Samsung's zHBM roadmap chooses AI memory over commodity DRAM — what does that substitution do to its own Galaxy margins and its net DRAM payoff?

Samsung's phone business lost money in the April–June quarter because of the price of memory. The same company's chip business turned in a record profit from selling that memory. Both numbers appear in the same earnings release, and both are caused by the same capacity decision.

For the three months to June 2026, Samsung's Device Solutions (DS) segment — chips, chiefly memory — posted KRW 89.2 trillion in operating profit. Its Mobile eXperience (MX) division, the Galaxy phone business, reported an operating loss of KRW 0.7 trillion. The swing was violent even by memory-cycle standards: MX had earned roughly KRW 2.8 trillion in the prior quarter, before component costs caught it.

The reconciliation is a two-market story inside one company. Samsung is shifting wafer capacity away from commodity DRAM — the run-of-the-mill memory inside phones, PCs, and standard servers — toward high-bandwidth memory (HBM), the premium stacked AI memory it sells to Nvidia and hyperscalers at soaring prices. That allocation is the driver of both sides of the ledger. It shrinks the commodity supply Samsung's own Galaxy handsets depend on, which is exactly what jacks up the contract prices MX now pays. Samsung is, in effect, taxing its phone division to fund its memory division's record margins.

One capacity decision, two markets

Start with the allocation. Samsung has stated it will manage its output by "prioritizing high-value-added products over commodity DRAM", and its roadmap for the next phase, unveiled at Hot Chips and SEMICON Taiwan, is a series of HBM iterations that eventually stack DRAM directly on the processor (zHBM), eliminating the interposer and reshaping how the wafers are used. This is a structural commitment to AI memory over generic DRAM, not a tactical trim.

The capacity arithmetic behind it is public. TrendForce noted that Samsung and SK hynix had already shifted roughly 20% of DRAM production to HBM; by August 2026, industry reporting said the three big suppliers had effectively sold out their entire 2027 DRAM and HBM output. Samsung's own memory division has said it can fulfill only about 70% of incoming DRAM orders, and it has reportedly declined a supply request from its own MX unit to defend external pricing. Whether or not the zHBM endpoint ships in 2028 or after 2029, the direction — lean wafers toward AI memory, starve commodity — is already visible in current-year numbers.

That shift lands on the two segments in opposite directions, which a short table makes plain.

Samsung segment	Role in the trade	Q2 2026 operating result
Memory (in DS)	Sells DRAM/HBM at rising ASPs	+KRW 89.2 trillion (DS)
Galaxy (MX)	Buys commodity DRAM at market prices	−KRW 0.7 trillion

The contract-price index is the bridge between the rows. Conventional DRAM contract prices rose roughly 90–95% quarter over quarter in early 2026, then 58–63% in the spring; Citi put Samsung's DRAM average selling price up 44% in Q2 alone. On the cost side, analysts estimate DRAM's share of a premium phone's bill of materials climbed from 14% a year ago to 23%. That is the margin compression: MX swung from a small profit to a loss as its biggest single component roughly tripled in contract price within a couple of quarters.

The transfer is a rounding error

The temptation is to read this as Samsung "paying itself" and call the deal a wash. It is not — and the reason is scale.

The Galaxy loss is KRW 0.7 trillion against DS operating profit of KRW 89.2 trillion — under 1%. Samsung sells an enormous volume of DRAM to the open market and only a fraction into its own handsets, and that external volume sells at the inflated prices. The internal purchase that hurts MX is reallocated, not destroyed; it shows up as the memory division's own revenue. Net of that transfer, kicking the commodity side to keep HBM scarce is unambiguously value-accretive to the consolidated company. Whatever price Galaxy pays, the memory division receives; the only true cost is the small amount Samsung would rather sell externally on margin than to itself, and it has shown it will ration internal allocations to protect the external price.

So from an investor's consolidated view, the Galaxy "cost drag" is the wrong place to focus. It is the tail, not the dog.

What turns the tax off

The variable that actually determines the net payoff is not Galaxy margin — it is whether the scarcity holds. And here the evidence is genuinely two-sided, which is why the framing deserves a falsification test.

The price momentum is already decelerating sharply: contract-price increases have gone from roughly 90–95% quarter over quarter to 58–63% to a forecast 13–18% for the third quarter, as consumer markets hit affordability limits. That is the taper signal. But it is a taper in the rate of increase, not a rollover; prices are still rising, and the supply side is holding the line. Suppliers have said the "memory super-cycle" may stretch well past 2028, 2027 capacity is fully booked, and there is no evidence yet of wafers flowing back toward commodity DRAM — indeed the disclosed direction is the opposite.

That asymmetry is the investment point. If the disciplined scarcity holds, Samsung keeps printing oversized memory margins and Galaxy stays a small drag — net, heavily positive. If prices roll over or capacity reverts to commodity, the Galaxy cost pressure heals, but only because the memory windfall shrinks with it; at consolidated level that is still net memory-positive to Samsung, just less spectacular. Either branch, the Galaxy drain erases the memory gain only in the arithmetic where 0.7 trillion is compared with 89.2 trillion.

The zHBM choice is therefore not really a trade-off an investor has to adjudicate between Samsung's two segments. It is a bet that Samsung, by steering wafers from commodity DRAM to custom AI memory, stays the primary beneficiary of the same shortage it is creating. Watch the DDR5 contract index and the memory division's willingness to keep rationing commodity capacity — not the Galaxy margin line. The day prices stop climbing or the wafers come back, the cost drag and the windfall will fall together, and the net still belongs to the memory division.