The tide data this month is hard to look away from. In early August, Korea’s DRAM exports were running at a unit price of $92,183 per kilogram — up 401 percent year on year, according to Korea’s trade statistics agency, measured over August 1–20. Let me put that number in terms that land: with spot gold near $4,620 per ounce, one kilogram of DRAM is now worth the equivalent of roughly 620 grams of gold. Memory is being priced by weight now, and the weight is goldweight.
We should worry — but measure first. This is not a headline to be alarmed at; it is a measurement to be understood. And when I measure it carefully, the story underneath is sobering in a different way than a price spike usually is.
The per-gram price is the signal
Let me check the arithmetic on that number, because it is the kind of figure worth verifying before it goes any further. Korea’s trade statistics agency reported DRAM export unit prices at $92,183 per kilogram for the first twenty days of August. Against spot gold at about $4,620 an ounce, a kilo of gold is worth roughly $148,600 — and one kilo of DRAM at $92,183 is the value of about 620 grams of gold. The claim in the headline is literal, not rhetorical.
I started writing this piece from the demand angle — the idea that AI servers are hoovering up every stick of memory they can find — and I dropped it halfway through. Demand is part of the picture, but demand does not explain a per-gram price. Gold is not priced by demand; it is priced by scarcity, by the slow trickle of supply. When a semiconductor gets priced like gold, you are looking at supply, not appetite.
A supply wave, not a demand tide
Here is where the data gets precise. The normal delivery window for DDR5 memory is about six weeks; by August, lead times had stretched to fifty weeks, according to a supply-chain data platform. That is a 733 percent extension of the wait. A lead time is the clearest single measurement of a supply crunch there is: it tells you how long the wave takes to arrive, and fifty weeks is nearly a year.
The same platform tracked server-grade 32GB DDR5 contract prices through August — from $730 to $900 in a single month, up 23.29 percent. Contract prices move more slowly than spot prices and they reflect negotiated, committed volume, so a 23 percent monthly move in the contract market is not noise. It is the tide line rising on the shore; the spot market moved first, and the contract market is following.
Then came the day the whole channel noticed. On the morning of August 17, Samsung Electronics — the largest memory maker in the world — pushed a price notice to its global distribution channels: 32GB DDR5 spot prices moved from $149 to $239, a jump of about 60 percent, with the 16GB and 64GB parts raised 30 to 50 percent as well. Sixty percent in one notice is not a gradual adjustment; it is a wave arriving. The distribution channel that had been quoting old numbers woke up to new ones.
What is crowding the water
The reason this matters — and the reason it is sobering rather than merely interesting — is that the memory wave is not an isolated event. The AI build-out that drove graphics processors into shortages is now crowding the same fabrication capacity and the same supply chain. Memory was the second thing to run scarce after computing, and the scarcity is compounding because capacity does not flex quickly. A fab line that is retooled for one product does not swing back to another in a quarter.
I have watched this kind of wave before — memory has always been cyclical, and I have said so in earlier pieces. But the current one is different in a way the numbers make visible. In past cycles, price spikes were driven by demand surges that corrected when demand normalized. This time the lead times and the per-gram prices are both pointing to the supply side: there is simply not enough capacity scheduled to come online to meet what is already contracted. The wave does not have a natural breakwater yet.
No, that is not quite right. There is a breakwater — it is just years away. New memory fabs take two to three years from groundbreaking to volume production, and the current wave of announcements will not become real capacity before the middle of the decade at the earliest. Until then, the market has to ration by price, and rationing by price is exactly what the $92,183 per kilogram number is.
The reprice ripples downstream
What follows from a per-gram price like this is a repricing of everything that contains memory, and the list is long. Servers, laptops, phones, automotive electronics, networking gear — every product with a DRAM socket has a memory cost that just moved. When a component’s price rises 60 percent in a notice, the finished good either absorbs it or passes it on, and the pass-through is what consumers eventually meet at the checkout.
The consumer-electronic cost story is the part of this I find easiest to understate and most important to state plainly. A supercycle priced in grams of gold does not stay inside the semiconductor industry; it flows downstream as quietly as a current. Device makers who cannot renegotiate contracts will price the memory into the product, and the product price is where the wave becomes visible to ordinary buyers who have never heard of DDR5.
What the evidence says so far
Let me put the evidence in order, the way a steady journal should. First, the per-gram price: $92,183 per kilogram, up 401 percent year on year, worth the equivalent of 620 grams of gold per kilo. Second, the lead time: six weeks stretched to fifty. Third, the contract market: server 32GB DDR5 up 23.29 percent in August alone. Fourth, the spot notice: 60 percent on the flagship part in a single day, with the rest of the line up 30 to 50 percent. Every one of those numbers says the same thing: supply is the constraint, and the constraint is binding now, not next year.
The data shows the trend; alarm adds nothing to the evidence. What the evidence requires is not panic but planning. If you buy memory, the price is what it is; the only useful question is how long the wave lasts. If you sell products that contain memory, the only useful question is what your contract position is and when it reprices. The one thing nobody should do is assume the old six-week lead time and the old price are coming back soon, because every measurement in the file says otherwise.
Who rides the wave, who pays for it
A steady journal also notes where the wave lands, because the pain is not distributed evenly. The makers of memory itself are, on the face of the numbers, the riders: a 60 percent price increase on the flagship part flows almost entirely to their revenue line, and the per-gram export price statistic is, in effect, their rate sheet. The suppliers of the manufacturing equipment behind them benefit on a lag, because high prices fund the capacity expansions that those suppliers sell. I do not have a verified figure for that equipment effect, so I will mark it as a directional observation rather than a number.
The payers are the device makers and, downstream of them, the buyers. A laptop, a phone, a server, a car’s infotainment system — every one of them carries DRAM, and every one of them just absorbed a component whose price moved 30 to 60 percent in a matter of weeks. The contract positions vary, which is the only reason the repricing is not instantaneous, but contract terms run out. When they do, the component increase lands in the bill of materials, and the bill of materials decides the shelf price.
The asymmetry is worth stating plainly because it determines who the evidence should worry on behalf of. The memory industry is having its best quarter in years; the device industry is having its least predictable procurement quarter in years; and the consumer sits at the end of the chain with the least information and the most exposure. That is not a complaint about anyone; it is just the shape of the current.
When the tide turns
A steady journal also notes what would turn this tide, so the watch can be useful. The first is capacity: when the announced fabs reach volume production, the supply side of the ledger starts to recover, lead times shorten, and the per-gram price stops being the headline. The second is demand: if the AI build-out pauses or cools, the pull on memory eases and the scarcity loosens from the other side. Either development is visible in advance — in capex announcements, in fab construction milestones, in lead-time reports — which is why watching those numbers is the responsible position.
I have been wrong about timing before, and I want to be honest about that. The memory cycle is old and the calls are hard; a cycle that looks long can still turn quickly when demand rolls over. But there is a difference between being wrong about the timing and being wrong about the direction. The direction here — supply-constrained, priced by scarcity — is the direction the data has pointed all month.
The other number to watch is the contract price for server memory, because it is the slowest and most committed measurement in the file. A 23 percent move in a single month in the contract market is the negotiation layer catching up to the spot market, and it is the layer that eventually sets the cost basis for the next quarter’s products. When contract prices flatten, the wave is cresting; when they start to fall, the tide is turning. That is the measurement I will be watching, because it is the one that does not panic and does not hype — it just negotiates.
The sobering part
The sobering part of this story is not the price. Prices rise and fall; that is what markets do. The sobering part is the implication that the scarcity is structural and the ripple is broad. A component worth more than its weight in gold does not get cheaper by wishing; it gets cheaper when the wave of supply actually lands, and that landing is years out.
So here is the steady-journal summary, in the plainest terms I can manage: memory is in a supply-constrained supercycle, the numbers are consistent across four independent measurements, and the consequences will show up in product prices rather than staying inside the semiconductor industry. Evidence before alarm — and the evidence, measured carefully, is the more unsettling of the two.
Alarm adds nothing to the evidence. What adds to the evidence is watching the lead-time number, because when fifty weeks begins to shorten, the tide is turning. Until it does, the per-gram price is the honest measurement of a market that is telling us, in the calmest possible way, that it does not have enough.