What Does Samsung Do? The Memory Invoice

STACK · Semiconductors · HBM · Status: Tightening · Last reviewed: 6 Nov 2026

Samsung Electronics gets paid on the memory line, which the company put at KRW 140.7 trillion for the quarter to 30 September 2026, while System LSI and the foundry remain inside a KRW 6.4 trillion remainder that was the same size a year earlier.

In short. Device Solutions did KRW 147.1 trillion of sales and KRW 106.5 trillion of operating profit. Memory was KRW 140.7 trillion of the sales. The other KRW 6.4 trillion, our arithmetic, is not split. Operating cash flow was KRW 98.6 trillion against KRW 26.9 trillion for plant. The company names capacity into 2027 as the limit, and it set a KRW 30 trillion dividend while paying KRW 2.5 trillion of dividends in the quarter.

Why this sits on the map: the mechanism is high-bandwidth memory. The logic wafer under that stack is a separate invoice, the HBM base die. Samsung runs both. One prints a sales line. The field of sellers is who makes HBM.

Numbers that matter

Dated 29 October 2026, quarter ended 30 September 2026. Memory sales and the cash-flow lines are from the earnings presentation. The dividend, capital spending and headline profit are also in that day's newsroom release. The KRW 6.4 trillion gap is our arithmetic. Unit sales include sales between units. The external review was not finished when the slides were issued.

What does it actually do?

Samsung Electronics is a group: phones and appliances, displays, Harman, and Device Solutions. Device Solutions is memory, System LSI and foundry. This quarter the sales line that moved was memory. The company says it pointed the mix at AI products amid capacity constraints, that server demand lifted prices, that HBM4 sales were scaled up, and that server SSD demand was met. It does not print HBM revenue, bit growth, or a memory operating profit.

The foundry paragraph has no won. The slides report a substantial improvement in revenue and earnings on HBM base-die sales and advanced nodes, plus 2-nanometre computing orders that include major hyperscalers. System LSI reports chip and image-sensor growth, also with no won figure. Device Solutions sales of KRW 147.1 trillion minus memory sales of KRW 140.7 trillion leave KRW 6.4 trillion (our arithmetic). A year earlier the gap was the same KRW 6.4 trillion, from KRW 33.1 trillion and KRW 26.7 trillion. In the second quarter it was KRW 6.7 trillion. Unit sales include sales between units, so that residual is not a clean external foundry invoice. It is the only size on the page for System LSI and foundry together.

Inside this group the base die can be a foundry sale into the memory business. The group shows the memory invoice and not the base-die invoice. The other arms are on the page. Device eXperience sales were KRW 52.1 trillion, with an operating loss of KRW 0.8 trillion. Samsung Display did KRW 8.0 trillion of sales and KRW 1.1 trillion of operating profit. Harman did KRW 4.3 trillion and KRW 0.6 trillion. Add those operating lines to Device Solutions: 106.5, minus 0.8, plus 1.1, plus 0.6, equals 107.4. The profit sits in Device Solutions. Memory is about 96% of that arm's sales, and the arm is about 99% of group operating profit (our arithmetic).

How does it make money?

Group sales were KRW 195.3 trillion, against KRW 86.1 trillion a year earlier. Cost of sales was KRW 60.0 trillion, against KRW 52.6 trillion. Sales rose KRW 109.2 trillion and cost rose KRW 7.4 trillion. About 93% of the extra sales stayed as gross profit (our arithmetic). Gross margin was 69.3%, against 38.9%. Operating margin was 55.0%, against 14.1%. Depreciation was KRW 11.9 trillion, against KRW 11.0 trillion. The cost base did not scale with the sales line.

The company says prices rose, led by servers. Bits are not printed, so the cost gap stays a price-and-mix fact next to a missing volume fact. Memory sales ran KRW 26.7 trillion, then KRW 120.8 trillion, then KRW 140.7 trillion. Device Solutions profit ran KRW 7.0 trillion, then KRW 89.2 trillion, then KRW 106.5 trillion. The company says memory set a record operating profit, and it does not print that profit. If the foundry improvement sits inside KRW 106.5 trillion, it is unsplit. If the foundry is still a loss, memory profit is larger than the arm. For the fourth quarter, memory says it will sell more HBM, server DRAM and SSDs. The foundry plans double-digit revenue growth from an unpublished base. For 2027, memory says demand is likely to tighten the shortage further versus 2026, and the foundry points to the start of operations at the first fab in Taylor, Texas, with no capacity and no month attached. Finance was a KRW 9.5 trillion expense, against KRW 4.7 trillion of income in the prior quarter. Net profit was KRW 75.1 trillion after tax of KRW 23.4 trillion. The slides do not explain the finance swing.

What limits it?

Cash is open. Operations generated KRW 98.6 trillion. Cash paid for plant was KRW 26.9 trillion. Cash on the company's broad line was KRW 214.8 trillion, debt KRW 17.2 trillion, net cash KRW 197.6 trillion. Net debt to equity was negative 32%.

Capital expenditure in the newsroom is a different line from cash paid for plant. The quarter was KRW 29 trillion, of which Device Solutions was KRW 27.2 trillion. Nine-month spending was KRW 56.9 trillion. The full-year expectation is about KRW 84.4 trillion, including KRW 76.8 trillion for Device Solutions, so the group's implied fourth quarter is about KRW 27.5 trillion (our arithmetic).

The limit the company names is supply. Memory managed its mix amid capacity constraints, and 2027 demand is likely to tighten the shortage further versus 2026. A plant cheque and a producing wafer are different clocks. The clean room is that mechanism. These materials do not date the extra bits. Inventories were KRW 72.0 trillion at 30 September against KRW 71.4 trillion at 30 June, a rise under 1% (our arithmetic), while sales rose 14%. Receivables rose from KRW 96.4 trillion to KRW 106.1 trillion, about 10%. Plant rose from KRW 224.4 trillion to KRW 234.0 trillion.

Who gets paid: Samsung, on the memory sales line. Shareholders, on a dividend set at KRW 30 trillion, KRW 4,601 a share. The cash that left was dividends paid of KRW 2.5 trillion, treasury stock of KRW 12.5 trillion, and debt down KRW 4.6 trillion. Of an estimated KRW 90–110 trillion of 2026 shareholder returns, KRW 35 trillion is already planned or completed and KRW 55–75 trillion remains, for a board in late January 2027. This page does not stack the new dividend on the KRW 35 trillion. Toolmakers and builders are paid from the KRW 27.2 trillion Device Solutions budget. Customers are who the company says is still short of product.

The foundry's other limit is disclosure. Double-digit fourth-quarter growth and a Texas start of operations in the 2027 outlook are in the materials. A sales line is not. What a 2-nanometre chip is explains the process. It does not fill in the won.

Who else does this?

SK hynix reported revenue of KRW 104.6155 trillion, operating profit of KRW 83.3338 trillion and an 80% operating margin for the same quarter, preliminary, on 28 October 2026. Samsung's memory sales line is about 1.3 times that revenue (our arithmetic). The scopes differ: hynix is a memory company, and Samsung's line includes sales between units. The margins that can sit together are Device Solutions at 72% and hynix at 80%. Samsung's 72% still contains System LSI and the foundry.

Hynix also prints what Samsung's slides leave blank. It expects HBM4 to pass HBM3E in shipment volume and revenue in the fourth quarter, and it says agreements with 14 customers cover more than 50% of 2027 and 2028 capacity. Samsung says it is building a sales base through long-term contracts, and it prints no share of output. Hynix's shortage language runs through 2030. Samsung's stops at 2027 versus 2026. Micron is the third seller, on its own page. Samsung says HBM base-die sales grew, and it does not say whether those dies are only its own stacks.

What would change this view?

A printed split would replace the inference. Memory profit well below KRW 106.5 trillion would mean System LSI and the foundry are carrying profit this page cannot see. A foundry loss would mean memory profit is above the arm. A foundry sales line would retire the KRW 6.4 trillion residual. If that residual shrinks while the company again says foundry revenue improved, the improvement is price, mix, or a sale between units. An HBM share of the KRW 140.7 trillion would show whether the mix is the stack or conventional server DRAM and SSDs. "Scaled up" does not say.

The price reading rests on KRW 109.2 trillion of extra sales against KRW 7.4 trillion of extra cost. If the next cost line catches up, or inventories build while sales stall, that reading weakens. If operating cash no longer covers the plant cheque and net cash of KRW 197.6 trillion turns down, the build has started to compete with the shareholder return. The board meets in late January 2027 to set the remaining KRW 55–75 trillion. A smaller dividend, or a 2027 capital plan several times the KRW 76.8 trillion Device Solutions budget, changes who is paid next. The presentation also warns that the external review can move this table.

Practical takeaways

Read KRW 140.7 trillion against KRW 147.1 trillion before the foundry paragraph. The KRW 6.4 trillion residual includes System LSI and sales between units, and it was the same size a year ago. The foundry paragraph has no sales line.

Sales rose KRW 109.2 trillion from a year earlier and cost of sales rose KRW 7.4 trillion. The company says prices rose and does not print bits. Cash is open: operating cash flow KRW 98.6 trillion, plant cash KRW 26.9 trillion, net cash KRW 197.6 trillion. The dividend set for the quarter, KRW 30 trillion, is a different number from dividends paid, KRW 2.5 trillion. The stuck step the company describes is supply into 2027.

Hynix's 80% margin is a memory company with an HBM4 volume claim and a printed share of 2027–2028 capacity under contract. Samsung's 72% is Device Solutions, foundry still inside it. The Korea Pulse is the short read beside this page.

FAQ

What does Samsung do in the memory layer?
Samsung Electronics sells memory, and it also runs System LSI and a foundry inside the same semiconductor arm. For the quarter to 30 September 2026 it printed memory sales of KRW 140.7 trillion and Device Solutions sales of KRW 147.1 trillion. The KRW 6.4 trillion difference is the rest of that arm. System LSI and the foundry have no sales line of their own.

Where does Samsung's profit sit?
In Device Solutions. That arm earned KRW 106.5 trillion of operating profit, about 99% of the group's KRW 107.4 trillion, which is our arithmetic. The company says the memory business set a record for both revenue and operating profit, and it does not print the memory profit on its own. The phone, TV and appliance arm recorded an operating loss of KRW 0.8 trillion.

How much HBM did Samsung sell?
The company does not say. It says HBM4 sales were scaled up significantly, and that server demand lifted prices across applications. The foundry section says HBM base-die sales grew, again without a won figure. A reader cannot see what share of the KRW 140.7 trillion was the stack.

What limits Samsung right now?
The company points to capacity. It says memory is managing its mix amid capacity constraints, and that 2027 demand is likely to tighten the shortage further versus 2026. Operating cash flow was KRW 98.6 trillion against KRW 26.9 trillion paid for plant, and net cash was KRW 197.6 trillion. The foundry's own size stays inside the KRW 6.4 trillion remainder.

Is this a recommendation to buy Samsung shares?
No. This page explains how Samsung makes money and where it can get stuck. It gives no buy or sell calls. Not investment advice. Do your own research.

Keep reading on the map

Layer: HBM, inside Semiconductors. Also on the layer: SK hynix and Micron. Explained: HBM, the HBM base die, who makes HBM, the clean room and 2-nanometre. Markets door: Korea, and this quarter's Pulse, who gets paid while Samsung builds the clean rooms. The HBM4 ask is a separate Korea note, what Samsung's ask is waiting on. Wire: the HBM4 ask, the base-die note, and hynix's quarter, the contracts and the return. Analysis: what Samsung's price assumes about the wafer pool.

Get the Wire — short daily notes when memory invoices, plant cheques and the HBM line move.

Not investment advice. Do your own research.

How we check this

Figures come from Samsung's 29 October 2026 earnings presentation and newsroom release, and from SK hynix's 28 October 2026 release, which marks its results as preliminary. Unit sales include sales between units, and Samsung's external review was not finished when the slides were issued. Secondhand transcript figures are not used. The residual, the profit shares, the cost gap, the fourth-quarter capital remainder and the inventory change are our arithmetic. Who gets paid, and supply as the stuck step, are our read. The featured image is a graded fab photograph with one empty chamber, not a Samsung factory. No buy or sell call.

Last reviewed: 6 Nov 2026 · Next review: after the reviewed third-quarter financial statements, or the fourth-quarter results, whichever prints a foundry or memory split first

Spot an error? Tell us.

Our editorial standards →

Sources

1. Samsung Electronics — third-quarter results, newsroom (29 Oct 2026) — group revenue and operating profit, Device Solutions revenue and operating profit, the DX loss, capital expenditure, the dividend, and the memory, foundry and 2027 language.
2. Samsung Electronics — 3Q 2026 earnings presentation (29 Oct 2026) — memory sales, the segment appendix, margins, cash flow, net cash, inventories and the pre-review notice.
3. SK hynix — 3Q 2026 results (28 Oct 2026) — revenue, operating profit, operating margin, the HBM4 volume claim, and the share of 2027–2028 capacity under long-term agreements. Marked preliminary by the company.

S

Sean

Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck.

Author page → · Editorial standards →