Samsung Booked an $80 Billion Quarter. Its Phone Business Lost Money Doing It.

The headline writes itself. On Thursday Samsung Electronics told investors it expects to book 107.4 trillion won in operating profit for the three months to September. At the day's exchange rate that is about $80 billion. Revenue lands near 195 trillion won, and the operating margin of 55.1 percent is the highest in the company's history, a figure you almost never see from a manufacturer.

No tech company has ever posted a larger quarterly profit. Yonhap, which called the record unprecedented among global giants including Apple and Nvidia, puts the only bigger corporate comparable at Saudi Aramco's roughly $86.5 billion in the second quarter of 2022. Samsung's operating profit for all of last year was 43.6 trillion won. It has already booked 254.1 trillion won in the first nine months of this one.

Then look at the share price. Samsung stock slipped 0.2 percent in early Seoul trade on the news and sits more than 25 percent below its June record. The market saw this quarter coming. What it is arguing about is not the size of the number but how long it can hold.

What Samsung actually disclosed

The guidance is deliberately thin, and that is the first thing to understand about it. Samsung publishes a single consolidated sales and operating profit figure, because Korean disclosure rules do not permit the headline estimate to be presented as a range. The filing lists the underlying band anyway: operating profit of 107.3 to 107.5 trillion won, sales of 194 to 196 trillion won. Everything about which division earned what is inference, not disclosure. The detailed results, with the divisional breakdown, land on October 29.

That silence is doing work. The consolidated number is a record. The composition of it is where the story turns.

Where the money came from

This is a memory price story before it is anything else. DRAMeXchange figures cited by the Seoul Economic Daily put the spot price of a mainstream DDR4 8Gb DRAM chip at $26 at the end of the third quarter, more than four times the $6.30 it fetched a year earlier and up 23.8 percent from the prior quarter. NAND flash was running at roughly eight times its year-earlier price. Sell a commodity into a curve like that and the margin follows on its own.

The newer and more strategic part is HBM4, the sixth-generation high-bandwidth memory that sits beside an AI accelerator and feeds it data faster than conventional DRAM can. Samsung shipped the industry's first commercial HBM4 in February at 11.7 gigabits per second per pin, and it is qualifying for Nvidia's next-generation Vera Rubin platform. Douglas Kim of Douglas Research Advisory estimates Samsung's HBM bit shipments grew close to 50 percent quarter over quarter. Counterpoint Research data puts Samsung's HBM revenue share at 33 percent in the second quarter, up from 21 percent in the first, while market leader SK Hynix fell from 58 percent to 50 percent. The gap that swallowed Samsung in the HBM3E generation is closing.

Where the money didn't come from

Analysts' estimates, not company disclosure, sketch the other half of the picture. The foundry and system LSI units are believed to have lost about 1 trillion won between them, dragged down by the fixed costs of the Taylor fab in Texas and utilisation rates that are still climbing. The Device eXperience division, home to the phones and consumer electronics, is estimated to have posted a second straight quarterly loss. Within it, the mobile and network businesses are thought to have lost more than 1.5 trillion won, with displays and home appliances adding around 500 billion won of red ink. Reuters reported the mobile business alone losing more than $1 billion, a wider loss than analysts expected.

The mechanism is not subtle. The same shortage that made the memory division's quarter is a cost shock to the division that buys memory. Samsung's phone business is being squeezed by the price of the components it has to install, and it is buying many of them from itself. The company built its name on a vertically integrated model where making your own parts was the advantage. In a shortage this severe, integration becomes the thing that transmits the pain.

The structural problem underneath

Two variables decide whether this is a supercycle or a squeeze, and neither of them is demand.

The first is supply, because high-bandwidth memory is a spectacularly destructive use of a fab. Research house Nathan Research Group, drawing on TrendForce data, estimates that HBM will consume about 22 percent of the top three memory makers' DRAM wafer starts in 2026 while returning under 10 percent of the world's DRAM bits, a ratio projected to reach roughly 30 percent of wafer input by 2027. Every stack of HBM removes something close to three units of ordinary memory output. Relief has to be built, and the capacity funded today lands in 2027 and 2028, after the shortage it is meant to fix.

The second is price momentum, and the curve is already bending. TrendForce expects conventional DRAM contract prices to rise 10 to 15 percent in the fourth quarter, against a roughly 60 percent jump in the second. Analysts polled by Reuters expect Samsung's fourth-quarter profit to grow 8.2 percent sequentially, down from 20 percent in the third quarter. Both Samsung and Micron have said they expect the supply imbalance to persist into 2028, but the rate of change is doing something different from the level.

There is currency in the mix too. A stronger won erodes the local-currency value of dollar-denominated memory sales, and several analysts have trimmed forecasts for exactly that reason. Kim Seok-hwan of Mirae Asset Securities summed up the mood: "The market's focus has shifted to whether the sharp earnings growth that started a year ago would be sustainable."

What this means

For a decade the received wisdom was that memory was the commodity business, the boom-and-bust tail of the semiconductor industry, and that the durable money sat in finished devices. Samsung just spent a quarter demonstrating the reverse. The division that sells the commodity is the profit centre. The division that makes the product is the cost centre. A company that once used its chip arm to underwrite its phone arm is now watching the chip arm price its own phone arm out of profitability.

None of this is a verdict on demand. AI infrastructure spending is still enormous and the build-out is real. It is a verdict on structure. Memory has become the pressure valve in the AI supply chain, and the pressure is being released unevenly, into the balance sheets of memory makers and out of the balance sheets of everyone who has to buy from them.

That carries a planning consequence for anyone building hardware. The memory line item in a bill of materials is no longer a rounding error to be renegotiated at the margin. It is the variable that decides whether a product ships at a profit or a loss, and it is being set by a market where a fifth of the world's wafers return a tenth of the bits. At DMC, we work with hardware companies that have to model this kind of volatility before the first unit ships: sourcing strategy, cost modelling, and production ramp planning when the component that decides your margin moves 60 percent in one quarter and 12 in the next. If your roadmap still assumes memory behaves the way it did in 2024, let's talk.