Rampage: How Ai Is Eating The World's Memory

Hi Professor, Thanks for the detailed feedback. I went through every inline comment and the rubric notes and tried to address them in the revision. A lot of it was straightforward stuff like defining H.B.M, D.D.R 5, and Gartner on first mention, swapping"by the truckload" for actual numbers (two to four terabytes of dram per A.I server, around 70% of high-end dram by 2026), and explaining what clean rooms and"tight" supply actually mean. A few changes ended up being bigger than I expected, though.

The thesis is the one I rewrote the most. My original version basically summarized the problem and then asked for regulation, which felt more like an overview than an argument once I went back to it. The new thesis tries to claim something more contestable: that we've been treating dram like a consumer commodity when it's already functioning as infrastructure, and that the market is structurally incentivized to keep the shortage going rather than just failing to fix it. The conclusion got rewritten to echo that instead of being a generic call to action.

In the demand paragraph I also added some reasoning I think helped the argument. The original draft assumed A.I models need a lot of memory because they're inherently memory-greedy, which I realized was hiding an assumption I hadn't checked. I added two specific techniques that disprove it: knowledge distillation (training a smaller model to mimic a bigger one) and quantization (storing weights in 8 bits instead of 32). Then I named the real reason companies don't bother with them, which is that hardware is cheaper than engineering hours.

The counterargument paragraph is the change I'm happiest with overall. In the first draft it was basically one dismissive sentence, so I rebuilt it to steelman the "markets will self-correct" position before showing why that doesn't hold up on the timeline that matters. The Jevons paradox piece in there is the part I'd defend hardest, since it answers the obvious "but what if A.I just gets more efficient" objection in a way that surprised me as I was working through it.

I also flipped a couple of sections around. "The Scarcity Is Manufactured" now goes demand to supply, so the section title lands as a payoff instead of being asserted upfront. "Who Gets Screwed?" opens with the I.D.C and Gartner numbers before the moral framing, which makes the moral part feel earned. Both read better that way.

Thanks again for pushing me on this. The comments hit a lot of the spots in the argument I'd been getting away with not fully thinking through.

Wony [Revised version begins next page]

Wony Nguyen

Audio by Paper2Audio.

When the ram Market Rammed.

If you're a photography or videography geek, you've probably heard of the little piece of technology inside your computer, so-called ram, that limits what you can do on your tiny laptop. If you're a gamer, you know all too well what ram is and how important it is for any of your wins. It may sound really tech-savvy until you realize that the cool Stanley cup you bought on sale for $30 can only exist because a sophisticated manufacturing system runs on ram. Think of it as your computer's short-term memory, so the more you have, the more tasks it can juggle at once. And it's not just inside your laptop: the factory robots, medical scanners, and assembly lines that build the stuff you use every day all run on it too.

Here's where things get personal. A.I operates on massive data centers that can only function with a gigantic amount of this same memory, specifically dram, the most common type of memory chip and the same kind found in P.C's, industrial equipment, and the servers running hospitals and supply chains. When A.I companies buy it up by the truckload, what's left for everyone else shrinks.

A single A.I server now ships with two to four terabytes of dram, up from 128 to 256 gigabytes in a traditional enterprise server, and A.I and server workloads are projected to consume roughly 70 percent of all high-end dram by 2026. The deeper problem isn't the shortage itself, it's that we've been classifying dram wrong. As a consumer commodity, it goes to the highest bidder, which is A.I. As what it has actually become, critical infrastructure on par with water and electricity, the picture changes: when one industry can corner the supply of something essential, it isn't competing in a market, it's privatizing infrastructure the rest of the economy depends on. The market won't fix this, because the chipmakers profiting from the shortage have every reason to keep it going. dram needs to be reclassified as a strategic public good and regulated like any utility, before the physical economy keeps subsidizing Silicon Valley's refusal to write efficient code.

The Scarcity Is Manufactured.

Start with the demand side. Industry analysts at Ram Exchange and TrendForce, two firms that track the memory market, report that A.I data centers now consume roughly 70% of all high-end memory chips produced globally, a complete inversion from just a few years earlier, when most of that output went to P.C's and smartphones. The per-server jump is what's driving it, and it's enormous, an order of magnitude more memory than a traditional enterprise server needed. The industry frames this as inherent, calling large language models"memory-greedy," but the assumption that bigger models require proportionally more memory doesn't hold up. Knowledge distillation, where a smaller"student" model is trained to mimic a larger"teacher," can shrink a model by roughly 40% while keeping most of its performance, as Distilburt showed for language models.

Quantization stores model weights using 8 bits instead of 32, a four-fold reduction in memory use with little accuracy loss. Companies aren't avoiding these techniques because they don't work, they're avoiding them because hardware is cheaper than the engineering hours to optimize. So "just scale it" has become the industry default not because it has to be, but because nobody is forcing the industry to choose otherwise.

You'd expect a demand surge this big to trigger a supply response, but the dram industry isn't built to flex. Fabrication requires billion-dollar clean rooms, ultra-sterile facilities where even a single speck of dust can ruin a chip with nanoscale features. According to Intel, building a new fab takes three to five years and roughly ten billion dollars, and a typical fab contains over a thousand multimillion-dollar tools. That's only half the story.

Network World adds the second half: only three companies, Samsung, S.K Hynix, and Micron, control nearly all of global dram production, and all three have chosen to redirect capacity away from the standard dram that everyday devices depend on toward the high-margin memory A.I data centers demand. The supply chain has been structurally tight for years, and the same three companies that built the bottleneck are now deciding who gets served first. That's what"manufactured scarcity" means: dram isn't naturally rare, a handful of companies have chosen who gets access to it.

Who Gets Screwed?

The consequences are already showing up in everyone's wallet. I.D.C, one of the largest technology market research firms, now projects P.C shipments to fall 11.3% in 2026 and smartphone shipments to drop nearly 13%, worse than even their most pessimistic earlier forecasts. Network World documents the same squeeze hitting buyers directly: Samsung raised the price of its 32 gigabytes Double Data Rate 5 (D.D.R 5) dram modules from $149 to $239, D.D.R 5 contract pricing has more than doubled since early 2025, and Gartner, a leading technology research firm, projects dram prices to climb another 47% in 2026 alone. I saw this firsthand last semester when I tried to upgrade my own computer setup, and the same kit I had been eyeing for months had jumped way past what I'd budgeted for. When manufacturing can't get dram, production stalls, costs rise, and those costs land squarely on consumers. The shortage is stopping us from making stuff, not just writing code.

And consider what people are paying for. A.I does have real applications in healthcare and research, but the bulk of this memory isn't going toward curing diseases. Most of it goes toward what is, hype aside, a luxury good: a marginally better chatbot, a slightly sharper image generator.

Meanwhile, the same dram runs medical scanners that catch tumors, robotic arms that weld car frames, and the embedded chip inside your washing machine. One side is chasing profit margins, the other is keeping the physical world running, and the luxury side is winning by a landslide.

Zoom out, and the pattern looks familiar. dram is becoming the invisible currency of the global economy, rippling through manufacturing the way oil ripples through entire nations, and almost nobody outside the tech industry is talking about it. So if the market created this problem, can the market fix it?

Who Is to Blame?

The strongest counterargument is that markets self-correct: higher dram prices should push chipmakers to expand production, A.I companies should eventually optimize for memory, and the value A.I creates should offset costs elsewhere. There's real evidence for the optimistic case. Samsung, S.K Hynix, and Micron have all announced new fabs, and the Chips Act, a 2022 federal law providing about $52 billion for U.S. semiconductor manufacturing, is accelerating construction.

In a normal market, that would be the end of the story. But the dram market has three structural features that block self-correction. First, the lead-time mismatch: fabs take three to five years to build, while A.I workloads scale on a timeline of months, so demand multiplies again before new supply arrives.

Samsung's newest line won't begin mass production until 2028. Second, margin asymmetry: High Bandwidth Memory (H.B.M), a specialized type of dram designed for A.I accelerators, commands far higher margins than conventional dram, giving chipmakers a permanent reason to prioritize data center clients. Third, the Jevons paradox: when a resource becomes more efficient to use, total consumption tends to rise rather than fall.

Leaner A.I models just free up budget for bigger ones, so efficiency gains expand demand instead of contracting it. The result is a market that looks like it's working but keeps squeezing the rest of the economy anyway. The factory that can't source parts now doesn't benefit from a correction in 2028.

So regulation has to step in. Treating dram as a "public good" doesn't mean nationalizing Samsung; it means demanding the same transparency and accountability governments already require of other shared resources. Water is the obvious example: governments can stop one company from draining an aquifer because everyone downstream depends on it. Utility companies aren't allowed to charge whatever they want when the lights must stay on for everyone. The F.C.C auctions radio spectrum instead of letting one broadcaster squat on every frequency.

dram deserves the same framework. A.I companies should be required to disclose how much dram they consume and show real efforts toward efficiency, and federally funded A.I projects should be required to adopt the memory-saving techniques described earlier. That would push companies to innovate instead of throwing more hardware at the problem.

Change the Paradigm!

The dram crisis isn't a shortage the market accidentally created and will eventually fix. It's a structural failure the market is incentivized to maintain, and it persists because we're still treating dram as a consumer product when it has quietly become 21st-century infrastructure. Every price hike at the checkout counter is the rest of the economy paying for that misclassification. Engineers can shoulder part of the fix by building leaner models, but the bigger move belongs to policymakers: stop letting one industry privatize a shared utility. The next time your laptop costs more than you expected, or a gadget you wanted is mysteriously out of stock, remember that a single A.I server now ships with up to four terabytes of dram, and data centers run thousands of them. That supply has to come from somewhere, and right now it's coming out of everyone else's pocket.

Works Cited

You have reached the end of the document.