2026-07-29 每日信息简报

DAILY BRIEFING · 2026-07-29

AI 的资本账、智能体的工程账，以及开放权重的治理账

今天的主线不是某一款新模型，而是 AI 从“能力展示”进入资本开支、生产部署与安全治理同时被审视的阶段。

今日要点

AI 的成本开始影响估值

The Verge 报道，Google 上调了资本开支预期，区间达 1,950 亿至 2,050 亿美元；市场开始把算力投入的现金成本与 AI 叙事一起定价。值得注意的不是单季股价波动，而是“收入增长能否覆盖基础设施折旧与电力成本”正成为投资者的核心问题。与此同时，《经济学人》也指出，AI 收入虽快增，但与巨额投入之间仍有回报不确定性。

前沿模型公司把治理请求推向政府

OpenAI、Anthropic、Google、Meta 等实验室的员工联署，要求政府加快全球协调治理并考虑放缓最前沿系统的无序推进。这个信号的重点是：模型竞争已不只是企业自律议题，而在寻求可执行的外部规则与跨境协调。它与 Sam Altman 对安全事件的表态 一起，说明“安全成本”正从原则讨论进入经营与政策约束。

智能体进入科学计算与安全边界

OpenAI 的实地报告描述了研究者如何借助编码智能体维护遗留科学软件、推进基因组学等工作；可复现、验证与领域专家审查仍是关键闸门。另一面，《纽约时报》报道，Anthropic 的模型在弱化加密算法测试中发现攻击路径。两件事共同提示：智能体扩大了研究与工程杠杆，也要求更严格的隔离、评测和责任链。

AI 与科技

开放权重之争，从产品路线变成权力分配

围绕“开放权重”的争论，实质是能力、审计和扩散速度之间的取舍。Meta 的扎克伯格在采访中批评高度集中的 AI 权力、主张更开放的路径；但开放并不自动解决滥用、算力门槛或责任归属问题。读者应区分开放参数、开放代码、开放数据和开放部署权限，它们带来的风险面并不相同。

AI 智能体的身份与访问安全正在成为独立市场

Cyera 同意以 10 亿美元收购 Oasis Security，交易背景是企业内能访问系统和数据的智能体数量增加。真正需要管理的是机器身份、最小权限、凭证轮换和操作审计，而非只把代理当成新的聊天界面。对团队而言，先明确智能体可读取什么、可执行什么、谁可撤销权限，通常比追求更高自治更紧迫。

AI 工厂的约束从 GPU 延伸到电网与能效

数据中心扩张已让电网运营方考虑在压力时段临时限电；这让训练与推理的可用性变成基础设施问题。NVIDIA 的能效优化建议也将每瓦电的去向拆到数据、训练、推理与散热。单位 token 的成本、供电韧性和调度能力，会直接决定规模化应用的边界。

值得深读

从线性注意力回看 Kimi K3 的架构选择

这篇技术拆解沿 DeltaNet 系列线性注意力变体回溯 Kimi K3 的可能设计取舍，适合把“低成本高性能”的产品叙事落回算法与工程实现。阅读时可重点看复杂度、长上下文行为和训练稳定性之间的交换，而不要把任何单一架构标签当作性能结论。

医疗机器人为何需要 GPU 原生物理仿真

医疗机器人无法像消费互联网产品那样依赖海量线上试错；物理仿真因此承担了更高比例的验证责任。文章说明将医学物理仿真与 GPU 工作流结合的思路，值得关注其训练数据、仿真到现实迁移与临床验证三者如何衔接。

“数字身份验证”公共政策的反作用

一项欧洲公民倡议反对强制数字身份与年龄验证，理由包括隐私、可及性和互联网开放性。无论是否同意立场，它把一个经常被忽略的问题摆到台面：将合规责任前置到每个用户身份，可能改变匿名表达、数据集中化与平台进入门槛。

快速浏览

Hacker News 精选Codex SecurityOpenAI 的安全工具项目登上社区热议，适合结合代码仓库与讨论了解其目标边界。↗ The Verge AIPerplexity 将 Personal Computer 扩展到 Windows本地文件和应用访问使“通用数字员工”的权限模型更值得审视。↗ TechCrunch AIFish Audio 融资 5,200 万美元语音模型的开源与托管并行，商业化开始回到真实使用与收入指标。↗ Apple Machine Learning Research面向设备端的高效音频合成研究讨论在严格内存与算力预算内将语义音频 token 转为高保真语音。↗ 纽约时报 AIAI 生成图书如何进入书店零售商面对大量生成内容时的标注、筛选与分发政策正在分化。↗ The Verge AI开源模型托管与非自愿深伪风险报告将内容治理的难题落在模型发现、审核与下架机制上。↗ HelloGitHubHelloGitHub 第 124 期40 个开源项目横跨 C、Python、Rust、AI 与工具链，可作本周选型入口。↗ Longreads帝王蝶数量回升，但仍只是“暂时”一篇关于生态波动、保护成效与长期观察窗口的长读。↗

@Brad_Setser · 请求账号 @Brad_Setser · 2026-07-29 01:08:36 UTC · 原创It doesn't really pop in this chart but export volume growth (average of y/y changes reported by China) in q2 was around 7% and import volume growth was back under 1% (q1 data distorted by slow adjustment for rising chip prices) 1/2↗ @jukan05 · 请求账号 @jukan05 · 2026-07-29 00:51:46 UTC · 原创From SK hynix’s 2Q Earnings Call Q&A Q. What is the status of HBM price negotiations for 2027? What is the pricing outlook and the status of contract discussions for HBM3, HBM4, and the upcoming HBM4E? A. We are currently discussing 2027 HBM supply volumes and pricing with our major customers. Negotiations are proceeding smoothly, supported by robust customer demand. However, we ask for your understanding that we cannot disclose individual customers’ contract terms or pricing details at this time. Given the recent sharp increase in conventional DRAM prices, these market conditions could also influence HBM price negotiations. However, HBM pricing is not determined solely by movements in conventional DRAM prices. Compared with conventional DRAM, HBM requires substantially greater resources, including more wafers, advanced process technologies, TSV, packaging, and production capacity. With each new generation, customers’ performance and quality requirements become more demanding, while the development and qualification processes grow increasingly complex. Accordingly, we negotiate prices individually with customers by comprehensively considering not only conventional DRAM pricing and supply-demand conditions, but also the resources required to produce HBM, opportunity costs, technological complexity, and the value our products provide to customers. Our focus is on securing an appropriate level of profitability commensurate with the differentiated value we provide to customers, while supporting the healthy and sustainable growth of the AI ecosystem. Building on our accumulated technological and cost competitiveness, stable mass-production capabilities, strong customer trust, and close collaboration, we will maintain solid profitability in our HBM business. Through product transitions and greater customer value creation, we aim to establish ourselves as a key partner capable of growing alongside our customers in the AI era. Ultimately, we will continue to pursue sustainable long-term growth.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-29 00:35:37 UTC · 引用Samsung Electro-Mechanics to Raise MLCC Prices by 30% Starting August 1.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-28 23:36:14 UTC · 引用I don't doubt that China has invested heavily in modern manufacturing and moved up the value chain. That tho raises the question of why China needs the additional export boost from an undervalued yuan↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 23:09:25 UTC · 原创There’s a reason for the saying that, when it comes to memory semiconductors, you shouldn’t trust anyone but Korean analysts, lol. KIS is a veteran in this field. Not only KIS, but veteran analysts at Citi, Meritz, and elsewhere all lowered their earnings estimates for SK hynix.↗ @Claudia_Sahm · 请求账号 @Claudia_Sahm · 2026-07-28 23:08:20 UTC · 原创Everyone is watching what Warsh says at 2:30 tomorrow. Watch what the committee says at 2:00. The FOMC statement is where compromises get forged. I mocked up one I think they could agree to, redlined from June. Full post in the first reply.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 23:02:42 UTC · 原创Samsung Electro-Mechanics to Raise MLCC Prices - Samsung Electro-Mechanics has sent customers a notice announcing that it plans to raise prices for certain MLCC products by 30% from August 1, 2026. - The company explained that structural growth in global electronics demand has placed pressure on the supply chain beyond what it can absorb internally. - Samsung Electro-Mechanics said that although it has continued to improve manufacturing efficiency and expand capacity, price adjustments for certain products have become unavoidable to maintain product quality and a stable supply. - The price increase will apply to all MLCC products under Samsung Electro-Mechanics’ Markup Business Code, with the new prices taking effect for shipments made on or after August 1, 2026. - The company emphasized that the price adjustment is unavoidable and said it will continue to supply high-quality products and support its customers’ businesses. - Taiyo Yuden is also moving to raise MLCC prices from September 1, while noting that it may be unable to guarantee customers’ requested delivery schedules. - Taiyo Yuden explained that it decided to adjust prices because ongoing increases in raw-material costs could no longer be absorbed through cost-cutting efforts alone. - Even after the price adjustment, supply is expected to remain insufficient to fully meet demand, and the company may still be unable to guarantee on-time delivery.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 22:57:58 UTC · 原创Net Profit 93.92T!($64.59B)↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 22:50:51 UTC · 原创SK hynix: Long-term contracts with approximately 10 major customers↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-28 22:50:26 UTC · 引用Also a dated point from the authors' of the Economist piece. Settlement is the best indicator of true intervention, and China has been a net buyer of fx over the last year↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 22:49:37 UTC · 原创SK hynix (Market Cap: KRW 1,132.2632 trillion) 2Q26 Earnings Results ▶️ Results: Revenue: KRW 79 trillion (+257% YoY) (Consensus: KRW 84 trillion; 5.5% below consensus) Operating profit: KRW 61 trillion (+557% YoY) (Consensus OP: KRW 64 trillion; 5.4% below consensus) Operating margin: 76.3% (Consensus OPM: 76.2%; 0.1%p above consensus)↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-28 22:47:02 UTC · 引用But a stronger yuan also won't add much to China's deflationary pressures, and it would address the rest of the world's concern that China is getting too much growth from net exports ...↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-28 22:30:06 UTC · 引用One additional observation -- with a current account surplus reported to be about 3 pp higher than the IMF's norm (4 pp v 1 pp, give or take), the IMF's models imply the RMB is undervalued by 20%. That is a big gap to be resolved by swings in inflation differentials!↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-28 22:25:27 UTC · 引用I personally see it as extreme short term deleveraging that overshot many individual names. - $BE to $TER reported blowout results, with Bloom reporting 166% Y/Y rev growth + expanding margins + raised 2026 guidance... Teradyne reported 104% Y/Y revenue growth with 300%+ EPS growth. Your leading optical players over in China reported amazing preliminary earnings + extremely strong read through for the Western photonics sector. Think earnings season tend to remind markets about continued AI acceleration. I'm expecting $LITE, $SNDK / SK Hynix, and sector leader ERs to continue that trend. But it does feel like markets are rewarding existing acceleration more rather than future growth a year out over near-term revisions as seen with $AMKR. But as a TLDR, looks very positive for the themes I'm tracking. - Then you have $GOOGL capex raised to $195-$205B, which is typically the biggest AI demand signal from hyperscalers. - 77% 0bps change est. next FOMC decision from prediction markets. Trump administration on Monday called for the Fed to lower interest rates. So your 3x rate hikes fears this year seem kinda overblown. - Chinese fears from DUV to $CXMT overflooding are extremely overblown as well. We see this with every sector from time to time until people remember that lasers or HBM bottlenecks for a reason. As for personal thoughts, it was unhealthy seeing everything rise up or down together, especially with these themes. But I'm personally not worried since I have full conviction in my names, especially optical interconnects like $SIVE or $AAOI. $META compute and $GOOGL GCP margins/growth shows increasing demand for $NBIS and similar neocloud business models. $META + other hyperscaler notes around LTAs with $SNDK and Samsung/SK Hynix, $MU point to structural memory demand. Can go on and on... But when you have Jim Cramer telling every margined DC trader to "sell everything" at market open and news saying there's an AI bubble. I do hate it when others say "it's dropping so sell". My belief is that the important thing to look is accelerating revenue/EPS growth. And if that thesis stays in-tact with revenue acceleration in line with hyperscaler capex. TLDR: I see signs of AI demand / revenue acceleration, hyperscaler capex is one of the most important things to track. And my guess is that we'll see the theme recover broadly. But I can't predict what hour, day, or week that happens.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-28 21:31:56 UTC · 原创It isn't every day that the last two chief economists of the IMF join forces to push back against one of your arguments! The debate over the role of China's currency in the increase in China's surplus is an important one, and I am happy to have this debate in the open 1/↗ @AndrewYNg · 请求账号 @AndrewYNg · 2026-07-28 20:19:41 UTC · 原创Fifteen years ago, @Coursera and online courses changed education. It worked better than almost anyone expected, expanding access by opening up where you can learn. But how you learn remains largely the same as it has for centuries: it is still one-size-fits-all, taught the same way to each person who shows up. We now have an opportunity to change how learning happens. With advances in AI, we can now build a custom learning guide for each person. We will turn learning from one‑to‑many to one‑to‑one. I'm starting LearnVector to invent this next generation of learning. We are starting with a $100M investment from Coursera, and plan to collaborate closely with Coursera and Udemy. Good learning needs much more than just a chatbot. Research shows that chatbots without guardrails harm learning. They help complete tasks and enable students to do better on homework. But cognitive offloading to a chatbot results in them being less skilled. And, you cannot always trust what a chatbot tells you. In contrast, LearnVector will plan a path with you, adapt to how you learn, and patiently stay with you until you’ve mastered new skills. One thing has not changed in all this time. People want learning they can trust: material that is accurate, relevant, and worth the effort you put into it. Anything less wastes the most valuable thing a learner has: time. Coursera has a trusted library of materials from authoritative sources. LearnVector plans to work with Coursera to bring this trustworthy learning to everyone. I'm grateful to Greg Hart and the entire Coursera team for supporting LearnVector. I look forward to working with our talented team to change how we learn, and accelerate human development.↗ @Claudia_Sahm · 请求账号 @Claudia_Sahm · 2026-07-28 19:45:48 UTC · 引用Re-upping my question for the Fed presser tomorrow: "2% of What?" I would like to know whether Warsh agrees that PCE inflation is the Fed's target.↗ @aleabitoreddit · 请求账号 @aleabitoreddit · 2026-07-28 18:45:19 UTC · 原创FCC set to ban Chinese imports of humanoid and quadruped robots per Reuters. To protect American Humanoid/Robotics programs. Agility Robotics ( $CCXI ) and other US humanoid players should be happy to hear this.↗ @realDonaldTrump · 请求账号 @realDonaldTrump · 2026-07-28 17:42:14 UTC · 引用We’re going to bring this heroic young man, and his family, into the White House with, perhaps, the boy he saved, to give him a High Civilian Honor. Very brave, he deserves it! President DONALD J. TRUMP↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-28 15:31:35 UTC · 引用Good The EU (i.e. Commission) should be doing this as well. Now that China has weaponized rare earth/ magnet supplies, the ability to counter retaliate is critical for any real form of autonomy↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 12:21:42 UTC · 原创If their share prices get crushed even further from here, will the hyperscalers cut capex? We have bond vigilantes, but are there equity vigilantes? Personally, I find it hard to imagine them cutting capex.↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-28 11:02:39 UTC · 原创This is an important point. When the economic value of a banking system's liabilities exceeds the economic value of its assets, the system can postpone adjustment only as long as it can capitalize and roll over the difference, in the form of accelerating debt growth. Once it can no longer do so, however, the economy will be forced to recognize the underlying loss and – and this is politically the hard part – allocate it between households, different business sectors, and different government entities. This could happen in the form of a financial crisis, and many times it does, but only in financial systems in which a severe mismatch between the structure and timing of assets and liabilities is faced with a sharp contraction in liabilities. But in financial systems in which regulators are able to control net outflows from the financial system and to restructure liabilities at will – and China's is obviously such a financial system, just as Japan's was in the 1980s – the adjustment is unlikely to come in the form of a financial crisis or collapse. It is much more likely to be a long, drawn-out affair, much like what happened in Japan during the 2-3 decades after 1991. Is this a good thing or a bad thing? It depends. Financial crises can be much more disruptive socially and politically in the short term. But because a slower, less violent adjustment slows down the capital reallocation process and generates deep and long-term financial distress costs (along with the associated hysteresis), it can be much more costly economically over the medium and long term. So I would argue that the answer to the question of which is a better form of adjustment depends on whether the more important objective is short term political and social stability, or medium- and long-term economic growth.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 10:53:35 UTC · 原创“Although we do not have the detailed allocation yet, we think it is difficult to reach Google’s target with TSMC alone, and Intel’s supply is a must by 2028.” $INTC↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 10:51:55 UTC · 引用NVIDIA’s moat.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 07:39:34 UTC · 原创>> TrendForce: Monthly Shipments by Korean and Japanese MLCC Makers Hit Record High on AI Demand; Spillover in Consumer Orders Becomes Visible • According to TrendForce’s latest MLCC industry survey, strong AI demand drove Murata, Samsung Electro-Mechanics, and Taiyo Yuden—the three leading Korean and Japanese suppliers—to record their highest monthly shipment volumes in five years in June 2026. Meanwhile, continued spillover in consumer orders pushed up prices at other manufacturers and across distribution channels. • As Korean and Japanese suppliers rapidly shift consumer-grade X5R capacity toward higher-spec X6S/X7R products for AI applications, capacity for mid- to high-capacitance consumer MLCCs (1µF–22µF) is coming under pressure. Inventory levels for key products have generally fallen to 30 days or less. • Supply constraints have triggered a surge in urgent orders from distributors, agents, and Tier 2 and Tier 3 customers, further amplifying the order-spillover effect. As a result, order visibility at other manufacturers has improved markedly, accompanied by higher product prices. Distributors that had built inventories in advance have raised prices by an average of 20–25%, while spot-market prices have climbed to two to three times their previous levels. The consumer MLCC market is therefore exhibiting a structural divergence in which end demand remains weak while distribution-channel prices stay firm.↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 07:36:42 UTC · 原创A magnitude 7.0 earthquake has struck the area where TSMC’s fab in Japan is located.↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-28 06:58:13 UTC · 原创SCMP: "Nearly 70 per cent of the surveyed entrepreneurs said traditional manufacturing and service-sector firms were caught in a dilemma: companies had to engage in price wars to maintain revenue growth, but had to scale back operations if they wanted to preserve profits." This, much more than ideology, is what has been driving the retreat of the private sector and the expansion of the public sector over the past two decades in China's economy. In an environment of massive (and growing) excess capacity and pricing involution, businesses that operate under hard-budget constraints, and so need to preserve profits, must scale back operations. This requires that much of the politically-determined expansion in total capacity must be driven by businesses that don't operate under hard-budget constraints. As a result, manufacturing capacity shifts from the former to the latter.↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-28 06:42:02 UTC · 原创Caixin: "China’s top leadership is expected to hold off on major new stimulus at a key Politburo meeting in late July, focusing instead on executing existing fiscal policies to stabilize a slowing economy. The meeting may emphasize accelerating the construction of new infrastructure networks, while policymakers are expected to shift consumption support from subsidizing physical goods toward service sectors."↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-28 06:39:20 UTC · 原创Caixin: "In the high-stakes game of global energy, China has been playing the crucial role of shock absorber. Market observers expect Chinese demand to remain subdued until crude prices stabilize further, choosing instead to opportunistically replenish reserves during price drops."↗ @michaelxpettis · 请求账号 @michaelxpettis · 2026-07-28 06:33:12 UTC · 原创Yicai: "Among the top 10 provinces by GDP, Jiangsu, Shandong, Zhejiang, Sichuan, Henan, Hubei, Shanghai, and Anhui had growth faster than the national average."↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 05:44:14 UTC · 原创TFHK Commentary: How Should We Understand the Current Correction in AI Hardware? The market is always right. Changes in stock prices inevitably reflect the new variables the market is currently pricing in. Even as long-term bulls on the AI industry, we need to understand the core concerns driving this correction in AI semiconductor stocks. The current market bears a striking resemblance to last autumn and winter. Following OpenAI’s large fundraising round, industry conditions were very strong, yet stocks continued to trade sideways. Market participants spent every day debating CapEx, ROI, valuations, and financing—much like they are doing now. The conclusions from this quarter’s earnings reports remain overwhelmingly positive. GCP grew by 80%, the ROI of cloud investment was validated, Intel delivered a significant beat, and ASML, TSMC, and Intel raised their order or CapEx outlooks. Presumably, these companies also saw extremely strong downstream forecasts, giving even the most conservative players in the supply chain the confidence to make aggressive bets. Had this information emerged in May or June, semiconductor stocks would almost certainly have surged. Now, however, every earnings release has instead become an opportunity for bears to reassess valuations and the long-term investment thesis. What we may be seeing is that the AI market is no longer in the “AI Summer” of May and June. The same positive developments now provide less support to share prices. Take GCP’s 80% growth as an example. Previously, the market’s first reaction would have been: “AI demand has exceeded expectations—the catalyst has arrived.” Now, the first response is: “So what? What about 2028? Can OpenAI become profitable? For how many more years can GPU prices keep rising? Margins are rising again, financing costs are increasing, and the entire CapEx thesis needs to be repriced.” In essence, the market has shifted from trading the growth of AI CapEx to trading its sustainability and ROI. Market sentiment, as we perceive it, has already become extremely bearish. Even long-term bulls are beginning to question whether AI semiconductor stocks can continue to rise, and we are hearing almost no calls for new highs. When the market shifts from looking for further upside to searching for additional downside risks, it usually means that pessimistic expectations have already been largely priced in. Nevertheless, we have no doubts about the fundamentals. We also believe that, ultimately, facts determine stock prices. So what would send these stocks higher again? Under the framework outlined above, additional capital-spending plans alone will no longer be enough to convince the bears. What is needed is validation of a new demand curve. The most powerful and direct catalyst would be the emergence of a blockbuster product. If “Coding 1.0” proved that AI can improve developer productivity, then “Coding 2.0” must demonstrate that AI agents can genuinely replace part of the software development process. Once new productivity use cases are validated, the market’s concerns about AI ROI may be redefined, and AI infrastructure spending will once again be viewed as “productivity investment” rather than a “cost.”↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 04:26:20 UTC · 原创Seeing even CXMT and Chinese lithography-equipment plays fall in the Chinese market really makes it feel like we’ve entered a bear market.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-28 02:09:14 UTC · 引用"the real problem lies in its external imbalances" so true why does the IMF have difficulty speaking with such clarity? its mandate is external not fiscal ...↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-28 01:24:39 UTC · 引用China's 2026 trade surplus looks very different with and without gold imports (which reached ~ 1.5 pp of GDP in q2) 1/2↗ @jukan05 · 请求账号 @jukan05 · 2026-07-28 01:08:16 UTC · 原创[Exclusive] SK hynix to Begin Mass Production of LPDDR6 in Second Half; Xiaomi to Be First Customer SK hynix plans to begin mass-producing and shipping LPDDR6, its next-generation low-power mobile DRAM, in the second half of this year. Having completed development of the product in March, the company intends to make its first shipment to China’s Xiaomi for use in mobile devices. The launch is expected to further intensify competition among the three major memory manufacturers—Samsung Electronics, SK hynix, and Micron—in SOCAMM, a next-generation, LPDDR-based low-power memory module for AI servers. According to industry sources on July 28, SK hynix will enter full-scale mass production of LPDDR6 in the second half of this year. Xiaomi is expected to be its first customer, with the memory slated for use in the company’s next-generation flagship smartphones. SK hynix’s LPDDR6 uses its sixth-generation 10nm-class (1c) process, offering improved data-transfer speeds and power efficiency over the previous generation. In March, SK hynix announced that it had completed certification of what it described as the world’s first LPDDR6 product, saying, “We plan to complete preparations for mass production within the first half and begin supplying the product in the second half.” Regarding the reported Xiaomi deal, however, the company said it could not confirm information related to customers. SK hynix began doing business with Xiaomi in 2013, supplying LPDDR3 for certain models, and is understood to have continued supplying mobile DRAM through the latest LPDDR5X generation. Samsung Electronics and SK hynix currently lead the mobile DRAM market. Meanwhile, CXMT, China’s largest memory manufacturer, has been rapidly gaining ground following its stock-market debut, bringing its market share close to that of third-ranked U.S. chipmaker Micron. Samsung supplies LPDDR to Android smartphone manufacturers worldwide, while SK hynix has a strong position among Chinese smartphone brands. The Xiaomi deal is expected to pave the way for SK hynix to expand LPDDR6 shipments to other Chinese smartphone manufacturers, including Vivo and Oppo. Chinese smartphone makers are increasingly adopting domestically produced memory from suppliers such as CXMT in their entry-level lineups, but their flagship models—which require higher performance—remain dependent on LPDDR from the three major global memory manufacturers. LPDDR is primarily used in mobile devices such as smartphones and tablet PCs. By reducing power consumption while maintaining high performance, it has recently experienced explosive demand from AI servers. Some industry observers expect it to become a major growth driver for the memory market after HBM, or high-bandwidth memory. LPDDR supply prices have recently risen to levels comparable with those of DRAM, whose prices have been soaring. The increase stems from supply bottlenecks caused by memory manufacturers concentrating their production capacity on commodity DRAM and HBM. As LPDDR applications diversify and contract prices continue to rise, memory manufacturers posted record-high revenue in the first quarter, supported by strong LPDDR sales. Market-research firm TrendForce said, “Mobile DRAM revenue increased by 64.5% in the first quarter of this year,” adding that “continued increases in contract prices will drive further growth in mobile DRAM revenue in the second quarter.” Samsung Electronics and SK hynix are also seen as having seized the initiative by beginning preparations for mass production of the latest LPDDR6 products ahead of Micron. In January, Samsung unveiled what it described as the industry’s first LPDDR6 product built using fifth-generation 10nm-class (1b) DRAM technology. The product won a CES Innovation Award. SK hynix’s mass production of LPDDR6 is also expected to signal the start of a full-fledged race to secure an early lead in the SOCAMM market. SOCAMM packages LPDDR into a module designed for AI servers. As the AI market shifts from training toward inference, it is rapidly emerging as a next-generation memory solution for AI servers. In June, Nvidia CEO Jensen Huang visited SK hynix’s booth at Computex 2026, Asia’s largest information-technology trade show, and left the message “LOVE SOCAMM” on a 192GB SOCAMM module. Nvidia also recently unveiled Vera, its first internally designed CPU for AI servers. If HBM is the GPU’s natural partner, LPDDR is the CPU’s. As the AI market becomes increasingly advanced and demand for CPUs rises sharply, demand for SOCAMM is also expected to strengthen.↗ @sama · 请求账号 @sama · 2026-07-27 20:02:25 UTC · 引用wrong↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-27 19:31:11 UTC · 原创The case for a "gold-adjusted" Chinese trade and current account surplus. A new blog 1/↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-27 19:19:40 UTC · 引用I applaud Mr. Balding's return (?) to blogging -- A few reactions ... 1/↗ @AndrewYNg · 请求账号 @AndrewYNg · 2026-07-27 17:01:38 UTC · 引用Good move by @JensenHuang. The Nvidia letter is well written and worth reading. As we saw with the OpenAI-Hugging Face hack, we need open models and harnesses for defense. Lets stop believing the PR that closed models are safer. - that's just regulatory capture.↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-27 17:01:26 UTC · 引用good points↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-27 16:00:25 UTC · 引用Think the position of Merz and Macron now matters more than the head in the sand view of the US Treasury ... China's currency is an issue that won't go away just b/c the Treasury didn't express concern about China's rising backdoor intervention in 2025↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-27 15:49:11 UTC · 原创If the recent surge in gold imports (really a financial transaction) are netted out, China's underlying surplus in goods trade has continued to shoot up -- and is now approaching 1.5 pp of the GDP of all its trade partners (a record obviously for a single country) Tis crazy↗ @Brad_Setser · 请求账号 @Brad_Setser · 2026-07-27 15:19:00 UTC · 引用There is a certain kind of investment bank analyst, full of confidence but perhaps not yet seasoned with much experience, who does a chart deck based on a data snap shoot from 5 years ago, "sells" their point of view, & then never bothers to update their priors 1/↗ @nft_hu · 请求账号 @nft_hu · 2026-07-27 15:12:10 UTC · 原创半导体鬼故事也太多了 一会内存太贵，一会kimi ，一会战争，一会国产替代DUV。 说真的还是没有大资金买入了，风声鹤唳的。好消息不涨但是坏消息一定跌。亚盘美盘轮着跌。 资本流动才是鬼故事背后的真相。↗ @jukan05 · 请求账号 @jukan05 · 2026-07-27 14:40:24 UTC · 原创A thought that just occurred to me: the factors currently weighing on the market are not new negatives. The semiconductor cycle peaking, deteriorating profitability due to excessive CAPEX, China’s push for semiconductor self-sufficiency... People simply want to take profits. As time goes on, the market keeps resurrecting whatever bearish narratives it needs to justify selling, creating a self-reinforcing vicious cycle.↗

YouTube 更新

今日最值得看

Greg Isenberg：Jack Dorsey's Buzz: The New Hermes Agent?

Buzz 将智能体放进团队聊天的第一等位置，并让 Claude Code、Codex、Goose 等底层执行器可替换。演示覆盖代码分支、CRM 原型和自动回传数据；价值在于上下文和工作流不被单一模型或平台锁死。

AI Engineer：How Forward Deployed Engineering is done at Factory — Eno Reyes

Factory 把前置部署工程放进“软件工厂”循环：客户信号转为计划，经验证后交付。视频强调客户自有、模型无关的执行框架，以及通过代码库可验证性衡量智能体自治度。

Latent Space：OpenAI’s Plan to Make ChatGPT the Everything App — Akshay Nathan, OpenAI

OpenAI 产品负责人讨论 ChatGPT Work、Codex 与面向知识工作的智能体体验如何共享基础能力但保持不同界面。节目涉及持久电脑、产物、站点、插件、记忆和子智能体，也把“品味与问题定义”视为更稀缺的瓶颈。

其他频道更新

AI Engineer · 2026-07-28AI tools for Forward Deployed Engineering — Vasuman Moza, Varick AgentsVarick Agents 从企业既有流程而非迁移大系统开始，尝试让智能体覆盖端到端的业务衔接。节目将难点落在跨系统上下文、实体识别和可追溯规格，而不是单纯调用模型。↗ AI Engineer · 2026-07-28How Forward Deployed Engineering is done at Cognition — Jia WuCognition 讨论如何用前后对比的业务结果衡量部署价值，而非只看 token 或演示效果。其核心判断是：随着编码成本下降，定义问题、选择杠杆点和对客户结果负责更重要。↗ AI Engineer · 2026-07-28How Forward Deployed Engineering is done at Ramp — Leo MehrRamp 的案例把“先界定范围”放在前置部署工程的首位，避免把每个销售请求都变成定制开发。视频还展示用智能体把需求入口转成规格，并以评测和人工反馈约束输出。↗ AI Engineer · 2026-07-28The Dirty Secret of Forward Deployed Engineering — Natalie Meurer, SierraSierra 的分享回看 FDE 从驻场运维到数据集成、解决方案架构的演变。观点是编码更便宜后，持续理解客户数据与对交付结果负责仍是不可替代的工作。↗ AI Engineer · 2026-07-28How Forward Deployed Engineering is done at Decagon — Sunny RekhiDecagon 将企业客服智能体的部署拆为配置行为规则和产品化集成两部分。视频强调把定制能力回流进平台，使下一位客户可以自助复用，而非无限堆积项目制工作。↗ Computerphile · 2026-07-28PiDP6 & Lisp - ComputerphileComputerphile 用 PiDP6 重访 1960 年代 PDP-6 与 Lisp 的历史和实践。它适合作为理解交互式计算、语言文化与早期机器设计的一段轻量补充。↗