# Memory Related Articles

HTX News Center provides the latest articles and in-depth analysis on "Memory", covering market trends, project updates, tech developments, and regulatory policies in the crypto industry.

JP Morgan Research Report Analysis: SK Hynix Fell 15% Last Week, Concerns Over HBM Pricing Overblown and Shareholder Returns to Land Ahead of Schedule

JPMorgan Research Report Interpretation: Addressing Concerns on SK Hynix's Recent Share Price Decline SK Hynix's stock fell 15% last week, underperforming the KOSPI (-5%) and Samsung (-9%). Key investor concerns centered on HBM pricing uncertainty, unclear shareholder return timelines, and a recently disclosed ₩54 trillion capital expenditure plan. JPMorgan's August 9 report addresses each point. Regarding HBM, JPMorgan refutes inaccurate media reports suggesting potential 50% price discounts for HBM4 in 2027. The firm's conservative model assumes <40% average HBM price growth by 2027, based on factors including memory suppliers prioritizing high-margin DDR5/LPDDR5/NAND in LTA allocations and the long-term partnership with key customer NVIDIA. The most significant near-term catalyst is shareholder returns. SK Hynix has committed to announcing additional shareholder return measures before the end of Q3 (September), earlier than its prior "year-end" guidance. JPMorgan anticipates a progressive policy, supported by an estimated cumulative free cash flow exceeding ₩800 trillion over three years and proceeds from the Kioxia stake sale. The substantial ₩54 trillion capex plan is for two new memory fabs to support its 2030 roadmap, not short-term aggressive expansion. The Yongin Y2 DRAM fab (₩35.2tn) and Cheongju M17 NAND fab (₩19.1tn) have construction starts scheduled for 2027, with clean room completion and investments stretching to 2028-2031. JPMorgan sees limited strategic value in a potential IPO for subsidiary Solidigm, noting SK Hynix's strong internal cash flow can fund capex without dilution, and an IPO might trigger dual-listing rule constraints in Korea. The report concludes last week's sell-off was an overreaction. JPMorgan maintains its Overweight rating and ₩2.75 million price target (~7x avg. 2026-2027 EPS), asserting the memory super-cycle thesis and SK Hynix's fundamentals remain intact.

marsbit6h ago

JP Morgan Research Report Analysis: SK Hynix Fell 15% Last Week, Concerns Over HBM Pricing Overblown and Shareholder Returns to Land Ahead of Schedule

marsbit6h ago

BofA Research Report Analysis: Storage Super Cycle Extends to 2027, Samsung May Launch Large-Scale Buyback and Special Dividend

Bank of America Research Report: Storage Supercycle Extends to 2027, Samsung Poised for Major Buyback and Special Dividend Bank of America's latest global memory tech report (August 7) forecasts a continued supercycle in the memory chip sector, extending to 2027. DRAM industry sales are projected to reach $573 billion in 2026 and $847 billion in 2027. NAND sales are expected to hit $357.5 billion in 2026 and $467.5 billion in 2027. HBM demand is a core driver, with its market size forecast to surge to $152.8 billion in 2027. Spot prices for NAND showed an unexpected strong rebound, with 256Gb products rising over 10%, while DRAM spot prices also increased 1-2% weekly. Supply for both remains tight. Samsung Electronics is anticipated to announce significant capital returns, potentially including a special dividend exceeding KRW 30 trillion and a share buyback over KRW 40 trillion, funded by 50% of its 2024-2026 free cash flow. SK Hynix is also expected to return 50% of FCF to shareholders. The supercycle is shifting from price-driven to volume-and-price driven. While ASPs may see a mild correction in 2028, sustained HBM adoption and AI server demand are reshaping industry profitability. The tight supply-demand balance is unlikely to reverse before 2027.

marsbitYesterday 04:43

BofA Research Report Analysis: Storage Super Cycle Extends to 2027, Samsung May Launch Large-Scale Buyback and Special Dividend

marsbitYesterday 04:43

Karpathy Says It Will Take Another Decade, But This Road Is Already Packed

The article discusses the intensifying focus on **Continual (or Lifelong) Learning** for large language models (LLMs), a capability seen as crucial for developing truly useful "AI colleagues." It references Andrej Karpathy's 2025 assessment that overcoming LLMs' lack of persistent memory and learning might take a decade. The core challenge is **catastrophic forgetting**, where learning new information erases previously acquired skills. Current research diverges into several technical paths: 1. **External Memory Systems:** Storing new knowledge in external databases (e.g., MemGPT, Letta's approach, Karpathy's "LLM Wiki"), akin to enhanced RAG. This is safe but doesn't "internalize" knowledge. 2. **Context Engineering:** Evolving the input context itself into a growing "playbook," as seen in ACE (Agentic Context Engineering), which uses execution feedback to refine instructions without weight updates. 3. **Continual Post-Training:** Carefully updating model weights (e.g., via LoRA) to internalize knowledge, using techniques like Self-Distillation Fine-Tuning (SDFT) to mitigate forgetting. 4. **Continual Pre-training:** Updating the base model with new corpus data, which is compute-intensive and prone to forgetting. 5. **Novel Paradigms:** More radical approaches redefining learning itself. These include models that generate their own training data and update instructions (e.g., SEAL), architectures with nested, multi-timescale learning (e.g., Google's Nested Learning/Hope), and the vision of an "Era of Experience" where AI learns from self-generated interaction data. The article suggests a pragmatic, hybrid future is likely: short-term, mutable knowledge handled by external memory/context, while long-term, solidified capabilities are encoded into model parameters via fine-tuning. While catastrophic forgetting remains unsolved, the field has evolved from a theoretical gap into an active arena with multiple competing approaches and emerging commercial products, all racing to bridge the gap to practical, continually learning AI agents.

marsbitYesterday 02:33

Karpathy Says It Will Take Another Decade, But This Road Is Already Packed

marsbitYesterday 02:33

AMD acquires Taalas: hardware AI manages without scarce HBM memory

AMD has agreed to acquire Toronto-based startup Taalas, which tackles a key bottleneck in AI inference: the constant need to transfer model weights from memory to the processor for each generated token. Taalas's chips eliminate this operation by permanently embedding the model weights into the transistors themselves. This data transfer is what currently limits inference speed and has made high-bandwidth memory (HBM) a scarce commodity. Taalas's first test chip, fabricated on TSMC's 6nm process, reportedly generated tokens for Meta's Llama 3.1 8B model at speeds 48 times faster than comparable Nvidia GPUs. Its architecture features a mask ROM section for fixed weights and SRAM for adaptable components. However, this design comes with a significant trade-off: each chip is permanently dedicated to a single model. Switching models requires a partial redesign and fabrication, a process taking about two months. While the acquisition is seen as part of AMD's rivalry with Nvidia in inference, its broader implication lies in challenging the assumption of a permanent HBM memory shortage. The AI memory market is currently booming, with HBM supply sold out through 2026. Yet, Taalas's technology demonstrates that the memory bottleneck is an engineering challenge, not an absolute physical constraint. This aligns with industry-wide efforts from companies like Nvidia (through model compression) and memory makers like Samsung and SK hynix (developing new packaging and storage technologies) to reduce dependency on scarce HBM. AMD's move suggests that the current high prices for memory, driven by AI demand, may not be sustainable. It highlights a growing engineering push against the premise of perpetual memory scarcity, reminding investors that memory has historically been a cyclical business.

cryptonews.ruYesterday 14:56

AMD acquires Taalas: hardware AI manages without scarce HBM memory

cryptonews.ruYesterday 14:56

Why the Memory Chip Shortage Is Reshaping Portfolios Across Semiconductors in 2026

The global memory chip shortage, driven by AI infrastructure demand, is reshaping markets in 2026. NAND and DRAM contract prices have surged up to 90% as production capacity shifts from consumer chips to high-bandwidth memory for AI. This has impacted consumer electronics, with even Apple raising prices. DRAM is central to the shortage, essential for all AI hardware, and its price action has made it a tradable commodity. Micron holds a key position as its high-bandwidth memory for AI accelerators is sold out into 2027, driving record revenue and making its stock a direct trade on the scarcity premium. SanDisk benefits similarly from the NAND shortage, with long-term supply agreements locking in elevated pricing. The core dynamic is a multi-year mismatch between surging AI demand and slow memory supply expansion, with supply agreements extending to 2027. For traders, the key question is the duration of elevated pricing. Platforms like WEEX now offer access to these assets via USDT-settled products, providing consolidated access to this cross-market narrative.

TheNewsCryptoYesterday 11:31

Why the Memory Chip Shortage Is Reshaping Portfolios Across Semiconductors in 2026

TheNewsCryptoYesterday 11:31

All Metrics Smashing Records, Yet Stock Prices Plunge Across the Board

Memory giants like Western Digital (WDC) and SanDisk (SNDK) reported blockbuster earnings in the summer of 2026, featuring毛利率 exceeding 80%, massive customer prepayments, and long-term supply agreements. Despite this seemingly perfect performance, their stocks plummeted post-earnings (WDC down 13%, SNDK down 7%), along with peers like Micron. The collapse highlights a core market rule: "good" isn't enough; results must beat already sky-high expectations. With valuations at peak "perfect asset" levels, even slightly conservative forward guidance triggered a sell-off. The market saw "peak performance" as a signal to exit. Beneath the stellar numbers, four反常 trends emerged: 1. **Financialized Pricing:** Customers provide百亿级 in upfront "interest-free deposits" to secure future capacity. 2. **Reversed Cost Curve:** Advanced DRAM (HBM4, DDR6) costs are rising per bit due to complex packaging, breaking Moore's Law. 3. **AI vs. Consumer Split:** Data center storage demand soars (+103% for SanDisk), while consumer electronics demand weakens under high costs. 4. **HDD Revival:** Hard drives, now used for AI agent context caching, see毛利率 near 55-57%. Underlying隐忧 persist. Soaring capital expenditure (CapEx) by SK Hynix and Micron risks future oversupply. Revenue growth is increasingly driven by price hikes, not surging shipment volumes (bit growth), making profits vulnerable to any price correction. In conclusion, while AI has created a long-term growth narrative, transforming storage into "strategic infrastructure," the market's violent reaction signals that peak valuations and expectations have left no safety margin. The周期 hasn't disappeared; it's merely wearing an AI disguise.

marsbit08/07 10:21

All Metrics Smashing Records, Yet Stock Prices Plunge Across the Board

marsbit08/07 10:21

ChangXin's "Peer": The Fate of Fujian Jinhua Integrated Circuit Co., Ltd. Is Regrettable

China's DRAM industry saw a pivotal moment with ChangXin's (CXMT) successful IPO. However, the fate of its 2016 counterpart, Fujian Jinhua Integrated Circuit, offers a stark contrast. Both were founded the same year with similar missions, massive investment, and 12-inch wafer fab goals to break into the DRAM market dominated by Samsung, SK Hynix, and Micron. Fujian Jinhua initially progressed faster by partnering with Taiwan's United Microelectronics Corporation (UMC) for 32nm DRAM technology. This strategy, however, led to a protracted legal battle. In 2017, Micron sued UMC and Jinhua for trade secret theft. The situation escalated in October 2018 when the U.S. Commerce Department added Fujian Jinhua to its Entity List, citing its imminent mass production as a threat. This resulted in an immediate halt of equipment, software, and technical support from American suppliers, followed by UMC suspending cooperation. Although Jinhua was eventually cleared of criminal charges in late 2023 after a nearly six-year legal saga, it missed the critical industry growth window. In contrast, ChangXin took a different path from the start, focusing on building its own R&D system and securing intellectual property, notably through a license for former Qimonda patents. While also facing U.S. scrutiny and initial heavy losses, ChangXin benefited from a more mature domestic supply chain when it reached mass production. It achieved profitability in 2025 and represents the rise of China's DRAM industry. Jinhua's story is a crucial lesson. It was the first Chinese DRAM company to confront the complex realities of international IP disputes, export controls, and supply chain vulnerabilities. Today, it has resumed operations with a 40,000 wafers-per-month capacity, aiming for 60,000. While it missed its initial opportunity, its experience informed the strategic evolution of later Chinese semiconductor firms.

marsbit08/05 12:31

ChangXin's "Peer": The Fate of Fujian Jinhua Integrated Circuit Co., Ltd. Is Regrettable

marsbit08/05 12:31

Despite the sell-off, Goldman Sachs remains bullish on Samsung and SK Hynix. Here's why.

Goldman Sachs maintains "Buy" ratings on Samsung Electronics and SK Hynix despite recent stock declines. Its bullish view centers on three core arguments. Firstly, it expects HBM (High Bandwidth Memory) pricing to re-establish a premium over conventional DRAM by 2027, with a projected blended ASP of around $2.9/Gb. This is driven by tight supply-demand dynamics, increasing manufacturing complexity for newer HBM generations, and the need to restore its historical price premium. Secondly, the volatility of the memory cycle is expected to moderate due to widespread adoption of 3-5 year Long-Term Agreements (LTAs) with key customers. These contracts, covering a significant portion of planned capacity, feature mechanisms like price floors, prepayments, and penalties, reducing supplier risk and improving earnings visibility. Thirdly, inventory levels remain low at key suppliers and major customers, providing a buffer against a sharp downturn. Furthermore, robust demand from enterprise SSDs for AI servers is seen offsetting weakness in consumer segments like smartphones and PCs, preventing NAND markets from slipping into oversupply in the near term. While risks exist—such as potential weaker AI demand or aggressive capacity expansion—Goldman Sachs believes the combination of HBM repricing, LTAs, and low inventory underpins a more stable earnings outlook for the leading Korean memory makers.

marsbit08/05 04:11

Despite the sell-off, Goldman Sachs remains bullish on Samsung and SK Hynix. Here's why.

marsbit08/05 04:11

Kimi K3, which used to require 16 B200s, now fits on just 8 AMD cards

This article highlights a key achievement for AMD in the AI hardware race. The company's MI355X GPUs, each with 288 GB of memory, successfully deployed the massive 2.8 trillion parameter Kimi K3 model on a single 8-GPU server. In contrast, the NVIDIA B200 (with 192 GB per card) required a two-server, 16-GPU setup to hold the model, leading to inter-node communication overhead. In performance tests for a 1024-input/400-output token task, the 8-card MI355X system achieved a total throughput of 952 tokens/s and a single-user generation speed of 118 tokens/s. This single-node throughput was approximately 3.8 times higher than the per-node average of the dual-node B200 setup (498 tokens/s total). While NVIDIA's B300 delivered higher absolute performance (1568 tokens/s on 8 cards), a cost-efficiency analysis based on assumed hourly rates showed MI355X offered better value per dollar. Notably, the deployment on AMD's ROCm software platform was relatively straightforward, requiring only minor fixes like patching a missing function for speculative decoding and a simple zero-padding workaround to optimize a prefill kernel for attention heads. This significantly reduced the Time-To-First-Token (TTFT). The article concludes that for extremely large models, memory capacity is becoming a critical differentiator. AMD's strategy of equipping cards with more HBM memory provides a tangible system advantage in single-node deployment efficiency and cost, posing a growing challenge to NVIDIA's CUDA ecosystem dominance.

marsbit08/04 11:22

Kimi K3, which used to require 16 B200s, now fits on just 8 AMD cards

marsbit08/04 11:22

SK Hynix and SanDisk Introduce a New Type of Intermediate Memory

On August 4, 2026, at the FMS conference in California, SK Hynix and SanDisk unveiled the first specifications for HBF, a new memory technology designed to bridge the gap between high-speed HBM memory and SSDs. Key HBF specs include a capacity of up to 512GB, bandwidth ranging from ~0.4 TB/s to 3.0 TB/s, and connectivity via the open UCIe interface for compatibility with various CPUs/GPUs. The specifications are released through the Open Compute Project, positioning HBF as an open industry standard rather than a proprietary one. This introduction suggests future AI servers will likely adopt a four-tier memory hierarchy: 1) HBM for the hottest data, 2) CXL for accelerating classic RAM, 3) HBF as a high-bandwidth, high-capacity layer using cost-effective NAND, and 4) NVMe SSDs for cold storage. In LLM infrastructure, HBF is intended to store large AI model context and weights that don't fit into the more expensive HBM and CXL memory. The HBF standardization partnership began in August 2025, with a consortium including Google and Tenstorrent launching in February 2026. Separately, SK Hynix also previewed its developing 375-layer 4D NAND (V10) at FMS, claiming a 2.5x performance-per-watt improvement, with mass production of eSSDs planned for early next year.

cryptonews.ru08/04 11:18

SK Hynix and SanDisk Introduce a New Type of Intermediate Memory

cryptonews.ru08/04 11:18

活动图片

Industry Newsright-arrow

Market Analysisright-arrow