and late 2010s of smaller tech companies trying to run production PostgreSQL on 8-16 GB of RAM because they didn't want to pay the cloud RAM tax. Many "cloud native" systems were designed under these (mostly artificial IMO) RAM constraints.
Is that because the amount of available memory is limited for a single process? You can always add more storage and storage access is relatively the same regardless of whether it comes from the SSD inside the server or sits in another rack. Storage is a pretty linear cost when you're a cloud host buying storage in the hundreds of PB numbers. Whereas for memory, if you want the whole thing, you need the whole server even if your process is light on CPU requirements.
It's amazing how consistently thr lower memory cost have expanded the set of economic viable applications : cheaper hardware doesn't just improve existing software it also enables software that was not possible before
The memory manufacturers have made an interesting mistake. The tech giants of the world will be working to replace them from the supply chain as soon as possible. China already makes it's own RAM albeit at 16nm but you can bet they are working to get down to 4nm.
DRAM hit a barrier at 10 nm a few years ago[0], so 16 nm is actually even closer to state-of-the-art. E.g. Micron newest node (1-γ) is their sixth at 10 nm [1] and their first EUV-based node.
The problem is that DRAM is fundamentally based on storing charge in a capacitor and how much charge a capacitor can store is a result of the geometry of the capacitor. So either someone will have to figure out a way to make the same size capacitor take up less space on the RAM chip (this is what broke the previous 20 nm barrier) or someone will have to invent a practical way of making RAM with less than 1 capacitor per bit.
China often exhibits tremendous internal market competition, so it's possible that different Chinese suppliers will race each other to the bottom (Chinese firms are really good at suriving on ultra-thin margins) making prices even lower than 10% below the premium providers.
Unfortunately, this is unadjusted prices, and 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.
This is probably the first real thing that is affecting me personally with this whole AI business. Having to pay more for device upgrades going forward. I hope the demand settles or new memory production offsets the demand.
The truly absurd part is that datacenters are barely being built and those that are built can't be turned on because they don't have enough power. Satya Nadella admitted recently that they have warehouses full of unused hardware because they a) can't get datacenters built and b) don't have enough power i.e. this whole RAM scandal is a bloody joke. If OpenAI goes bust (their financials are a huge mess and lots of red flags - they might not even survive to IPO!) and then what will happen to all those "inked" deals to buy all that RAM?
turns out things are not that bad! we just rolled back to 2010.
oh, wait, now every app is a browser instance. shit.
EDIT: so, how did I arrive at 2010, you ask? I looked at DDR5 pricing and found the closest pricing per GB in the past. this turned out to be DDR3 memory. I think it's totally fair since it was the latest and greatest thing back then, much like DDR5 is now. although, if we compare DDR3 to DDR3, we still roll back pretty far - a very close to current price was spotted in 2018, '17, 15, '13, and '11.
Will they..? It seems equally (or perhaps more) likely that we'll increasingly see vibe coded browser or Electron based applications as the bar is now lower to build such a thing.
yeah but you also have commercial licensing with Qt specifically :))
or we are going to see an explosion of vibe-coded GPL apps.
anyhow, the likes of Linear and Notion ain't gonna abandon web and go Qt. or!! if we are very lucky, we can see a native app framework that ticks all the boxes of a modern UI framework and is permissively licensed, but we need this crunch to stay there for years.
That doesn't apply so long as you are willing to accept the LGPL. In practice that means you can statically link everything except QT so that the end user is free to drop in a modified QT version if he would like.
If you are going to vibe code you can just pick any language you want. I had a go vibecoding in Rust and it worked perfectly fine. Even better than vibe coding in JS/Python because the type hints give the LLM a faster way to check progress.
Can you back that up with anything about semi-recent nodes? The voltages are so fragile that I'm not convinced you would actually save space once you adjust the design to handle more levels.
If it were possible, it would have been done already. The issue is the capacitors are already tiny, and barely can prevent a single bit decaying before refresh.
do you have a reference to exact / realistic scaling laws for the leakage currents as function of capacitor/dielectric dimensions and access transistor dimensions?
using 4 (or 2^N) voltage levels stores 2 (or N) bits, so we can afford to make the structures larger
why would this approach make sense for NAND flash but not DRAM?
You could also do a computing pr dollar graph - which would be a similar sharp decline over the past decades - however it won’t show anything like the memory price spike of the past few years.
I guess ‘per GB’ doesn’t really capture it, because the base number of gigabytes available to people (ie- the smallest compatible RAM kit you need to build a computer) and the base number of gigabytes you really need (OS bloat, feeling responsive, etc) have gone up so much.
A perfect example of how graphs are often misleading. $/GB is a totally useless unit value because it's an arbitrary size. The unit needs to be tied to the relative usefulness for its time. The y axis should be something like $/average workstation memory or $/requirement for common compute task. It's obvious that ram is expensive right now, but it's not expensive per GB. It's expensive relative to what you need to accomplish a useful task.
But relative usefulness is entirely subjective, making it a meaningless unit. Depending on your use case you may need 256 GB or 0.5 GB.
The audience who would benefit from hypothetical $/usefulness would be people who don’t know what memory is and don’t know what’s inside of their computers, or what it does. This is a fine audience to be in and to serve, but obviously not the audience of that website and not HN.
If you think that audience is under served for memory market statistics, I encourage you to make such a website and serve that audience.
For people on HN, who do you know what memory is, $/GB is a fine metric.
This is assuming that the wide variety of use cases are evenly distributed and that larger use cases are not mostly just a lot of duplicated smaller use cases. If I have a website I will need X amount of ram. If you run a much larger website offering a comparable service you will need some multiple of X, but you don't actually need much more ram per user (assuming you're also accounting for extra infrastructure and not just the web servers). It's the same task just scaled. Relative usefulness is not subjective, you could look at a variety of tasks in different industries. Windows server 2012 had a minimum requirement of 512 MB. Windows server 2025 has a minimum requirement of 2 GB. That's 4x for the same task which totally distorts $/GBs usefulness for being able to tell you anything helpful economically. It's obviously good to collect this data, but you need to pair it with some kind of demand data for it to actually tell you anything.
> you need to pair it with some kind of demand data for it to actually tell you anything.
Again, this is entirely dependant on who is consuming the statistic and for what purpose. For some use cases, yes demand data will be quite crucial. For others it will not. It's quite apparent the site's author doesn't see this as crucial and for the purposes I need to consider memory pricing, I agree.
I use a Thinkpad T530 for reasons that are very important to me. It is the only laptop that I have, so it is what I use for every manner of portable computing.
It still does all the things I want it to do, including using modern websites with modern browsers on modern operating systems (including Windows 11).
The 2012 computer running a modern linux install will still work fine. I'm talking more about the specs, specifically memory. I had 8gb of ram in my computer in 2012, the Macbook Neo released this year still has 8gb and is usable for modern day tasks.
We don't _need_ that much ram, we just found new things to do with more.
Gas is priced in $/gal, not dollars per mile or hour of lawn mowing or whatever. The resource and the use are completely different concepts and the resource owner/producer cares not of the buyers purpose for it.
This is extremely misleading and not very useful. It makes little sense to use pricing per GB during decades when RAM was at most in MBs. In that case, why not talk about price per TB or PB? Then the line will look pretty much flat and horizontal.
With respect... I am surprised such a low-quality analysis is published on stanford.edu . What is compared here? What is the purpose of this? What are the conclusions of the analysis? Heck, where is the analysis? By what logic are the prices per GB(!) comparable between 1960(!) and 2026?
I am sorry to being rude, I just don't understand this publishing beyond getting the media exposure.
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 I saw an IBM CE about to do an upgrade on a 370 system at Cal Berkeley. He had a box with several carefully-packed, large circuit boards. "So, is that a megabyte?" I asked. "Yup, that's a meg."
reply