Bank of America said in a weekend memory industry survey that Samsung Electronics has already placed 60% to 70% of its memory sales under long-term supply agreements, with terms that favor suppliers. According to the report, quarterly price cuts are usually limited to no more than 5%, while price increases can range from 10% to 20% and are effectively uncapped. The bank said most LTAs signed with major U.S. technology companies use a five-year rolling structure that can be renewed around the end of the first contract year, creating long-term customer lock-in. BofA said the arrangement improves revenue visibility for Samsung’s memory business while preserving pricing flexibility during tight supply cycles. It also pointed to continued demand growth from AI servers and the time still required for memory makers to expand output. Separately, DRAMeXchange data showed sharp gains across spot and contract markets, including 16Gb DDR5 spot prices rising 733% year over year to $51, NAND 1Tb wafer spot prices climbing 415% to $26.4, and 64GB DDR5 module contract prices moving above $1,480. The bank said August pricing is being supported by stronger restocking demand, OEM inventory builds for new products, and falling channel inventories.

Fujian Jinhua Integrated Circuit Co. Ltd., once seen as one of China’s three major memory projects alongside Yangtze Memory Technologies and ChangXin Memory Technologies, is back in focus after years of disruption. The company was hit by a U.S. export blacklist and criminal charges just as its DRAM production line was coming together, bringing progress to a halt and pushing it out of the public eye for years. That legal overhang began to lift at the end of 2023, when Micron Technology reached a global settlement with Fujian Jinhua and both sides withdrew lawsuits worldwide. On Feb. 27, 2024, a federal court in San Francisco ruled that prosecutors had failed to prove Jinhua stole Micron trade secrets. The decision cleared the company of the charges that had defined much of its recent history. The article traces Jinhua’s rise from a state-backed strategic project launched in Jinjiang in 2016, its early technology partnership with United Microelectronics Corp. (UMC), the central role of executive Chen Zhengkun, and the impact of years of sanctions on its production roadmap. It also places the company inside the economics of the DRAM industry, where scale, capital intensity, yield management and intellectual property disputes have long shaped the global pecking order. Jinhua is still far behind larger Chinese peers. Its 12-inch fab is producing about 40,000 wafers a month, with expansion to 60,000 planned for 2026, and the company remains on the U.S. entity list. But with local state backing, a niche DRAM focus and more than 1,007 related patents, it has not dropped out of the race.

Samsung Electronics and SK Hynix are accelerating their investment in high bandwidth memory, or HBM, and other next-generation memory technologies as the two South Korean chipmakers work to preserve their lead in advanced memory. According to ChainCatcher, Samsung said its second-quarter research and development spending hit a record 16 trillion won, roughly $11.6 billion. The company also said it plans to expand its lineup of high-end memory products, including HBM4, HBM4E, and DDR5. SK Hynix, meanwhile, is developing next-generation technologies such as HBM5, hybrid bonding, and iHBM. The company is expected to spend 40 trillion won, or about $32.6 billion, in capital expenditures this year. The latest figures show both companies are continuing to commit large sums to memory development and manufacturing capacity as competition around advanced chips intensifies.

P Equity Research said in a post on X that UBS expects global cloud service providers to spend $190.6 billion on DDR memory in 2026, with that figure projected to climb to $448.9 billion in 2027. Based on the figures cited in the post, DDR memory spending would increase by about 135% year over year. UBS also expects overall memory spending in 2027 to rise by roughly 127%. The figures were shared by P Equity Research and cited by Odaily in a brief news update.

UBS said spending on storage is rising quickly as demand from AI GPU vendors and hyperscale cloud providers accelerates and average selling prices for storage products move higher. The bank estimates that the world’s top 11 hyperscale cloud service providers will spend a combined $1.0403 trillion in capital expenditures in 2027, with $761.3 billion of that going to storage, lifting storage’s share to 73%. UBS also outlined a steep climb in overall capex and storage spending over the next two years. Total capex is projected to rise from $504 billion in 2025 to $873.9 billion in 2026 and then to $1.0403 trillion in 2027. Over the same period, storage spending is expected to jump from $72.8 billion to $335.7 billion and then to $761.3 billion, with year-over-year growth of 361% in 2026 and 127% in 2027. By product, UBS expects HBM, DDR and NAND spending all to increase sharply, while average prices per Gb or GB also move up. The bank added that storage makers will need to expand capacity to meet demand, though fab space, equipment delivery times and labor constraints remain major bottlenecks.

Changxin Memory Technologies, or CXMT, is nearing the final stage of research and validation for its LPDDR6 product, according to BlockBeats, which cited Yicai and unnamed industry sources on Aug. 1. The development is described as a key step before mass production. Earlier market chatter in March said Changxin Storage, under Changxin Technology, was pushing ahead with LPDDR6 commercialization and had already sent samples to core customers. That report said the product could be released and moved into mass-production introduction in the second half of 2026. The same information outlined the first LPDDR6 product's specifications: a design data rate of 12,800 Mbps, a base operating rate of 10,667 Mbps in coordination with the system-on-chip, 16Gb die capacity, and a 1295-ball package using package-on-package stacking. It also said the chip shows clear improvements over LPDDR5X in low-power design and RAS features, referring to reliability, availability and serviceability.

UBS said in a July 29 research note that the market’s reset in SK Hynix shares is hard to justify, arguing that the company’s current valuation implies a long-term return profile far below its own forecasts. The bank said SK Hynix is trading at 1.66x 12-month forward price-to-book, a level that implies a long-term ROE of 18.9%, while UBS projects average ROE of 40.2% for 2027 through 2031. UBS kept its buy rating on the stock and cut its target price to KRW 3.0 million from KRW 3.2 million. The note said SK Hynix shares have fallen 52% from their June 22 peak, even though the stock is still up 115% for the year. UBS also pointed to faster-than-expected progress on long-term agreements, stronger demand tied to AI agents, and supply conditions that remain tight across memory markets. It estimated HBM shipments and capacity will keep growing into 2027, while free cash flow could support buybacks and broader shareholder returns. UBS said it has long expected SK Hynix to return 50% of free cash flow to shareholders through a mix of dividends and repurchases.

SK Hynix’s latest earnings miss is being read as more than a quarterly stumble. It points to a deeper change in the memory market as production keeps moving from commodity DRAM to high-bandwidth memory, or HBM, for AI accelerators. According to analyst Bubble Boi, the problem is straightforward: HBM pricing has been mostly flat in the near term, while spot prices for commodity DRAM have continued to rise. That leaves suppliers replacing a product with improving pricing by one that is no longer seeing the same uplift in average selling prices. The broader point is that the old “win on both sides” setup may be breaking down. For the past two years, rising HBM demand consumed DRAM wafer capacity, tightening DRAM supply and lifting prices at the same time. That gave memory makers high-margin HBM exposure plus DRAM price tailwinds. Bubble Boi argues the next phase looks different. As each new NVIDIA accelerator carries more HBM — from 80 GB on H100 to 148 GB on B200 and 288 GB on Rubin — the need for server DDR DRAM falls. At the lower end, flash offload using NAND is also starting to replace part of DRAM workloads in inference. In that setup, vendors with weaker HBM execution, especially Micron, may face the hardest pressure.