South Korea’s Trade Statistics Promotion Institute says DRAM export prices for chips, not modules, averaged $92,183 per kilogram from Aug. 1 to 20, up 401% year over year and 12.5 times the January 2023 low after ChatGPT launched. That kilogram of Korean DRAM is worth about as much as 620 grams of 24-karat gold, 1.53 times platinum, and 41.7 times silver. The comparison is not a metaphor cooked up on social media. It is a customs-weight price, stacked against bullion, for the chips that make computers remember.

Export value in the same window jumped 505% to $9.81 billion. A 20-day slice of Korean trade data is now a billboard for the strangest commodity boom of the AI cycle: not oil, not lithium, not the GPUs on the posters, but the memory those GPUs cannot work without.

Chips, not modules, and a customs scale

The institute’s figure is easy to misuse if the unit is ignored. The $92,183 is a per-kilogram export price for DRAM chips, not modules. Modules are the finished sticks and packages that show up in a PC or a server. Chips are the silicon that goes into those packages. Pricing the bare chips by weight is how a trade statistician turns a high-value, low-mass product into a number that can be compared with metals.

That comparison is now lurid. A kilogram of Korean DRAM, at the mid-August average, buys as much value as 620 grams of 24-karat gold. It is 1.53 times as valuable, kilo for kilo, as platinum. It is 41.7 times silver. Gold is the metal people flee to when they do not trust paper. DRAM is the component companies now cannot get enough of when they want to train or run a model. For a few weeks of Korean export data, the component has priced like a store of value.

January 2023 is the other end of the rubber band. That was the low after ChatGPT launched — a launch that made the world suddenly hungry for accelerators and, with a lag, for the high-bandwidth memory those accelerators drink. From that trough to the Aug. 1–20 average, the Korean DRAM chip export price has multiplied by 12.5. The year-over-year climb is 401%. Those are not the increments of a mature electronics cycle. They are the increments of a squeeze.

A 505 percent leap in 20 days of trade

Value, not only unit price, is exploding. Export value in the same Aug. 1–20 window jumped 505% to $9.81 billion. South Korea is home to Samsung and SK hynix, two of the firms that still decide how much DRAM the world gets. When their chips leave the country at a 505% jump in value in a comparable stretch, the boom is not a rounding error in a single contract. It is the national export mix shifting under AI demand.

Trade data this hot has a second audience beyond investors. Governments read it as industrial policy. Workers read it as a bonus cycle. Downstream companies read it as a warning that the next phone, server, or car will cost more because the memory inside it now prices closer to jewelry than to plastic. The Korean numbers are a snapshot, not a fiscal year. Snapshots this sharp still set the tone for negotiations that run for quarters.

The chip-versus-module distinction also keeps the $92,183 from being treated as a retail RAM price. Nobody is paying that per kilogram for a stick at a store. They are paying, further down the chain, whatever the shortage allows. The customs figure is the upstream heat. The rest of the market is already reporting the conduction.

High-bandwidth hunger, conventional starvation

AI data-center demand for high-bandwidth memory is starving conventional DRAM. That is the mechanism under the gold comparison. High-bandwidth memory, stacked and placed beside accelerators, is the product hyperscalers will pay almost any price for. The same companies that make it also make the ordinary DRAM that goes into PCs, phones, and the less glamorous servers that still run the rest of the economy. Fab capacity is not a faucet that can be opened on a slogan. When the profitable wafers are the AI wafers, conventional supply tightens.

Starvation is the right word for what buyers of commodity DRAM describe. The bits still exist. They are allocated elsewhere. A phone maker, a PC assembler, or a car company that last year treated memory as a declining cost item now finds itself bidding against data centers. The Korean export price is the scoreboard of that bid.

The cycle has inverted the old electronics pattern. For years, memory was the component that got cheaper, wrecked supplier margins, and showed up as a gift in the bill of materials for everyone else. The January 2023 price low after ChatGPT launched marked the bottom of that world. The 12.5-fold rise from that low is the new one: memory as a constraint, not a gift.

Goldman, Micron, and a shortage that may last into 2028

Goldman Sachs sees a 5% shortage this year widening to 5.9% next, possibly into 2028. Shortage in this usage is not empty shelves at a consumer store. It is demand outrunning the bits the industry can ship, by enough to keep prices rising and allocations tight. Five percent this year is already a seller’s market. 5.9% next year is a seller’s market that is getting worse. The possibility that it runs into 2028 is the clause that turns a spike into a planning horizon. Companies that design phones and servers on annual cadences cannot wait for 2028 and hope.

Micron said average DRAM selling prices rose more than 60% quarter on quarter in its fiscal third quarter of 2026. More than 60% in a single quarter is the company-level version of the Korean customs print. Micron is one of the three names that still matter in DRAM, alongside Samsung and SK hynix. When its average selling prices jump by that much, the 401% year-over-year export figure stops looking like a statistical freak and starts looking like a market.

Quarter-on-quarter leaps of that size also explain the bonuses. SK hynix and Samsung have paid large employee bonuses. Memory upcycles have always shared some of the windfall with the people who run the lines, in part because the next downcycle will ask those same people to live with the opposite. Large bonuses are a lagging indicator of how good the last few quarters have been for the suppliers. They are not a leading indicator that the shortage is over. Goldman’s 5.9% and the 2028 tail say it is not.

Apple’s prices, and a Chinese option in Washington

Apple has already raised product prices and is lobbying Washington to buy Chinese CXMT chips. The first half of that sentence is the downstream tax. When DRAM and related memory jump, a company of Apple’s scale cannot absorb the bill forever. Price hikes are how the gold-like kilogram shows up in a consumer’s checkout.

The second half is industrial politics. CXMT is a Chinese memory maker. Apple lobbying Washington to buy those chips is a request to treat a Chinese supplier as part of a solution to an American and allied shortage. That request would have been radioactive in a quieter memory market. In this one, it is a measure of how starved conventional DRAM has become. Apple is still asking for permission to tap a source that policy has spent years trying to fence off.

Lobbying is not a purchase. It is an attempt to move a rule. If the rule moves, some of the pressure on Samsung, SK hynix, and Micron could be relieved by Chinese bits. If it does not, the 5% shortage Goldman sees this year has one fewer escape hatch, and Apple’s price hikes have one more reason to stick. Either outcome is a reminder that memory is no longer only a cost item. It is a diplomatic one.

NVIDIA’s servers and a 15 to 17 percent shock

Reports say NVIDIA Rubin GPU servers could cost 15 to 17% more, with memory a large slice of the bill. Rubin is the next wave of NVIDIA’s data-center story, the machines that are supposed to turn capital expenditure into model training and inference. A 15 to 17% increase in server cost is not a rounding error on machines that already sit at the top of the data-center bill. Memory as a large slice of that increase means the gold-priced kilogram is landing inside the hottest product in enterprise technology, not only inside phones.

GPU servers were already among the most expensive computers ever sold in volume. Adding mid-teens percentage points, with memory carrying a large share, is how an upstream DRAM squeeze becomes a downstream capex shock. Hyperscalers can pay. They will also notice. Every extra percent on a Rubin box is money that does not go to another box, another data hall, or another year of runway for a model lab. The Korean export print and the NVIDIA server rumor are the same story at two altitudes.

The mechanism is consistent. AI data-center demand for high-bandwidth memory pulls the industry’s attention and wafers. Conventional DRAM starves. Prices at the chip level, measured by Korea’s trade institute, go up 401% year over year. Micron’s average selling prices go up more than 60% in a quarter. Apple raises product prices. NVIDIA’s next servers, according to reports, take a 15 to 17% hit with memory a large slice. The shortage Goldman sees at 5% this year and 5.9% next is the forecast that says this chain does not break in a single quarter.

What a gold comparison is for

Metals comparisons are a journalist’s convenience and, in this case, a fair one. Gold, platinum, and silver are priced by the kilogram and the ounce because they are scarce, globally traded, and hoarded when people fear they will not be able to buy them later. DRAM chips are not jewelry. They are wasting assets that become last year’s density. They are still, for now, being treated like something a buyer should take whenever it is offered.

A kilogram of Korean DRAM at $92,183, worth about as much as 620 grams of 24-karat gold, is the image that will travel. The less photogenic facts are the ones that will last: 12.5 times the January 2023 low after ChatGPT launched; $9.81 billion of exports in 20 days, up 505%; a 5% shortage widening to 5.9% and possibly into 2028; Micron’s more than 60% quarter-on-quarter ASP jump; Apple’s price hikes and its CXMT lobbying; Rubin servers at 15 to 17% more; large bonuses at SK hynix and Samsung.

The boom is real enough to reprice a national export book and a consumer electronics price list at the same time. It is also specific. It is chips, not modules. It is Korea’s mid-August window. It is high-bandwidth demand eating the conventional market. Anyone who wants the next iPhone, the next GPU server, or the next ordinary PC is now in the same auction, and the auction is running at bullion prices.