Samsung Electronics has locked up most of its memory chip capacity for 2027 through 2029 under long-term agreements, with some customers pushing supply guarantees out to 2031. AI data center expansion has driven shortages across HBM, commodity DRAM and enterprise NAND, with major makers meeting only 60% to 70% of demand and 2027 commodity DRAM wafer capacity effectively sold out, strengthening memory vendors' pricing power.
HBM
高带宽内存
High-bandwidth stacked memory for AI accelerators, scarce and costly, acting as the bottleneck resource in inference hardware.
Curated Intelligence
The Information reported that Masayoshi Son approached Wall Street financiers about funding a takeover of SK Hynix, with a roughly 40% premium pushing the deal value toward $1 trillion. A SoftBank spokesperson said the transaction is no longer under active consideration. SK Hynix is a key HBM supplier for AI accelerators, so any deal would reshape the AI compute supply chain, though SoftBank's stretched balance sheet, Korean regulatory resistance and SK Group's lack of interest in selling are major obstacles.
Micron has moved the first manufacturing tool into its Singapore HBM advanced packaging facility, targeting initial output in early 2027 under a roughly $7 billion investment running past 2030. The plant is central to Micron's HBM capacity expansion for AI compute demand, alongside an advanced wafer fab slated for output in the second half of 2028.
Google is restructuring its TPU supply chain: on some new-generation TPUs it designs the compute die itself, MediaTek handles I/O chip design, and Google buys HBM directly from memory makers, bypassing Broadcom to cut intermediary costs. Broadcom still participates in other high-end flagship TPU programs, creating a dual-track model. The move reflects hyperscalers reducing NVIDIA GPU dependence and reshaping HBM pricing dynamics.
Positron's Titan servers abandon scarce HBM for cheap smartphone-style DRAM, letting a single machine hold models of up to 32 trillion parameters and making memory capacity per watt the key metric. Inference cloud Parasail will be the first to deploy Titan at scale, starting at 5MW, claiming 4X better cost-per-token than next-gen NVIDIA/AMD systems and targeting trillions of cost-optimized tokens per day by late 2027.
Driven by surging global AI chip demand, South Korea's cumulative exports surpassed $1 trillion this year, making it the fourth country after Germany, the US and China to reach that scale, up sharply from last year's record $709 billion. Memory giants Samsung Electronics and SK Hynix are the main beneficiaries, yet domestic consumption remains weak and the export windfall has not reached households. The Lee Jae-myung administration plans a 162.3 trillion won Future Response Fund to channel chip-related excess tax revenue into AI, biotech and other new growth industries.
South Korea is on track to join the $1 trillion export club on the back of memory-chip shipments from SK Hynix and Samsung Electronics, yet the KOSPI has fallen more than 40% from its June peak, with AI-linked stocks leading the slide and pushing Q3 consumption into its first contraction in nearly two years. Semiconductors employ just 0.8% of the workforce, concentrating the boom's gains, while Seoul plans a KRW 162.3 trillion future-response fund and support for services as a second growth engine.
TrendForce unveiled its 2027 memory market outlook in Seoul, projecting that new NAND capacity will push flash prices down starting in Q2 2027, while general-purpose DRAM prices will not decline through end-2027 due to limited new capacity and HBM crowding out supply. The firm also expects HBM to rise from 23% to nearly 30% of DRAM wafer output at the three major memory makers, with HBM4 pricing climbing from $2.05 to $3-5.5 per Gb.
Samsung has kicked off a 144 trillion won long-term investment in South Korea's Chungcheong region spanning advanced semiconductor packaging, OLED panels, AI server substrates and batteries. The centerpiece is a fifth HBM packaging line at Onyang, slated for June 2028 production, aimed at turning the Onyang-Cheonan complex into a frontier HBM hub for AI semiconductors.
A federal judge in the Northern District of California denied the joint motion to dismiss filed by Samsung, SK Hynix and Micron, ruling that plaintiffs plausibly alleged a Sherman Act Section 1 conspiracy to restrict DRAM supply and inflate prices. The suit claims conventional DRAM prices surged 697% in 18 months as the three firms shifted wafer capacity to HBM, while together holding over 91% of global DRAM revenue. No merits ruling has been made; the companies must respond by December 7.