Samsung Delivers $59 Billion Profit in a Single Quarter with AI Memory, Forecasts Tightness Until 2028
Chip division earned more than the sum of the last three years combined. Adding SK Hynix, the Korean duopoly posted $104 billion in operating profit in one quarter and warns that supply will remain tight.
Samsung Electronics reported an operating profit of 89.4 trillion won, equivalent to about $59 billion, in the second quarter of 2026, a staggering 1,181% increase year-over-year and the group's third consecutive quarterly record. Consolidated revenue reached 171.5 trillion won, a 130% jump on an annual basis and also a historic peak. The stark assessment for the quarter: in three months, Samsung earned more operating money than it accumulated in the combined three calendar years of 2023, 2024, and 2025.
The engine of this machine is clear when broken down by division. The Device Solutions unit, which houses the memory and semiconductor operation, delivered an operating profit of 89.2 trillion won on 127.5 trillion won in revenue, accounting for nearly the entire group's result. The smartphone division returned to operate in the red during the period, something Samsung treated as an acceptable collateral given the windfall from the B2B side of the balance sheet. If the amount set aside for a special performance bonus in the semiconductor area, around 10.5% of that unit's profit, is removed, the operating profit would have approached 100 trillion won, an unprecedented figure for any quarter in the company’s history.
SK Hynix Confirms the Duopoly
The assessment gains importance when combined with the results from SK Hynix, released a day earlier. The rival Korean company reported revenue of 79.3 trillion won and an operating profit of 60.5 trillion won in the second quarter, increases of 257% and 557% respectively, with an operating margin of 76%. Together, Samsung and SK Hynix posted 150 trillion won, approximately $104 billion, in operating profit in a single quarter, highlighting the earnings collected by the Korean memory duopoly amid the consumption of HBM4 and DDR5 by Nvidia’s GPUs and the proprietary chips of AWS and Google.
Samsung was explicit in stating that server memory demand is likely to remain robust, supported by AI infrastructure capital expenditure and the introduction of agent systems that expand the rack footprint. More concerning for enterprise buyers, the group warned that supply tightness extends into 2027 and that a growing number of customers are negotiating multi-year supply contracts to secure capacity. Samsung projected that the memory chip squeeze could last until 2028, a scenario that impacts DIMM prices, server delivery times, and refresh planning in private and regional data centers.
Where the Effect Hits
Price pressure is spreading through three familiar avenues for CIOs. In the United States, AWS raised its 2026 capital expenditure to $220 billion and attributed part of the revision to the increase in memory prices, the same rationale used by Meta’s CFO to justify the higher spending floor announced this week, between $130 billion and $145 billion. In Germany, SAP and local providers operating sovereign cloud have already been reporting margin pressure in infrastructure contracts, a trend expected to intensify. In Japan, MUFG and Mizuho are renegotiating refresh schedules for core banking systems with integrators like Fujitsu and NTT Data, sensitive to memory price fluctuations. In India, delivery hubs of TCS and Infosys, supporting offshore operations for global clients, are dealing with equipment priced in won and dollars, a dynamic that has become less favorable throughout the quarter.
Cynical readers may raise valid arguments, and it’s worth naming them. Analysts from Morgan Stanley and JPMorgan continue to point out the risk that the race for capacity may replicate, on a larger scale, the boom-bust cycle seen with Micron in 2018 and 2019, when oversupply caused DRAM prices to plummet by more than 40% in twelve months. A reasonable counter-argument is that today’s bottleneck is HBM, a highly specialized product, rather than standard DRAM, and that the learning curve for new suppliers of HBM4 is long. Still, the lesson from the last cycle calls for a sober reading rather than a linear celebration.
The question that remains on the table is not whether the Korean duopoly profits from AI. It is for how long and at what price for the enterprise buyer. A senior architecture consultant today needs to reevaluate the TCO of modernization projects assuming more expensive memory, longer delivery times, and multi-year contracts as the norm, not the exception.