AI's Memory Hunger Just Gave Smartphones Their Worst Quarter in 14 Years

The numbers that landed on Monday are the kind the smartphone industry has not seen in more than a decade. Global smartphone shipments fell 20 percent year on year in the third quarter of 2026, the worst third quarter the industry has recorded since 2012, according to preliminary estimates from Counterpoint Research. Nearly every major brand shrank. Yet people did not suddenly stop wanting phones. The phones got too expensive to build.

Samsung held the top spot with a 23 percent market share, helped by demand for its Galaxy Z Fold8 series and stable component supplies for mid-range devices, though its shipments still slipped 3 percent. Apple crossed 20 percent share in a third quarter for the first time, reaching 22 percent, despite a year-on-year decline in shipments; the absence of a base iPhone 18 model meant a smaller launch-month channel fill than the full iPhone 17 lineup a year earlier. OPPO and vivo each fell more than 20 percent, and Xiaomi collapsed by more than 30 percent, dragging its share to 10 percent as it cut exposure to lower-priced devices and some overseas markets to protect margins.

Behind every one of those declines sits the same bottleneck. AI infrastructure projects are cornering global supplies of memory chips, driving up production costs. Manufacturers raised retail prices, cut shipments, and prioritized profitability over volume. Consumers, in turn, are holding on to their devices longer than analysts estimated, increasingly considering refurbished devices, Counterpoint senior analyst Shilpi Jain said.

The culprit is the AI memory boom

The week before the shipment figures landed, AMD chief executive Lisa Su gave the game away during a visit to South Korea. Meeting reporters on October 7, she described global chip market conditions plainly: memory supply is tight. She named Samsung Electronics and SK hynix as core supply chain partners and stressed that AMD values relationships lasting three to five years or more. In a tight market, long contracts are the only reliable currency.

Samsung's preliminary third-quarter earnings, announced the following day on October 8, showed how large the profits from that shortage can be. The company put operating profit at 107.4 trillion won, effectively all of it from memory. Meritz Securities estimated that memory generated about 109 trillion won in operating profit, including 80.1 trillion won from DRAM and 28.7 trillion won from NAND flash, covering losses estimated in the trillions of won from non-memory operations like the Device Experience division, foundry, and System LSI.

The striking feature of the quarter is that price increases, not volume, drove the gain. Meritz estimates Samsung's DRAM and NAND shipments rose only 3.4 percent and 4.2 percent from the previous quarter, while average selling prices jumped 20.1 percent and 19 percent. Volume growth stayed in the single digits, but the sharp price gains sent profits soaring.

The most profitable business in chips right now is not AI accelerators. It is the memory that feeds them.

How memory got this expensive

The mechanism is structural rather than cyclical. Samsung, SK hynix, and Micron, the three companies controlling more than 95 percent of global DRAM production, have concentrated wafer capacity on high-bandwidth memory and server DRAM for AI accelerators. HBM, the stacked memory that sits directly next to AI processors, consumes more than three times as many wafers as conventional DRAM of the same capacity. Every wafer dedicated to HBM is one that cannot supply phones, PCs, or conventional servers.

TrendForce estimates that HBM accounted for about 18 percent of the top three suppliers' DRAM wafer input at the end of 2025, and forecasts 22 percent by the end of 2026 and 30 percent by the end of 2027. KB Securities projects that commodity products, excluding HBM, will fall to 59 percent of global DRAM capacity next year, down from 65 percent this year and 73 percent last year. NAND faces similar pressure as demand expands for enterprise solid-state drives used in AI data centers.

According to DRAMeXchange, the contract price for DDR4 8-gigabit DRAM, a commodity product, rose more than fourfold to 26 dollars in September this year from 6.30 dollars a unit in September last year. Over the same period, the price of 128-gigabit multi-level cell NAND jumped about eightfold to 30.60 dollars from 3.80 dollars.

And the squeeze is expected to deepen. The industry expects Samsung's prices for HBM4, the sixth-generation high-bandwidth memory it became the first company to mass-produce, to roughly double next year from this year's levels, and TrendForce projects global HBM average selling prices will rise 121 percent next year.

The smartphone crash is not a demand crisis, it is a supply crisis, and the supply is sitting inside AI data centers.

How the phone industry is adapting, painfully

The adaptation strategies are already visible, and they are ugly for buyers. Apple quietly removed the 512GB memory option from the Mac Studio this week; the machine now tops out at 256GB, the price of the 96GB-to-256GB upgrade rose from 1,600 to 2,000 dollars, and shipping times stretch into May. Xiaomi is retreating from low-margin devices and some overseas markets altogether. Samsung's relative resilience is partly explained by its own memory fabs: it can supply itself when rivals cannot.

The second-order effects will keep compounding. A single Nvidia Blackwell B200 accelerator requires 192 gigabytes of HBM, and one AI server with eight GPUs holds more memory than thousands of consumer devices combined. Every hyperscaler order that lands with Samsung, SK hynix, or Micron is capacity that never reaches a phone assembly line. Some customers are already adjusting the amount of memory they install in their products to cut costs, which means the phones of 2027 may ship with less memory than the phones of 2025, at higher prices.

What comes next

Analysts say Samsung's earnings have yet to peak. Data provider FnGuide shows the consensus estimate for Samsung's full-year operating profit at 366.4 trillion won, rising to 538.9 trillion won next year, an increase of about 47 percent. Meritz Securities forecasts an even higher 632.4 trillion won. The market expects the memory windfall to keep growing.

This is the deeper lesson for anyone waiting for prices to normalize. The cycles of 2017 to 2018 and 2021 resolved themselves within about 18 months because manufacturers simply built more capacity. The 2026 cycle is different: manufacturers are spending heavily on capital, but the new capacity is targeted at HBM, not commodity DRAM. New wafer fab investments built today produce HBM, not DDR5, because HBM margins justify the investment and DDR5 margins do not.

Relief has three possible paths, and all are slow. Dedicated HBM capacity takes two to three years to come online. AI demand could moderate, though no analyst is forecasting that near term. Or efficiency gains, such as 300-plus-layer NAND, could gradually improve supply economics. The most likely outcome is a prolonged period of elevated memory prices, with consumers paying more for phones and holding them longer. The AI boom is real, but it is being financed, in part, by the phone in your pocket.