Somebody put two numbers side by side on 20 September and the hardware internet has been chewing on them since. A square millimetre of 1b-generation DRAM: $0.654. A square millimetre of wafer off TSMC’s N2 line: $0.424.
Memory silicon, in other words, now carries half again the value per unit of area as the most advanced logic process in volume production. That inverts the assumption every hardware conversation has rested on for twenty years — the die doing the thinking is the expensive one, and the memory sitting next to it is the commodity you buy by the stick.
The chart came from Kurnal, of Kurnal Insights, and Tom’s Hardware wrote it up on the 22nd. It is worth walking through, because the arithmetic is sound and the conclusion people are drawing from it is not quite the one the numbers support.
The arithmetic
A 300-millimetre wafer has a gross area of about 70,686 mm². Divide a wafer price by that and you get a nominal price per square millimetre.
Kurnal takes a TSMC N3 wafer at $20,000, which works out at $0.283/mm², and an N2 wafer at $30,000, or $0.424/mm². Both figures come from media leaks rather than a price list — TSMC quotes per customer and per volume, and a separate analyst estimate last year put 3nm nearer $18,000, so treat $20,000 as the top of a range rather than a fact.
The memory side is calculated differently. Take a DRAM price of $1.50 per gigabit and multiply it by how many gigabits a generation fits into a square millimetre:
- 1y DRAM, at 0.219 Gb/mm² → $0.329/mm²
- 1z DRAM, at 0.273 Gb/mm² → $0.410/mm²
- 1b DRAM, at 0.436 Gb/mm² → $0.654/mm²
I ran the multiplications. They come out exactly as published, and 1b lands at 1.54 times the N2 figure. Nothing is being fudged.
The two numbers are not the same kind of number
Here is where the reading gets away from the chart.
The logic figure is a foundry charge. It is what a customer pays TSMC to run a wafer through the fab, before Nvidia or Apple or AMD adds anything of its own on top. The memory figure is a market price per bit, in the middle of the worst memory shortage in a decade, with the memory maker’s margin already baked in.
So the comparison is not memory manufacturing cost against logic manufacturing cost. It is what DRAM area currently sells for against what leading-edge logic area currently gets billed at. That is a statement about who is capturing the value in this shortage, and the answer is the memory makers. TrendForce put first-quarter DRAM industry revenue up 81% quarter on quarter, and on 7 September had conventional DRAM contract prices still climbing 13 to 18% for the quarter, with consumer parts worst hit because suppliers had pulled supply toward servers.
The comparison also leaves out packaging entirely, and packaging is the side where logic spends heavily — advanced substrates, interposers, the whole CoWoS-shaped queue that has been the other bottleneck of this cycle. Neither figure accounts for edge exclusion, dicing lanes, test structures or yield. They are nominal area prices, and the analyst said as much.
Why the newest node looks the dearest
The detail that gives the game away sits in the DRAM column. 1b is the newest, densest generation on that list, and it comes out as the most expensive per square millimetre. 1y, the oldest, is the cheapest.
Manufacturing cost does not behave that way. A denser node exists precisely so that a bit costs less to make; if density made silicon dearer per area nobody would ever shrink anything. The numbers come out in that order because the price per bit is held flat at $1.50 across all three generations, so more bits per millimetre mechanically means more dollars per millimetre.
Which tells you what the figure actually is. Revenue density. The newest node is not expensive to run — it is the best at converting a fixed quantity of wafer area into bits that can be sold at today’s price. In a shortage, that is the most valuable property a fab can have, and it is why every memory maker is pushing capacity onto its leading node instead of adding conventional lines.
What lands on the desk
Strip the analysis away and a plain fact remains: memory area is the scarce, expensive thing right now, and it is being rationed accordingly.
That single fact is the machinery under most of the hardware stories of the past fortnight. It is why a headset ships at a thousand dollars. It is why Raspberry Pi has quietly locked its boards to their factory RAM size in firmware — the resale arbitrage that policy kills only exists because a memory chip is now worth stealing off a board. It is why eight-gigabyte graphics cards refuse to die at the bottom of the stack. Acer’s chief executive put the peak at mid-2027 and called a shortage running to 2030 impossible, which is either a read of the supply curve or a man talking his own book, depending on how charitable you are feeling.
For anyone building assets, the thing being repriced is an assumption rather than a component. For a decade the answer to a texture budget problem was that the next generation would have more memory. CD Projekt could build Night City at twice the texel density of The Witcher 3 partly because that was a safe bet. It is not a safe bet now. Streaming pools, mip budgets, virtual texturing, the whole apparatus built for the years when memory was the cheap part of the board — all of it is about to matter more than it has since the PS3, and for a reason that has nothing to do with graphics and everything to do with what hyperscalers are willing to pay for a gigabit.
Per-area figures, wafer price assumptions and bit densities from Kurnal Insights, posted 20 September 2026 and written up by Tom’s Hardware on the 22nd; the multiplications were re-run here and match. DRAM contract price movements from TrendForce. Wafer prices are leaked estimates, not published rates, and none of the figures account for yield, dicing or packaging.