Notes ·
RAM Prices May Be Showing Us the Shape of the AI Boom
Joshua MorrisPCPartPicker's memory price trends are remarkable because the change does not look like ordinary inflation. For much of 2024 and 2025, DDR5 pricing moved within a fairly predictable range, and then, beginning around the fall of 2025, the curves changed dramatically. Reporting based on that data shows 64GB DDR5-5600 and DDR5-6000 kits sitting a little above $200 for much of the preceding period before climbing to roughly $800 and $900 respectively by the end of 2025. A typical 32GB DDR5-6000 kit that averaged around $125 earlier in 2025 had already passed $250 by late that year. By June 2026, the cheapest 32GB DDR5 kits tracked in the United States were approaching $375.
The hypothesis worth testing is that this is not simply a temporary shortage of RAM — AI is changing what memory manufacturers choose to manufacture. High-bandwidth memory used by AI accelerators consumes substantially more manufacturing capacity than ordinary DDR5. Micron has said its HBM production has roughly a 3-to-1 trade ratio with DDR5, meaning producing a given amount of HBM consumes about three times the wafer capacity that could otherwise produce conventional DDR5, and Micron expects tight DRAM and NAND supply conditions to persist beyond 2026. TrendForce is seeing the same thing: the largest DRAM manufacturers are prioritizing server products, squeezing consumer memory supply, with AI-driven demand expected to grow faster than new supply through 2027.
That supports the hypothesis that AI infrastructure is indirectly making ordinary computers more expensive. It is not that a data center is buying the same DIMM someone would put in a gaming PC — the connection happens upstream. The same manufacturers, cleanrooms, wafers, and capital can produce different kinds of memory, and scarce capacity goes toward the products with the strongest demand and highest margins.
There are facts that keep me from concluding prices will simply continue upward forever. Consumer demand is already weakening. TrendForce describes the DRAM spot market as subdued, with buyers refusing to chase high prices. That looks like demand destruction — at some point people stop upgrading, buy less memory, or decide 32GB is good enough. DDR4 production has been winding down, so not every increase is AI. Retail charts can exaggerate shortages when inexpensive kits disappear from inventory. Canadian PCPartPicker prices also reflect Canadian retailers and currency. And enormous capacity is being built — Micron increased planned U.S. investment to more than $250 billion through 2035, while SK Hynix announced roughly $38 billion of expansion in South Korea. Fabs take years.
My guess is that we are somewhere in the middle of an unusual memory supercycle. If AI demand keeps outrunning capacity, consumer RAM could stay painfully expensive through 2027 and possibly longer. But another outcome is entirely plausible: hyperscalers slow spending, models become more memory-efficient, or new inference architectures need less expensive memory per user, all while new fabs come online. Suddenly the equation reverses.
DRAM has historically been an exceptionally cyclical commodity. Manufacturers expand during shortages, supply catches demand, inventories rise, and prices collapse. If the AI boom continues, elevated memory prices may become one of the hidden costs of the infrastructure behind it. If the boom slows before all this new capacity is absorbed, these nearly vertical price charts could look just as dramatic going the other way.
The evidence right now supports the shortage. It does not yet prove that today's prices are sustainable. That distinction may tell us a lot about where the broader AI infrastructure boom goes next.