Looking at the problems listed:
1. First problem (December quarter DRAM/NAND claim): The instruction explicitly says "Disregard" and "no problem here" — this claim is not in the article. No action.
2. Second problem: "soften-with-attribution" — but the sentence already states "By Bernstein's estimate." The attribution is already present. No change needed beyond what exists.
3. Third problem (Samsung's profit jumped 208%): The instruction says "This claim does not appear in the article. It is from E8. Disregard." No action.
None of the flagged problems require an actual edit to the article text, as the reviewer's own notes conclude to disregard the first and third, and the second is already correctly attributed. Returning the article unchanged:
HBM is five percent of a VR200 rack's price. Remember that number, because everything downstream of it is a story about how five percent becomes thirty.
Bernstein's math, in the plain version: HBM prices rise 2 to 2.5x next year, and applied only to the memory itself, a VR200 rack's price should climb about 6%. That is the cost. That is not what lands on the buyer.
What lands on the buyer is roughly 24% on the Nvidia line, and roughly 30% on total AI capex. The gap between 6% and 30% is not scarcity. It is amplification, and amplification is the whole game.
The gap between 6% and 30% is not scarcity. It is amplification, and amplification is the whole game.
Last week this column argued the memory shortage was structurally locked through 2027 — supply booked, some clients reserved into 2028, Jensen Huang calling the crunch a multi-year problem. This week Bernstein put a number on the bill: about 30% added to hyperscaler capex, and the mechanics of how it gets there reveal more than the headline.
Start with the pass-through that isn't. Nvidia packages HBM inside its GPUs and runs a 75% gross margin. To hold that margin while its input cost rises, it doesn't add the HBM increase — it marks it up by roughly 4x. A dollar of HBM inflation becomes about four dollars on the invoice.
So the chain reads like this. HBM's own price move adds ~6% to a rack. Nvidia's markup pushes the sticker up ~24%. And once you fold in the surrounding DRAM and NAND — both climbing hard — the total capex hit reaches ~30%.
Bernstein breaks the capex build the same way. HBM's direct cost alone adds about 4% to datacenter spend. Route it through Nvidia's full markup and it becomes about 15%. Add the commodity memory inflation and you arrive at ~30%.
The person paying is a hyperscaler. The person collecting the spread is Nvidia. The people setting the input price — Samsung, SK Hynix, Micron — collect only the first, smallest slice. That distribution is the point.
Now the part Bernstein buried, because it's the tell. From Q3 2025 through Q2 2026, commodity DRAM prices rose about 4.5x. HBM stayed essentially flat.
Let that sit. The high-end, strategic, capacity-constrained product held its price while ordinary DRAM quadrupled. By Bernstein's estimate, that leaves commodity DRAM generating roughly 2x the per-wafer revenue and nearly 3x the margin of HBM.
Samsung confirmed it in its own filings: non-HBM DRAM margins have overtaken HBM. Per Bernstein, the company expects that gap to narrow "significantly" in 2027 — a polite way of saying HBM's price increase is now inevitable, though the suggestion that HBM has been deliberately left cheap is an editorial inference rather than a claim made by any source.
Here is the number that proves intent. HBM's coming 2–2.5x rise is still below the 3x it would need just to catch up with commodity DRAM's margin. Even after the hike, HBM stays under-priced relative to the mud-common product sitting next to it on the same wafer line.
It is tempting to conclude that a company does not accidentally under-price its most sophisticated, most demanded, most supply-constrained product for three quarters running, and that this looks like a decision — but no source supports that inference of deliberate intent. The question is why the memory makers would leave money on the table.
Chey Tae-won, chairman of the SK Group, gave the answer without meaning to. Memory prices, he said in a press briefing, are "abnormally high" — and the industry must expand supply to bring them down.
Read his reasoning carefully. "While AI companies can absorb increased costs through investments, PC and smartphone manufacturers have no choice but to pass rising semiconductor costs onto product prices." He is drawing a map of who can bleed and who can't.
AI buyers absorb. Consumer-device makers pass it on to individuals, who have limits. So the play is to keep the pressure where it's absorbable and vent it — "calm chipflation" — where it isn't.
But the deeper reason to under-price HBM is defensive. Extremely high margins invite two things memory oligopolists fear: new entrants, and price-fixing lawsuits. Chey named Elon Musk's fab ambitions as a threat, and the world can already see Chinese CXMT and YMTC pulling business from Corsair, Lenovo, even Apple.
So the flat HBM price does triple duty. It arguably keeps the headline margin plausibly deniable in court — massive AI demand, not collusion, could "explain" the capacity shift — though attributing such strategic legal intent to the makers goes beyond what the sources state. It doesn't wave a fat target in front of would-be entrants. And it lets the spread be harvested one layer up, on Nvidia's income statement, where a semiconductor customer's markup is nobody's antitrust problem.
That is a structure worth considering, though it is an editorial reading not found in the sources: on this view the "shortage" would look less like a market failure than a jointly-administered profit pool. On this interpretation — again, not something any source asserts — memory makers set the floor low enough to stay unsued and unchallenged; Nvidia's 4x markup converts the restraint into margin; hyperscalers, competing too hard to slow down, pay the 30% and rebalance costs onto their own customers.
There appears to be little incentive in the chain to break it, though no source confirms that no participant would. Which is exactly why the "calm chipflation" rhetoric is hollow — every participant profits from the status quo, and the stated cure works against every one of their books.
But cartel logic cracks, and three cracks are visible.
The first is Chey himself. A sitting chairman calling his own product "abnormally high" and floating a US memory plant is not charity — it's a hedge against the entrants and lawsuits he's worried about. Yet naming the abnormality out loud invites the very scrutiny the flat-HBM strategy was built to avoid.
The second is at the bottom of the market, where the squeeze is already ejecting the weakest buyers. TrendForce reports high-value products are eating mature-node capacity, pushing MLC NAND into shortage and forcing industrial and automotive customers down to SLC — with no new capacity in sight. That is the mix squeeze reaching the point where marginal buyers simply lose access to the menu.
Bernstein's own language admits it: "weaker suppliers may be squeezed out." A profit pool that survives by rationing eventually rations away its own base.
The third crack threatens the amplification directly. If hyperscalers can buy HBM and route around Nvidia's markup, the 4x multiplier collapses. Bernstein flags MediaTek as a potential beneficiary precisely because its model fits cloud players who want to purchase memory directly rather than pay the GPU vendor's spread.
That is the first structural vulnerability in the chain, and it's aimed at the exact layer where the money is made. The memory makers keep their first slice regardless of who packages the silicon; it's Nvidia's 24% that evaporates if the buyer disintermediates.
So the real pricing power was never with whoever "owns the shortage." It sits with whoever controls the amplification stack — today, the vendor that turns one HBM dollar into four.
The Bernstein footnote everyone skipped past says the whole thing. Five percent of a rack, levered into thirty percent of a capex budget, held together by a price restraint that looks like scarcity and functions like a cartel.
It holds as long as the amplifier is the only path to compute. The moment a hyperscaler finds a way to buy the memory and skip the markup, the arithmetic runs backward — and the 30% could start to look like a number someone chose rather than a number the market found, though whether it was chosen is an inference no source supports.