As AI computing power demand continues to reshape the DRAM industry cycle, three major financial reports have provided answers. Recently, Samsung, SK hynix, and ChangXin Memory Technologies (CXMT) released their results, confirming the sustained high prosperity in the memory sector: SK Hynix's Q3 revenue rose 328% year-on-year; Samsung's memory business revenue surged 427% year-on-year; CXMT's year-on-year growth rate remained at a high level of 589% to 619%. Amid this boom, questions arise: How long will this AI cycle last? Can physical AI scenarios such as autonomous driving and robotics become the second curve for memory market growth?
Korean Duo: The Engine of Profit Explosion Lies in "AI Memory"
SK hynix reported Q3 revenue of CNY equivalent to KRW 104.62 trillion, a 328% year-on-year increase, surpassing the KRW 100 trillion mark for the first time. Operating profit reached KRW 83.33 trillion, surging 632% year-on-year, with the operating margin rising to 80%, setting a new historical record for a single quarter. Samsung also delivered impressive results, with single-quarter operating profit reaching KRW 107.44 trillion, up 783% year-on-year, making it the first Korean company to break the KRW 100 trillion barrier. Its semiconductor division alone generated KRW106.5 trillion in quarterly revenue with a 72% margin, effectively accounting for the entire company's profits.
Both financial reports addressed market concerns about whether "AI infrastructure investment can be sustained."
Starting with SK hynix, its 80% operating margin defined the quality of this report. DRAM shipments grew approximately 10% quarter-on-quarter, while Average Selling Price (ASP) increased by 20% quarter-on-quarter. NAND ASP also rose by 20%, driven by an increased share of high-value-added server products. HBM4 shipments began fully ramping up in Q3, alongside increased sales of server DDR5. On the NAND side, 321-layer TLC enterprise SSDs continued to contribute incremental volume.
SK hynix explicitly guided that in Q4, both shipment volume and revenue share of HBM4 would exceed HBM3E, becoming the main product. Currently, SK hynix has signed long-term supply agreements with 14 core customers (including major global cloud service providers), locking in over 50% of capacity for 2027–2028. Some customers are already initiating early negotiations to extend the typically five-year agreement terms.
Turning to Samsung, its memory business generated KRW 140.7 trillion in Q3 sales, a massive 427% year-on-year surge. Its HBM revenue share climbed from 21% in Q1 2026 to 33% in Q2, narrowing the gap with SK hynix from 37 percentage points to 17 percentage points within one quarter. Samsung plans to increase monthly HBM wafer input from 180,000 wafers to 250,000 wafers and push the share of the HBM4 series to 80% of total shipments.
Notably, both Samsung and SK hynix are expanding production cautiously, refusing to repeat the path of blindly expanding general-purpose DRAM capacity seen in the previous cycle. SK hynix's capital expenditure for 2026 is expected to approach KRW 50 trillion, with nearly all new advanced process capacity directed toward HBM and AI server DRAM.
Overall, HBM and enterprise-grade DRAM/SSD are the core pillars of performance for Samsung and SK hynix, with long-term contract lock-ins becoming the norm. Armed with ample cash flow, these Korean giants are simultaneously investing in HBM technology and packaging capacity while securing long-term supply agreements, essentially balancing expansion impulses with profit safety margins during a period of high prosperity.
CXMT: Net Profit Contraction, Trading Current Investment for Technological Positioning
CXMT's Q3 revenue was approximately CNY 114.7 billion to CNY 119.7 billion, maintaining a high year-on-year growth rate of 589% to 619%. However, quarter-on-quarter growth slowed from 95.89% in Q2 to 15% to 20%, and net profit increased only 1% to 4% quarter-on-quarter.
While CXMT's Q3 revenue saw significant year-on-year growth, quarter-on-quarter momentum eased, and profit growth nearly stalled. This is partly due to a narrowing in DRAM price increases. Additionally, impacts from minority interests, front-loaded depreciation of new production lines, and expensing strategic investments in HBM in the current period collectively suppressed short-term profits. In other words, CXMT is currently in a phase of "trading current profits for future capacity and technological positioning," rather than experiencing weakening profitability.
In the past two months, CXMT has frequently updated its technology and R&D progress, including mass production of the latest generation Low-Power Double Data Rate 6 (LPDDR6) memory in late August, mass production of the fifth-generation DRAM technology platform (G5) in September, and an announcement in October that it will launch next-generation DDR5 RDIMM products using the 4F² DRAM architecture by year-end.
Currently, CXMT has established a foothold in the general-purpose DRAM and mobile LPDDR markets, capturing cyclical benefits that still have room for continuation. Data shows that the proportion of HBM wafer starts by the three major manufacturers—Samsung, SK hynix, and Micron—to total DRAM wafer starts has risen from approximately 18% at the end of 2025 to 22% at the end of 2026, and is projected to reach 30% by the end of 2027. The "crowding-out effect" of HBM on traditional DRAM capacity is intensifying. This is the structural reason for the persistent tightness in PC and smartphone DRAM supply and subsequent price hikes, which also represents an opportunity for CXMT.
However, regarding high-end server storage and the HBM track oriented towards communication computing infrastructure, CXMT still needs continuous technological catch-up and ecosystem breakthroughs. Currently, CXMT's HBM3 samples have been sent to domestic AI chip makers and remain in the pre-mass-production validation stage; the mass production target for 12-layer HBM3E is set for 2027.
How Long Will the Memory Bonus Last? Robotics May Open a New Curve
Under this super-cycle in memory, profit margins for Samsung Electronics, SK hynix, Micron Technology, CXMT, and YMTC (Yangtze Memory Technologies Co., Ltd.) have all hit historic highs. But how long can this upward trend persist?
Slowing price hike rhythms have triggered peak concerns. Sigmaintell Consulting stated that in Q3 2026, global memory market supply and demand remained tight, with prices across all categories continuing to rise quarter-on-quarter, but the magnitude of increases has clearly slowed, with some memory chips seeing quarter-on-quarter gains narrow to within 10%. TrendForce also predicts that general-purpose DRAM contract prices in Q4 this year will rise 10%–15% quarter-on-quarter, a significant drop compared to the approximately 60% quarterly surge seen in Q2.
Furthermore, long-term agreements account for an increasingly large portion of overall manufacturer shipments; these contracts include built-in price cap mechanisms, which also moderate the speed of price increases. The sustainability of AI capital expenditure is also being questioned. Meta recently hinted at setting caps on AI capital spending, triggering a sharp sell-off in the memory chip sector.
However, supply gaps extend the imaginative space for the prosperity cycle. SK hynix stated that the memory market shortage will persist until 2030, while Samsung judges that the gap will maintain until 2028. From a capacity perspective, it takes 2–3 years for wafer fabs to move from construction to stable output. Effective incremental capacity from memory manufacturers will not be released until the second half of 2027 or 2028, meaning the duration of the supply-demand gap may exceed previous market expectations.
Looking ahead to the memory market, SK hynix indicated that short-term demand will be led by enterprise-grade and consumer-grade AI, while medium-to-long-term growth will be driven by physical AI applications such as autonomous driving and robotics, ensuring the market continues to grow.
Tesla CEO Elon Musk recently stated that future robot demand for memory (DRAM) could exceed the sum of all other applications on Earth. A Counterpoint report also predicts that humanoid robot demand for DRAM will grow 20-fold by 2030, with average per-unit capacity increasing from 19GB to 39GB. Analysts estimate that if Tesla produces 10 million Optimus robots annually, DRAM consumption alone would amount to approximately 720PB of memory.
Beyond humanoid robots, high-level autonomous driving is another critical scenario where "physical AI" consumes memory. To achieve Level 4 and above full autonomy, vehicles must process vast amounts of data in real-time from dozens of cameras, LiDARs, and millimeter-wave radars.
Overall, AI has defined this memory bull market. Korean manufacturers firmly control pricing power in the HBM track through first-mover advantages and mature ecosystems. Domestic memory manufacturers are sharing in the price hike bonuses via the general-purpose DRAM track while continuously investing resources to catch up with high-end products. In the medium-to-long term, as AI moves from data centers into the physical world, the incremental LPDDR demand brought by humanoid robots and autonomous driving is expected to open a second growth curve for the memory industry. For domestic memory enterprises, this presents both a new wave of demand opportunities and a long-term challenge to continually narrow the gap in high-end technology.
-END-
Editor/Layout: Zhu Wenfeng
Proofreader: Mei Yaxin
Supervisor: Liu Qicheng