Nvidia has launched its latest Rubin CPX GPU aimed at applications in the era of large-scale inference for AI. According to research firm SemiAnalysis, the launch represents a new direction in GPU development, leveraging decoupling and collaborative division of labor to meet AI inference demands. In addition to demonstrating design flexibility and foresight, the Rubin CPX GPU could also signal a fundamental transformation in the inference domain and the industry roadmap as a whole.
As SK Hynix kicks off its second-half 2025 recruitment drive with a generous bonus package, the move has generated significant buzz among Samsung Electronics employees, especially those with shorter tenures. Some are now openly weighing the prospect of jumping ship.
As Nvidia raises the performance threshold for fourth-generation high-bandwidth memory (HBM4), reports suggest that Micron Technology may be struggling to meet these standards. Industry analysts in South Korea suggest that Samsung Electronics stands to benefit from this shift.
Huawei unveiled its next-generation Ascend AI chip roadmap at the 2025 Connect summit, highlighting a milestone for China's semiconductor ambitions. Starting in early 2026, the Ascend 950 PR will integrate Huawei's self-developed high-bandwidth memory (HBM), marking a pivotal advance in the country's drive for independence in AI chip memory systems.
SK Hynix has announced it is the first in the world to establish a mass production system for sixth-generation High Bandwidth Memory (HBM4), signaling the start of a new technological arms race in next-generation memory chips.
Huawei is pressing ahead with its in-house chip and AI infrastructure strategy despite escalating US–China tech tensions. At Huawei Connect 2025 in Shanghai, deputy and rotating chairman Eric Xu unveiled the company's first detailed Ascend AI chip roadmap in six years, highlighting Huawei's advances in high-performance computing and long-term ambitions in AI infrastructure.
Global DRAM shipments saw their sharpest increase in nearly two years during the second quarter of 2025, solidifying SK hynix's lead over Samsung Electronics. The surge was driven by the rapid expansion of artificial intelligence, with high-bandwidth memory, or HBM, emerging as the key growth engine for the sector.
Hybrid bonding machines are set to become critical by 2027, when chipmakers move into mass production of 20-layer, sixth-generation high-bandwidth memory (HBM4). Global equipment suppliers are racing to bring their systems to market ahead of that transition.
In the first half of 2025, SK Hynix recorded the highest operating profit and the largest corporate tax payment among South Korean companies. The surge in demand for high-bandwidth memory (HBM), fueled by advancements in artificial intelligence (AI) applications, significantly contributed to the semiconductor manufacturer's financial performance, according to The Chosun Ilbo.
Driven by strong AI demand, memory contract prices for NAND and DRAM rose sharply by an estimated 15-20% in the fourth quarter of 2025, breaking traditional year-end price decline patterns. Supply shortages led to aggressive procurement from cloud service providers, with high-stack 3D NAND products nearly sold out.
Puya Semiconductor (Shanghai) announced on September 15, 2025, that it will acquire a controlling stake in Zhuhai Noah Changtian Storage Technology through a cash deal, giving it indirect ownership of SkyHigh Memory (SHM). The planned acquisition expands Puya's non-volatile memory portfolio by adding 2D NAND and derivative products to its established NOR flash, EEPROM, and MCU businesses, creating a more complete product lineup.
The United Kingdom demonstrated a prominent presence at SEMICON Taiwan 2025 last week, reinforcing its position as a key player in innovative technology. The UK Department for Business and Trade and Innovate UK, the national innovation agency, led a delegation of over 30 emerging companies across semiconductor manufacturing, design, testing, and materials sectors to the event.
South Korea's information and communications technology (ICT) exports in August 2025 reached a record high for the period compared to previous years. Due to rising memory semiconductor prices and increased investment in artificial intelligence server infrastructure, semiconductor export values also reached an all-time high.
The increasing demand for big data storage driven by artificial intelligence data centers is expected to intensify a supply shortage of QLC NAND Flash memory, potentially resulting in a significant deficit by 2026. Major US cloud service providers are planning to replace traditional hard disk drives with enterprise-grade solid-state drives on a large scale starting in 2026, which will drive explosive growth in demand for high-capacity QLC SSDs. This transition is anticipated to narrow the price gap between SSDs and HDDs to roughly three times.
Nvidia is preparing a potential shake-up in the AI semiconductor memory market with the introduction of its next-generation AI processor, Rubin CPX, which is designed for large-context processing workloads and expected to debut in late 2026. Unlike its flagship AI accelerators that rely on HBM, Rubin CPX will feature 128GB of GDDR7, a move that industry observers say could rebalance demand between the two leading memory types, according to SemiAnalysis, Wide Daily, and ET News.
The Ministry of Economy, Trade and Industry (METI) of Japan announced on September 12, 2025, that it will provide Micron Technology with subsidies amounting to as much as JPY536 billion (US$3.63 billion) for the development and mass production of next-generation DRAM at its Higashihiroshima factory. The assistance is intended to promote advanced research and expand production capacity.
China's DRAM maker CXMT is reportedly developing compute express link (CXL) DRAM modules, a move that could escalate rivalry with Samsung Electronics and SK Hynix as the CXL market takes off in 2026. According to South Korea's EBN, CXMT has outlined its CXL roadmap to Korean semiconductor component suppliers and begun early design work, with one major substrate partner confirming initial joint R&D on CXL substrates.
Nvidia has reportedly abandoned its initial SOCAMM1 low-power DRAM module for AI workloads, shifting focus to SOCAMM2 with faster speeds and broader supplier collaboration, including Micron, Samsung, and SK Hynix. This move signals a strategic push to address technical challenges and advance AI memory solutions ahead of mass production in 2026.
Memory manufacturers are swiftly following SanDisk's announcement of a 10% price increase for NAND products, signaling a broader industry move toward higher pricing amid expected supply shortages starting in 2026. Micron has paused NAND Flash and DRAM pricing quotes for a week, while market sources anticipate potential NAND Flash price rises of up to 30%.
SK Hynix, a key supplier of HBM to Nvidia and other global clients, is reportedly scaling back its aggressive investment plans in thermo-compression bonders (TCB), equipment crucial for HBM production. Industry watchers say this shift could reshape the competitive landscape for suppliers such as Hanmi Semiconductor, even as the broader market for bonding equipment is set for strong growth in the coming years.
Micron Technology is reportedly deepening its ties with the Trump administration to secure a stronger foothold in the global semiconductor market. According to South Korean media, this move could intensify competitive pressure on non-US rivals such as Samsung Electronics and SK Hynix.
The global memory industry saw simultaneous growth in volume and pricing during the third quarter of 2025, with DRAM and NAND memory prices rising amid ongoing supply constraints. Market participants anticipate that the DDR4 supply-demand imbalance will persist for another two to three quarters, leading to price increases for both DDR4 and DDR5 into early 2026. Chinese manufacturers have also raised NOR Flash prices, while Taiwanese suppliers are preparing price hikes starting in the fourth quarter to offset rising costs.
Nvidia has reportedly requested Samsung Electronics Co. to significantly expand its procurement of GDDR7 graphics memory, a move expected to enhance Samsung's memory business revenue. South Korean sources indicate that while Samsung currently trails in supplying fifth-generation high bandwidth memory (HBM3E), its exclusive GDDR7 supply and mass production capabilities could help secure its position in Nvidia's upcoming sixth-generation HBM (HBM4).
By September 2025, half-year earnings reports from 10 Chinese memory firms are in. Nine posted revenue growth, underscoring solid top-line momentum. Profitability, however, diverged sharply: companies could generate sales but struggled to convert them into lasting profits.
The rising demand for AI servers and high-end GPUs has fueled growth in large-size 2.5D interposers, and Hanmi Semiconductor is moving quickly to capture this opportunity. At SEMICON Taiwan 2025, the South Korean company debuted its "2.5D big-die TC/FC bonder," designed for AI chips and capable of packaging interposers up to 120mm x 120mm—over six times larger than conventional semiconductor sizes.
