DRAM Is Worth More Than Gold. The Supply Squeeze Is Only Getting Worse.

The Price of Memory Just Outpaced Gold

One kilogram of Korean DRAM can now buy 620 grams of gold. The comparison sounds like a provocation — DRAM is a manufactured chip, not a mined commodity. But the math is stark. Since January 2023, when the export unit price for DRAM (excluding modules) bottomed out at roughly $7,400 per kilogram, it has climbed to $92,183 per kilogram this year. That is a 12.5-fold increase. Over the same period, gold rose 2.4 times. DRAM has outperformed the precious metal by more than fivefold.

The same kilogram of DRAM now trades at 1.53 times the price of platinum and 41.7 times the price of silver. Prices hit a record in July, surging 270% year over year according to the Bank of Korea. Micron reported a similar trajectory, with its DRAM average selling price jumping more than 60% in a single quarter for fiscal 2026.

But the headline number — $92,000 per kilogram — obscures an important detail. DRAM is not a uniform good. A kilogram of high-density HBM packs far more value than a kilogram of consumer-grade DDR5. Product mix has shifted. The mainstream DRAM product TrendForce tracked in early 2023 was DDR4 8Gb (1 gigabit). Its current benchmark is DDR5 16Gb, which holds twice the data in roughly the same footprint. Capacity per chip has doubled. That generational upgrade accounts for part of the price increase, but even after adjusting for it, industry executives say the price slope is still dramatically steeper than fundamentals would predict.

The Real Problem: AI Is Eating Memory Capacity

Here is what most English-language readers will miss: the DRAM price surge is not driven primarily by a lack of factories. It is driven by a bottleneck in what those factories are allowed to make.

Goldman Sachs now projects the global DRAM supply deficit will widen from 5.0% this year to 5.9% next year — revising its earlier forecast upward by 3.4 percentage points for 2027. The bank expects the shortage to persist into 2028.

The culprit is HBM — high-bandwidth memory. This is the stacked DRAM that every AI accelerator needs. NVIDIA’s H100, H200, and the upcoming Rubin-class chips all depend on it. HBM is not made from the same wafers as standard DRAM in a straightforward way. It requires specialized packaging — multiple DRAM dies stacked vertically and bonded together — which consumes disproportionately more fabrication capacity per gigabit delivered.

TrendForce estimates that by the end of next year, HBM will claim 30% of total DRAM wafer input from Samsung, SK Hynix, and Micron combined. But on a bit-count basis, HBM will supply only 13% of total output. In other words, nearly a third of industry capacity is being absorbed by a memory type that delivers less than a sixth of the bits. The remainder — the HBM overflow — goes to standard DRAM used in PCs, smartphones, and general-purpose servers. That segment is where the squeeze shows up in everyday devices.

Goldman’s revision this week was aggressive. The bank had previously forecast a narrower gap. The upward adjustment reflects a reassessment of how fast AI accelerator demand is converting into HBM wafer consumption, and how slowly new DRAM capacity is coming online.

The Timeline Problem

Memory makers are building new fabs. Samsung is expanding its Xi’an facility. SK Hynix is investing heavily in its Icheon and Giheung campuses. Micron is bringing online its Yokohama plant. But each new line takes 18 to 24 months to reach meaningful volume, and HBM qualification adds further delay. The current surplus of production capacity simply has not arrived.

Micron’s own guidance, delivered in June, projected that the tight supply environment would last through 2027. SK Hynix president Kwak Noh-jung stated the same week that 2027 could be the most severe shortage year in memory industry history. These are not cautious forecasts. They are warnings from companies that know their own capacity plans best.

Who Loses When AI Monopolizes Memory

The implications extend well beyond server racks.

Console manufacturers are already feeling the pressure. Sony and Nintendo both use DRAM in their systems, and the cost of every unit is rising. A console maker buying millions of chips at current prices cannot easily absorb the markup. Margins will compress, or prices will rise — or both. The next generation of gaming hardware may arrive more expensive than expected.

PC makers face the same dynamic. DDR5 prices have climbed so sharply that some OEMs are reconsidering upgrade timelines. A segment of the market that was supposed to shift entirely to DDR5 is now pausing, waiting for prices to stabilize. That demand suppression could eventually help, but it will not happen quickly enough to ease the next 12 months.

The most consequential development, however, is happening inside NVIDIA’s engineering teams. Industry sources tell Pulse that the company and several large cloud service providers are now reviewing how much HBM their next-generation AI accelerators will carry. The question is not just performance — it is whether a chip designed for 192 gigabytes of HBM can be re-engineered for 96 gigabytes without destroying its competitive position. If NVIDIA cuts HBM configurations, it would free up a meaningful slice of memory capacity for the rest of the industry.

That decision has not been made. It may not be made this year. But the fact that it is being discussed signals how serious the supply crunch has become. Even the dominant AI chipmaker is feeling the constraint.

What to Watch Next

Three developments will determine whether the shortage breaks or deepens.

First, when Samsung and SK Hynix bring their newest HBM production lines to full yield. HBM3E qualification rates remain a key variable. Any delay pushes the timeline further out.

Second, whether NVIDIA and the CSPs actually reduce HBM per chip. Even a partial reduction would release wafer capacity that could ease standard DRAM tightness within six to nine months.

Third, how demand itself responds to price. If server buyers continue ordering at inflated prices, the feedback loop that normally moderates memory markets — demand destruction — will not trigger. Prices stay high. Shortages persist.

The gold comparison is a useful shorthand for understanding the scale of the shift. But the real story is structural: AI has created a memory demand profile that the industry’s existing capacity model cannot serve without sacrificing something else. Right now, that something is everything that is not an AI accelerator.