SEOUL – The companies building the data centers that make large-scale AI possible cannot get enough of a specific type of memory chip. That constraint, invisible in most earnings calls and almost never the subject of a mainstream product announcement, just handed Samsung Electronics its best quarter in company history.
Samsung’s Device Solutions division, which makes the memory and logic chips the rest of the technology industry depends on, posted operating profit of 89.2 trillion won in the April-to-June period, a figure that represents more than a 250-fold increase from the same quarter a year ago, when the division was barely generating returns above its own cost of capital. Total company operating profit reached 89.5 trillion won, also an all-time high. Revenue came in at 171.5 trillion won, up 28 percent from the first quarter and a record for any single quarter in the company’s history, according to a results release published Thursday by Samsung Electronics.
The memory type at the center of the surge is HBM, or High Bandwidth Memory, a stacked architecture designed specifically for the parallel computation that training and running large AI models requires. Standard DRAM, the memory in a laptop or phone, moves data one stream at a time. HBM stacks multiple dies vertically and connects them with thousands of tiny wires called through-silicon vias, multiplying the bandwidth available to an AI accelerator by an order of magnitude. Samsung said server products drove AI demand in the quarter “despite capacity constraints,” meaning the company could have sold more than it was able to make.
What makes the Q2 result notable beyond the headline numbers is where Samsung is positioned in the next product generation. The company said it scaled up HBM4 sales during the quarter and, separately, shipped what it called “industry-first” HBM4E samples to major customers. HBM4E is the generation after HBM4, and Samsung’s claim of being first to deliver samples would mark a meaningful shift if it holds. SK Hynix, Samsung’s South Korean rival, established itself as the dominant HBM3E supplier in recent years, becoming the primary memory source for Nvidia’s current generation of AI accelerators. HBM4E samples in major customers’ hands suggests Samsung is competing aggressively for the generation that will define AI hardware through 2027 and beyond, though samples and high-volume production are not the same thing.
Samsung shares and SK Hynix shares have both been battered in recent weeks on anxiety about Chinese chipmaking competition. The Philadelphia Semiconductor Index entered bear market territory after an unverified report that a Chinese state-backed firm had begun mass production of domestic chipmaking equipment circulated through markets, with both Korean memory makers shedding more than 15 percent in a single session. Thursday’s earnings, covering the quarter that ended before those fears surfaced, tell a different story about the underlying business: AI infrastructure spending is producing returns that most of Samsung’s corporate history cannot match.
The depth of that infrastructure spending showed up five days before the earnings release as well. On July 25, Samsung and Broadcom signed a memorandum of understanding to expand collaboration in memory and foundry technologies for AI infrastructure development. Broadcom is one of the largest suppliers of custom AI accelerators to cloud providers that want chips tuned to their own workloads rather than relying on Nvidia’s general-purpose architecture. Memory that works efficiently with those custom chips is exactly what Samsung’s semiconductor division sells.
The risk embedded in Samsung’s record quarter is a version of the same risk embedded in every AI infrastructure story: the demand driving it is itself driven by capital expenditure decisions made at a handful of very large companies. Meta Platforms posted record quarterly revenue of $60.8 billion this week but watched its free cash flow collapse to $784 million as $31.1 billion in AI infrastructure spending consumed operating profits. The hyperscalers are collectively betting that AI infrastructure built today will generate returns that justify the cost. Samsung’s chip division is the direct beneficiary of that bet. If the capex cycle slows and one or more major cloud providers decides the returns are not arriving fast enough, the memory market that just produced 89 trillion won in quarterly profit could soften faster than Samsung’s production lines can be redirected.
Samsung’s forward guidance does not anticipate that turn. The company said it expects “robust server demand from continued AI infrastructure investment and agentic AI adoption” in the second half of 2026, and described the market as remaining undersupplied despite moderation in mobile and PC memory demand. Production constraints are expected to persist. That language is, of course, supplied by Samsung, which has an obvious interest in presenting the demand outlook favorably. What it cannot control is the speed with which HBM supply eventually catches up to demand, or the pace at which AI hardware architectures evolve in ways that change the specific memory geometry they require.
One thing Samsung has not disclosed: which “major customers” received HBM4E samples. The company’s results materials do not name its AI memory buyers, and its relationship with Nvidia, the most important AI hardware company in the HBM supply chain, has not been publicly characterized by either party. Whether Samsung has regained meaningful Nvidia volume with HBM4, or whether the HBM4E samples represent a different customer set entirely, is not answered by Thursday’s release. Nikkei Asia, which reported on the all-time high results, noted that questions about how long the AI boom will last are clouding the company’s longer-term outlook.
What is answered by Thursday’s results is the question of whether AI memory demand is real enough to show up in a major company’s financials in an unmistakable way. The DS division’s operating profit has moved from a number that barely justified calling it a business to a figure that accounts for virtually all of Samsung’s total corporate earnings in a single quarter. That shift did not happen because Samsung invented something radically new. It happened because the rest of the technology industry decided that AI was worth spending on at a scale that memory supply cannot yet match. The chip sitting between that decision and its consequences is, right now, a stack of DRAM dies manufactured in South Korea.
