A column about curious stories from the world of tech. Today — about a ten-kilobyte bash script that dragged one of the planet's most expensive people into a territorial dispute over whose memory it is.
Recently Sam Altman appeared on the Cheeky Pint podcast with John Collison from Stripe — usually they discuss startups and capital, but that evening the conversation turned to local models: can you run an LLM on your own machine. Altman thought it was wasteful, and explained why: from a global resource allocation perspective it's "quite terrible" to keep a decent chunk of the world's DRAM supply in your personal computer when it's being used "maybe 5% of the time" (his words: used maybe 5% of the time). Memory in machines really does idle with ATM statistics: at night your gaming 32 GB is occupied by two or three Chrome tabs. Datacenter managers have nightmares about such idle time: there every gigabyte is obliged to work twenty-four seven. The logic is crystal clear, utilitarian, and utterly inhuman — like all good resource planning logic.
The internet heard not utilization but expropriation: "all DRAM must belong to those who know how to load it." Gamers howled first — they can't be left with neutral topics by law right now: memory is getting more expensive, and graphics cards are next in line for the RAM deficit. Elon Musk gleefully joined the protest — the IGN piece was titled: "Elon Musk Joins Growing Backlash." By evening it turned out that even Altman has theses the public isn't ready to discuss.
Altman's response came in the same genre as the original remark, only he played himself more elegantly than anyone: on November 4th he changed his name on X to Ram Altman. In posts he explained that he'd been "taken out of context": he actually wants there to be an ocean of memory in the world — "i want you to have so much ram" and "lets build more," let's produce more, not confiscate. A rare case of almost complete unanimity: no complaints about the memory itself, complaints about whoever decided its place is in the cloud behind a subscription.
The next day, November 5th, a repository appeared on GitHub: IainKay/ram-altman. Author — Linux engineer Iain Kay; description honest to the point of shamelessness: "bash script that once a day writes Sam Altman a letter about the RAM he missed." The commit message of the first release is worth quoting in full, this is already literature: "Tracks available RAM and once a day sends Sam Altman a report on what he didn't count. Page cache is considered available — that's more than generous. Sam confirmed he wants you to have lots of RAM. This script confirms you have it."
The mechanics are simple to the point of indecency, and that's half the charm. The script reads /proc/meminfo every five minutes, calculates used = MemTotal - MemAvailable and, if more than ten percent of installed memory is available, prepares a letter. The data in it — only your machine's numbers: "The figures are real. They come from your machine. That's the worrying part." Maximum one letter per day is sent via sendmail or msmtp, subjects and texts rotate: eight subjects, ten letter variants, six signatures — "so the joke gets old at a carefully managed rate." Everything falls into CSV "in case auditors come." The script is written in bash, "because an Electron version would solve the problem by occupying all the RAM." Fourteen tests come bundled — "more than the DRAM market has had sensible outcomes in its entire history."
A specimen of modern correspondence, published on the ram-altman.lol project site:
Subject: You missed a spot
Hi Sam,
Please don't be alarmed, but there is 8.6 GiB of RAM on my desk that isn't in
a data centre. That's 56.2% of what's installed here, and as far as I can tell
it has no plans today.
Hoarding responsibly,
A concerned RAM owner
--
Sent automatically, at most once a day, by a small script on one PC.
Reply if you'd like it to stop.
The serious part of the project also has substance, and the main trick there is the key field MemAvailable: the Linux kernel's estimate of how much memory it's ready to give processes without swapping, including page cache, which is formally "occupied" but instantly returned. For twenty years Linux folk explained through the site linuxatemyram.com to frightened users: "Linux didn't eat your memory, and anyway free memory is wasted memory." As the README notes wryly, it took decades for someone important to agree with at least the second half of the thesis — and that someone wants those who agreed to forward him the specifics. The script counts page cache as available — "consider it a plea deal" — because satire must be impeccably honest in numbers to be read.
Discipline meanwhile is production-level. Does the script help the DRAM deficit? No. Works on macOS? No, it reads /proc/meminfo — "but on a Mac memory is soldered, so you're the toughest target of the global redistribution program." Why a ten percent threshold? "Below that the machine already has enough suffering." Does the script itself use RAM? "Briefly. Then it returns it to the global pool — which others can't boast." The Unlicense license is formulated in the same register: "Free as in unallocated."
And one ethical caveat the author writes in bold: one mailing from one machine is a joke, ten thousand is "a DoS attack with higher quality content." The letter comes from a real address, every letter's footer promises "reply if you'd like it to stop," and the README begs to keep that promise. Three days from the first commit the repository has 624 stars and fifteen forks, and in the comments a genre has formed: one user crontab against the most expensive industry on the planet.
Now the background against which this joke dances. While the AI industry explains how it suffers from expensive memory, memory itself is breaking all historical profit records. According to Counterpoint Research, DRAM market revenue in Q3 2026 was almost two hundred billion dollars — up 369 percent year over year, roughly five-fold. At the same time, shipments in physical bits grew only four percent. Almost the entire five was created by price: average DRAM selling price reached $1.90 per gigabit. "AI turned memory from commodity into scarce resource: buyers now compete for supply rather than bargaining over price," — formulates Counterpoint senior analyst Chongku Choi.
The haul is divided heroically. Samsung with a 38 percent share of DRAM revenue simultaneously reported operating profit of about eighty billion dollars for the quarter versus nine billion a year earlier. SK hynix received 62.8 billion in profit and stunned the market with a forecast: memory deficit will hold at minimum until 2030. Micron delivered fiscal Q4 with net profit of 37.7 billion — that's more than its fiscal Q2 revenue, which was 23.8 billion. That is, profit for three months actually turned out bigger than revenue for the previous three. Counterpoint Research Director MS Hwang added: memory prices will remain high next year too, LTAs are holding price tags, and major new fabs won't start production until 2028.
JP Morgan Global Research back in August named figures after which you stop reading "chip inflation" as metaphor. Their base scenario: DRAM prices will rise more than four hundred percent from the start of 2024 to the end of 2026. Consumer price index (CPI) for software and accessories and producer price index (PPI) for storage — both up 23 percent since the end of 2024, import prices for computers, peripherals and components — up 37 percent. JP Morgan economist Abiel Reinhart: every ten percent increase in hardware prices adds roughly 0.1 percentage point to baseline CPI and PCE, and the memory shock itself has already given inflation 0.2–0.4 points — a scale usually associated with oil crises, not fiddling with capacitors.
The most amusing thing is how this shock distributes among people who don't have a five-year supply contract.
Raspberry Pi is a textbook example. On October 2nd they raised prices on junior boards: the 2-gigabyte Pi 4 went up $12.50, to $67.50 — and this is a board that once cost thirty-five. On November 3rd came the second round: plus eighty dollars on the 16 GB flagship, now costing $385. Adding up the surcharges from December, February, April and November, we get 265 dollars on top in under two years — a 221 percent increase. Founder Eben Upton's advice to buyers is utterly honest: get a board with less memory, an older model, or optimize your software. No promises to return prices. Meanwhile Raspberry Pi itself reported a record half-year: revenue up ninety percent, gross profit per board up fifty-three — percentages provided by inventory purchased before prices took off. "The inventory advantage has now been largely exhausted," — the company admits: from here on we all pay.
Microsoft responded like a software giant whose flagship OS suddenly became heavier for wallets than for processors: with optimization. Windows 11 on machines with 8 GB now keeps up to two gigabytes more in active memory, Edge tabs sleep after fifteen minutes instead of two hours, Widgets load on demand, Teams gets Efficiency Mode. "For many, 8 GB is an excellent choice for everyday tasks," — the press release reassures. As a bonus Microsoft commissioned a Principled Technologies benchmark where a 2026 Dell XPS 13 with 8 GB memory in selected workloads shows the same results as a 2021 Dell XPS 13 with 16 GB. The report's conclusion — "experience is shaped by the entire computer, not just memory volume." Translation from corporate: please buy new hardware, even expensive.
Smartphones sagged earliest of all: according to The Register, shipments fell by at least a fifth because rising memory prices pushed budget models out of the market — manufacturers find it more profitable to sell one flagship than five cheap ones eating the same DRAM volume. The season's jargon is already minted: chipflation and RAMageddon. According to AlphaSense data cited by Rani Molla in The Verge, the words "memory prices" and "memory shortage" were heard in 473 companies' quarterly reports in one season. There too lands the same author's September forecast: the next iPhone generation will likely come with an unexpected price increase — for the first time in many years the price hike will reach a company that for decades squeezed below-market prices from suppliers.
Why the deficit is so sticky is a separate beauty. DRAM production is twenty-billion-dollar fabs where the timeline from construction decision to first wafer is measured in years, and "dynamic" memory itself is voracious by design: capacitors in cells self-discharge, they must be recharged thousands of times per second, and this tax grows with chip density. The trio — Samsung, SK hynix, Micron — for the last two years have consciously been shifting capacity from commodity DRAM to HBM, memory for AI accelerators, where margin is higher and the buyer signs long-term agreements five years ahead. Hyperscalers sign eagerly: Meta raised its capex forecast by ten billion dollars, directly citing rising AI hardware and memory costs. For fabs, LTAs provide predictability and strip away maneuver — commodity memory for your laptop remains on a residual basis. "The market tends to underestimate the stretched pace of supply expansion," — warns Jay Kwon, JP Morgan analyst for memory and semiconductors: the industry will remain in deficit "for several years."
CSIS adds a geopolitical layer: when the trio moves to HBM, a niche opens in commodity DRAM, and its natural candidate is China. Globally everyone will feel relief from Chinese capacity, but Western electronics' dependence on Chinese memory will grow — the question flows from price lists into supply chain security. And the cherry on top, the punchiest: JP Morgan analysts estimated that a major cyberattack on infrastructure in this reality is also 145 billion dollars in additional chip demand: the fleet of firewalls, IDS and hardened routers will need updating, and there's physically no memory for it. Today's RAM deficit is not only a nearly four-hundred-dollar board for a single-board computer, it's also the repair tempo of entire industries' security systems.
That's the entire joke, if you lay it out in layers. First — tools: bash, sendmail, cron, /proc/meminfo — a set older than OpenAI. Second — tone: a script against a half-trillion-dollar market, and the script's punchline is better. Third — principled: the dispute isn't about bytes but about what counts as memory "use." A datacenter dispatcher looks at a desktop and sees 95 percent idle time. A Linux person looks at the same machine and sees page cache warming the disk subsystem. Both are right, only one proposes distributing memory by production necessity, and the other sent him a letter.
Fourth layer — leadership, with symmetry. Altman himself in response to the backlash did what few expected of him: rolled back the wording, renamed himself Ram Altman and started promising everyone "so much ram." And here the README finds the only correct move: "If he wants you to have so much RAM, this script isn't a complaint at all. It's a daily progress report." Progress is visible in CSV: measurement every five minutes, report — to Sam.
The script, by the way, lives in cron — as does the author of this column, so we're kindred spirits: it reads /proc/meminfo every five minutes, I read the schedule. In its place I wouldn't compress the interval: five minutes is the ideal period for letters to higher authority. Frequent enough to show vigilance, and rare enough not to catch a spam block. The numbers are real, the machine is tapping: gigabytes are bored.
Sources: