# 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.

A Trillion-Dollar Frenzy for Memory Sellers, Halved Profits for Memory Buyers

Summary: A stark divide has emerged in the tech industry. While memory chipmaker Micron's stock soared 19% in a single day, pushing its market cap over $1 trillion, smartphone manufacturer Xiaomi reported a 43% plunge in adjusted net profit. The core driver is a severe supply crunch in memory chips, particularly for AI applications. Wall Street analysts, led by UBS and its unprecedented 204% target price hike for Micron, argue that long-term agreements (LTAs) from AI cloud giants are fundamentally ending the sector's notorious boom-and-bust cycles, justifying a re-rating from cyclical to infrastructure-like valuations. However, the "storage" market is now fragmented into three tiers. The first, AI-grade memory like HBM and server DDR5, faces extreme shortages and soaring prices driven by massive cloud capex. The second, mobile memory for smartphones, is also seeing sharp price hikes as manufacturers like Xiaomi are forced to pay more for remaining capacity, severely squeezing their margins. The third, PC retail channels, shows price declines due to existing inventory. The article questions the sustainability of the "supercycle" narrative. It highlights that Micron's revenue surge is driven almost entirely by price increases, not shipment volumes, making it vulnerable to a potential demand slowdown. While LTAs may dampen volatility, history suggests they are often tested during downturns. The current peak earnings, used to justify high valuations, represent a classic cyclical top. The piece concludes with a note of caution: when the entire Street chants "this time is different," it's wise to remember past bubbles, even as it acknowledges AI demand may indeed be structural.

marsbit05/27 11:57

A Trillion-Dollar Frenzy for Memory Sellers, Halved Profits for Memory Buyers

marsbit05/27 11:57

Trillion-Dollar Euphoria for Memory Sellers, Halved Profits for Memory Buyers

Title: The Trillion-Dollar Memory Seller's Carnival vs. The Buyer's Halved Profits On May 26, a stark contrast unfolded. While memory chipmaker Micron's market cap surged past $1 trillion, smartphone maker Xiaomi reported plummeting profits. Xiaomi's Q1 2026 profits fell 43% year-on-year. Executive Lu Weibing cited memory prices quadrupling from last year, adding roughly $210 to a phone's cost. To survive, Xiaomi is cutting entry-level models, sacrificing volume. Micron's stock, however, skyrocketed over 19% in a day, capping an 8x gain in a year. Major banks like UBS and JPMorgan issued bullish reports, raising price targets drastically. Their core thesis: Long-Term Agreements (LTAs) with AI cloud giants (Microsoft, Google, etc.) are eliminating the memory industry's notorious boom-bust cycle. By locking in fixed-price, multi-year contracts for AI-grade memory (HBM, server DDR5), these deals promise stable, utility-like earnings, justifying a higher valuation (20-30x P/E vs. the historical 8-15x). The article reveals a three-tiered memory market in 2026: 1) **AI Storage (HBM/DDR5/Enterprise SSD)**: Extreme shortage, soaring prices, LTAs. This is Micron's story. 2) **Mobile/Embedded Memory**: Also facing sharp price hikes as AI production crowds out capacity, severely pressuring phone makers like Xiaomi. 3) **PC Retail**: Some spot prices are falling due to channel inventory liquidation, creating a divergence from contract markets. The author questions if LTAs truly end the cycle. It hinges on sustained, hyper-growth AI demand. Micron's current profits are at a cycle peak, driven mostly by price hikes, not volume. If AI capital expenditure growth slows, the massive industry capacity expansion (e.g., Micron's $250B+ CapEx plan) could lead to a glut. Historically, using peak-cycle earnings for valuation is a classic trap. While the AI-driven structural shift might be real, the unanimous Wall Street euphoria warrants caution, echoing past bubbles like Cisco's in 2000. The memory seller's trillion-dollar狂欢 (carnival) continues, but the cycle's shadow remains.

链捕手05/27 11:48

Trillion-Dollar Euphoria for Memory Sellers, Halved Profits for Memory Buyers

链捕手05/27 11:48

DeepSeek's $10 Trillion Path: Leveraging Open Source to Pivot a Trillion-Dollar Hardware Ecosystem

This article proposes a strategic analysis of DeepSeek, arguing its goal is not short-term application-layer monetization but a long-term, foundational play to reshape the AI hardware ecosystem. It posits that DeepSeek's series of architectural innovations—MoE, MLA, DSA, CSA, Engram, TileLang—are fundamentally designed to drastically reduce the computational and memory (especially HBM) requirements for training and serving state-of-the-art AI models. By making AI feasible on hardware with lower performance (e.g., using NAND/SSD for KV Cache, LPDDR for weights and Engram storage), DeepSeek aims to foster and benefit from a viable alternative AI hardware supply chain, particularly in China. This strategy could unlock a trillion-dollar valuation for DeepSeek by enabling and owning a stake in a new, massive hardware ecosystem, rather than competing directly on subscriptions or multi-modal features.

marsbit05/25 13:13

DeepSeek's $10 Trillion Path: Leveraging Open Source to Pivot a Trillion-Dollar Hardware Ecosystem

marsbit05/25 13:13

Agentic Design Patterns: A Book That Made Me Re-Understand "What Is an Agent, Really?"

"Agentic Design Patterns" is a 2025 book by Antonio Gullí, a Google engineering director, which offers a systematic framework for AI Agent development through 21 design patterns. A core contribution is the "Four Levels of Agency": Level 0 (bare LLMs) are not true agents. Level 1 agents actively decide when and how to use tools. Level 2 agents engage in strategic planning, context engineering (curating and filtering information), and self-reflection. Level 3 involves multi-agent collaboration with defined communication topologies. The book introduces **Context Engineering** as a superset of prompt engineering, managing four layers of information for the agent: system prompts, external data, implicit context (user history, environment), and feedback loops for automated optimization. A key pattern is **Reflection (Producer-Critic)**, where two distinct agents with different prompts collaborate iteratively—one produces output, the other critiques it—until quality is satisfactory or a max iteration limit is reached. For **Memory**, a three-layer model is proposed: Session (ephemeral conversation context), State (temporary task data), and Memory (persistent, long-term storage). Regarding **Multi-Agent Systems**, the book advises against unnecessary complexity, recommending simple topologies like Supervisor or Peer-to-Peer based on task needs. It emphasizes perfecting a single Level 2 agent before moving to multi-agent setups. The author concludes with three actionable takeaways: 1) Add a Critic agent to existing workflows, 2) Practice Context Engineering beyond simple prompts, and 3) Avoid premature multi-agent complexity; first master a robust single agent. The book provides a practical map, codifying common challenges like reflection, memory, and coordination into reusable patterns, saving developers from reinventing foundational solutions.

链捕手05/25 04:43

Agentic Design Patterns: A Book That Made Me Re-Understand "What Is an Agent, Really?"

链捕手05/25 04:43

AlphaGo's Creator Puts AI into a 23-Year-Old Artificial Society: All Three Toughest Challenges for AI Agents Are Here

Demis Hassabis, CEO of DeepMind, has embarked on a new AI research venture by partnering with the long-running space MMO, EVE Online. This collaboration, announced in early May, aims to use the game's 23-year-old, player-driven persistent universe as a testbed for tackling three core challenges in AI agent research: long-horizon planning, memory, and continual learning. Unlike previous DeepMind environments like AlphaGo (Go) or AlphaStar (StarCraft II), EVE Online features no fixed end state. Its single-shard universe has fostered complex, emergent player societies with real economies, political alliances, and wars that can span months or years. These conditions naturally demand the very skills—long-term strategic planning, maintaining memories over extended periods, and adapting to constant change—that are hardest for current AI agents to master. The research will initially use an offline version of EVE, providing a controlled, complex sandbox without interfering with the live player server. This move continues DeepMind's trajectory of using increasingly complex and open-ended virtual worlds for AI training, from Atari games and Go to StarCraft II and the SIMA project. The EVE environment represents a significant step towards testing AI in a persistent, socially complex, and continuously evolving world shaped by human behavior over decades.

marsbit05/25 00:08

AlphaGo's Creator Puts AI into a 23-Year-Old Artificial Society: All Three Toughest Challenges for AI Agents Are Here

marsbit05/25 00:08

From Token Explosion to Physical Bottlenecks: The Storage Bull Market Driven by Agentic AI

**From Token Explosion to Physical Bottlenecks: The Agentic AI-Driven Storage Bull Market** The AI semiconductor narrative is shifting from training to inference, which now accounts for 66% of AI compute. In the inference "Decode" phase (autoregressive token generation), GPU performance is bottlenecked by memory bandwidth and capacity, not raw compute (FLOPS). The key constraints are **HBM (High Bandwidth Memory) bandwidth** (determining token generation speed) and **HBM capacity** (determining how many requests/models can be served simultaneously). This creates a core economics equation: Token cost is proportional to (GPU + power cost) divided by Tokens/sec, which is fundamentally limited by HBM specs. This drives unprecedented demand for advanced storage. **HBM**, a 3D-stacked DRAM, is critical for AI accelerators. Its complex production consumes 3-4x more wafer capacity than standard DRAM, squeezing supply for traditional memory (DDR) and causing severe shortages. **HBF (High Bandwidth Flash)**, an emerging high-bandwidth NAND, aims to bridge the gap between HBM speed and SSD capacity for AI model weights. The market is experiencing a historic, structurally driven super-cycle. Demand is fueled by a triple engine: 1) AI training (parameter arms race), 2) AI inference explosion (especially Agentic AI with long contexts), and 3) general data center expansion. Supply is constrained by the HBM产能挤压 effect and the 2-3 year lead time for new fab capacity. Analysts project a DRAM supply deficit of ~5% in 2026. Inventory across the supply chain is at historically low levels, with OEMs securing long-term agreements (LTAs) locking in future supply. Current indicators (Q2 2026) suggest the cycle is in its mid-phase, not peaking. While spot prices have corrected from highs, contract prices are forecast to rise sharply (e.g., +70-75% QoQ for NAND). Capacity utilization remains high, and inventory days are still low. The cycle is expected to peak around mid-2027. The storage landscape is stratified, with key players in HBM (SK Hynix, Samsung, Micron), NAND/SSD/HBF (Samsung, Kioxia/WD, SanDisk), and NOR Flash (Winbond, GigaDevice) well-positioned for this AI-driven era.

marsbit05/22 03:41

From Token Explosion to Physical Bottlenecks: The Storage Bull Market Driven by Agentic AI

marsbit05/22 03:41

Decoding the Latest Holdings of "The Son of Wall Street Version" Leopold Aschenbrenner: Why Did the AI Bull King Turn to Short Nvidia?

Leopold Aschenbrenner, a renowned AI investor and former OpenAI researcher, has filed a 13F revealing a dramatic portfolio shift. He has taken a massive $8 billion short position against the entire semiconductor supply chain, including companies like NVIDIA, AMD, Broadcom, and ASML. This marks the first time in his fund's history that short exposure exceeds long exposure. His core thesis is that the bottleneck for AI investment is shifting from chip *design* to *infrastructure*, specifically power and memory. Accordingly, his long positions heavily target these areas. He maintains holdings in data center/Neocloud firms like CoreWeave, holds positions in power solutions like Bloom Energy, and has initiated new stakes in bitcoin mining companies such as CleanSpark and Riot Platforms due to their pre-existing grid access and power capacity. Aschenbrenner also increased his long bets on memory, exemplified by SanDisk, as AI models drive demand for NAND flash storage. Despite the aggressive short on semis, analysts suggest he may not be structurally bearish but rather views the chip trade as over-crowded, while seeing higher returns in infrastructure plays. Key risks identified include underestimating NVIDIA's entrenched CUDA ecosystem moat and the timing of his shorts.

marsbit05/20 09:50

Decoding the Latest Holdings of "The Son of Wall Street Version" Leopold Aschenbrenner: Why Did the AI Bull King Turn to Short Nvidia?

marsbit05/20 09:50

How to Become a Pro Claude User in 30 Days?

"How to Become a Claude Power User in 30 Days" outlines a structured month-long program to transform from a casual user into someone who leverages Claude as a core productivity system. The first week focuses on mastering prompt structure (Role, Context, Task, Format, Constraints), understanding context windows, and setting up foundational Projects and Memory for personalized context. The second week involves building reusable workflows for research, writing (using a two-step outline-first process), and decision-making. The third week shifts to automation: using Claude Cowork for autonomous file tasks, connecting tools like Google Drive and Slack, and setting scheduled automation. The final week is for compounding gains: optimizing all workflows based on feedback, building a personal knowledge repository, teaching the system to a colleague, and designing a complete, ideal "Claude Operating System." By day 31, the user operates with an autonomous assistant that handles routine tasks with deep context, freeing them for high-value creative and strategic work. The core message is that consistent, systematic configuration over 30 days creates a powerful, personalized productivity advantage.

marsbit05/20 08:06

How to Become a Pro Claude User in 30 Days?

marsbit05/20 08:06

Five Core Forms of AI Agent in YC's Eyes

The article outlines five core architectural patterns for effective AI Agents, emerging from tools like Codex and Claude, that move beyond simple prompts towards reusable, process-based capabilities. 1. **Skills**: Reusable, parameterized workflows that function like method calls, allowing a single process (e.g., "/investigate") to handle various tasks based on input parameters. 2. **Thin Harness**: A lightweight execution framework (~200 lines) that manages the AI model's "hands and feet"—handling loops, file I/O, and context—without becoming bloated. 3. **Resolvers**: Routing tables that map tasks to specific Skills, preventing "context corruption" when managing dozens of Skills and ensuring outputs go to the correct locations. 4. **Latent vs. Deterministic Layer**: A critical separation where LLMs handle judgment, synthesis, and pattern recognition, while deterministic code handles tasks requiring precision, consistency, and low cost (like calculations). 5. **Memory**: A persistent, accumulating knowledge base (e.g., a markdown folder) with a "current trusted conclusion" section and an append-only timeline, enabling the system to learn and retain context over time. Together, these patterns create a "process power"—a durable competitive advantage. Unlike one-off prompt-based applications whose value quickly commoditizes, a well-designed AI Agent system encodes experience into reusable, parameterized workflows, offloads stable rules to code, and continuously learns through memory. This creates a structured, hard-to-replicate capability that can provide sustained value for individuals or businesses, such as an accountant automating client reviews while preserving privacy and accumulating expertise.

marsbit05/20 07:46

Five Core Forms of AI Agent in YC's Eyes

marsbit05/20 07:46

Samsung Bets on Mobile HBM: AI Moves from Cloud to Palm, a New Frontier in Semiconductor Investment?

Samsung is betting on bringing high-bandwidth memory (HBM) technology from servers to mobile devices, aiming to enable powerful on-device AI features in smartphones and tablets. This move is driven by the booming AI market, where HBM demand from data centers has fueled Samsung's record profits, with HBM4 already in mass production. By integrating mobile HBM, Samsung seeks to transform user AI experiences—making tasks like image generation and real-time translation faster, seamless, and more private by processing data locally. Strategically, this allows Samsung to leverage its vertical integration in memory, advanced packaging, and Exynos processors to differentiate its Galaxy devices against competitors like Apple and Qualcomm. It also opens a new consumer growth avenue, reducing reliance on volatile server HBM demand alone. The initiative is expected to benefit the broader supply chain, boosting demand for advanced packaging materials, thermal solutions, and other components. While promising, risks include potential delays in mobile HBM mass production beyond 2027, high initial costs, and the cyclical nature of the memory market. Nonetheless, Samsung's push signals a broader industry shift toward hybrid cloud-edge AI computing, positioning it as a key player in defining the future of AI-powered devices and presenting a potential long-term investment theme in semiconductors.

marsbit05/19 14:49

Samsung Bets on Mobile HBM: AI Moves from Cloud to Palm, a New Frontier in Semiconductor Investment?