독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-28154 게시물

새로고침

Samsung Announces Shift to 'Co-design' Model for AI Memory, Details zHBM Roadmap

주제 · 三星zHBM共同设计战略중대
2026-10-28 13:10 GMT+8

At the Tech Summit 2026 held in Seoul on October 28, 2026, Kim Kyung-yoon, Executive Vice President of the DRAM Development Group at Samsung Electronics' Device Solutions (DS) Division, announced a paradigm shift in the company's artificial intelligence (AI) memory business model. The strategy moves away from traditional commodity mass production or simple customer customization toward a 'Co-design' approach, where memory solutions are jointly developed with semiconductor clients from the planning stage.

Kim explained that while current High Bandwidth Memory (HBM) products are customized for specific clients, the heterogeneity of future GPU architectures requires deeper integration. "Customers' GPUs vary significantly in structure," Kim stated. "It is impossible to combine a uniform HBM product with various customers' AI semiconductors." This necessitates a new operational model capable of supporting 'high-mix, high-volume' production through collaborative engineering.

Central to this strategy is zHBM, a next-generation memory architecture featuring vertical stacking of HBM directly on top of the GPU, contrasting with the current 2.5D side-by-side placement. Samsung projects that zHBM will deliver up to 8 times the performance and 3 times the power efficiency of HBM5, which is scheduled for release in 2028. Commercialization of zHBM is expected after 2029.

This announcement highlights the evolving dynamics of the AI hardware supply chain. While recent reports indicated NVIDIA pushing for standardized base dies via TSMC to reduce validation burdens, Samsung's emphasis on 'Co-design' suggests a counter-strategy: leveraging its integrated capabilities in memory, foundry, and logic design to lock in closer relationships with key AI chipmakers. Kim noted that success in the zHBM era depends on comprehensive capabilities across these domains, an area where Samsung holds a competitive advantage over pure-play memory vendors.

출처:미래를 보는 창 - 전자신문

리서치

UK Government Proposes 'AI Covenant' Requiring Data Centers to Invest in Communities for Grid Priority

주제 · 英国AI数据中心社区契约중대
2026-10-28 19:05 GMT+8

According to TechRadar, the UK government is considering establishing a new "AI covenant" that would require data center developers to meet specific criteria before construction, including investing in local communities. This move aims to alleviate public dissatisfaction with AI infrastructure expansion and support Britain's ambition to become a global leader in AI technology.

Under the proposed framework, developers who meet the "covenant" standards could receive priority access to power grids and fast-tracked planning approvals. Potential forms of investment include direct payments to residents or funding for local infrastructure. UK AI Minister Kanishka Narayan stated that the government would not allow the AI buildout to purely serve "foreign firms and foreign billionaires," seeking instead a "distinctly Labour approach to British AI" that better serves the public.

This policy comes amid significant local opposition to AI data centers in the UK. For instance, a planned 600MW data center in Fife, Scotland, has attracted over 1,600 objections, while a 450MW project in North Wales faces a petition with more than 26,000 signatures. Additionally, Microsoft recently withdrew from residency plans at the UK's largest data center due to concerns about grid connection delays until the early 2030s.

The "covenant" is currently under consideration, with no concrete plans for implementation details, legal binding force, or timelines announced. For investors, this suggests that compliance costs for UK AI infrastructure projects may rise, but they might also accelerate deployment by securing critical resources like electricity.

출처:TechRadar

리서치

M&A Advisor Itay Sagie: Revenue Growth May Mask Declining Enterprise Value; Exit Decisions Should Track Valuation Multiples

중대
2026-10-28 19:00 GMT+8

Strategic advisor Itay Sagie had just finished a call with a potential client—a company performing excellently but not even considering a sale, as it hopes to achieve greater growth over the next two years before exiting. The conversation prompted him to consider an issue many entrepreneurs overlook: revenue can rise while enterprise value falls.

He illustrates this point with a hypothetical SaaS scenario: a company with $15 million in annual recurring revenue (ARR) valued at 6x ARR has an implied enterprise value of $90 million. Two years later, ARR grows to $20 million, but the valuation multiple compresses to 3.5x revenue, causing enterprise value to drop to $70 million. Recurring revenue increased by 33%, yet value declined by 22%. Sagie explicitly notes this is merely an illustrative assumption intended to demonstrate how multiple compression can mask operational growth.

Why do multiples change? Sagie categorizes factors influencing revenue multiples into three groups. Qualitative factors include management quality, founder dependence, competitive intensity, product differentiation, and strategic relevance to buyers. Absolute financial metrics include revenue, EBITDA, and cash flow. Relative performance metrics include growth rate, margins, net revenue retention, and gross revenue retention. The key dynamic is that as companies mature, absolute performance can continue to improve, but growth rates tend to weaken—this is a common source of multiple compression.

The resulting practical advice is: M&A preparation should begin while the company is winning. A sale requires preparation, buyer engagement, negotiation, and due diligence, during which buyers continuously reassess performance and forecasts. Waiting until a slowdown becomes visible weakens negotiating leverage and invites scrutiny of management’s projections. Furthermore, strategic buyers’ needs follow their own timetable; high valuation placed today on a specific technology, customer base, or geographic footprint may quickly dissipate due to internal product launches, competing acquisitions, or budget changes.

Sagie acknowledges that no one can reliably identify the valuation peak in advance, and waiting can still create value—when additional scale attracts more buyers, profitability improves, or customer concentration decreases. However, he believes boards should review estimated enterprise value alongside the operating plan, modeling scenarios where revenue grows while the multiple holds steady, expands, or contracts. They should also consider how debt, cash, dilution, liquidation preferences, and transaction terms affect actual shareholder proceeds. Every decision to delay a sale embeds a judgment about future valuation, and that judgment deserves its own graph.

출처:Crunchbase News

리서치

Investment Commentator Klement: Research Shows Gold Reserves Do Not Significantly Protect Exchange Rates Against Unexpected Federal Reserve Rate Hikes

주제 · 黄金储备汇率保护效果중대
2026-10-28 15:00 GMT+8

Gold bulls sometimes argue that central banks should increase their gold reserves, with one reason being that when the US Federal Reserve makes unexpected monetary policy adjustments, gold can simultaneously protect against both domestic currency depreciation and US dollar depreciation. Joachim Klement cites research by Aizenman, Saadaoui, Uddin, and Yago to quantitatively test this argument, concluding that the results do not support the superior protective role of gold.

The research team examined the exchange rate performance of emerging market central banks when the US Federal Reserve unexpectedly raised interest rates by 10 basis points. The core mechanism is intuitive: if a central bank holds more foreign exchange reserves, it can sell foreign currency and buy back its own currency when facing depreciation pressure on its domestic currency, thereby stabilizing the exchange rate. Data shows that for every one standard deviation increase in foreign exchange reserves (roughly corresponding to 20 percentage points), the depreciation of the exchange rate decreases by 0.04 percentage points.

However, the key lies in the currency composition of the reserves. For every one standard deviation increase in US dollar reserves, the depreciation of the exchange rate decreases by 0.1 percentage points—an effect more than double that of general foreign exchange reserves. The impact of non-dollar foreign exchange reserves is minimal and statistically insignificant. Klement points out that this aligns with intuition: since central banks are countering US dollar appreciation pressure, directly holding US dollar liquidity naturally provides the strongest intervention firepower.

So, what about gold? Gold is traded in US dollars, so for central banks, it is essentially an indirect US dollar exposure. The study analyzed foreign exchange reserves and gold separately and found that while the exchange rate buffering effect of gold reserves was numerically similar to that of general foreign exchange reserves, its statistical uncertainty was significantly higher, making the effect statistically insignificant. In other words, gold did not provide additional protection beyond directly holding US dollars.

Klement draws rather restrained conclusions from this analysis: First, foreign exchange markets seem to view gold as an indirect US dollar asset that can be sold to generate US dollar liquidity when needed; second, gold reserves do not truly enhance exchange rate stability compared to directly holding US dollars. He also emphasized that this analysis yields no hard-and-fast conclusions.

For investors, if the core rationale for increasing gold holdings is to hedge against the impact of unexpected Federal Reserve tightening on domestic currencies, empirical evidence does not support gold's protective function over US dollar reserves. Regarding exchange rate volatility caused by unexpected Federal Reserve policy, gold has not demonstrated statistically significant independent buffering capacity.

출처:klementoninvesting.substack.com

리서치

John Garrett, Author of Investment Masters Class: Listening Is a Compound-Interest Source for Management That Is Severely Undervalued

중대
2026-10-28 13:11 GMT+8

Warren Buffett was once asked about the management fads that seem to sweep through business schools and boardrooms every few years. His answer was typically understated: “There are these fads in management… obviously, listening to your customer and things like that… nothing makes more sense. But it’s hard to write a 300-page book that just says, ‘Listen to your customer.’” John Garrett, author of Investment Masters Class, believes this line captures the dilemma of many truly important ideas in business: they are too simple to sell. Nobody builds a consulting practice around listening. Yet when he reviews his notes on hundreds of businesses studied and the coaches and captains who built great teams, a recurring finding emerges: when the builders explain what made the difference, surprisingly many do not talk about vision or persuasion, but about listening—not as a soft skill or a pleasant personal quality, but as a discipline practiced daily and embedded in how the company operates.

Garrett distills this argument into a mechanism of information flow. Ken Iverson, who built Nucor into the most efficient steelmaker in America, pointed out that a reputation for good listening encourages others to proactively bring information. Garrett derives from this a distinction with direct significance for investors: a leader known for listening is brought information, while a leader known for not listening is brought reassurance. Over ten to twenty years, the difference between these two information flows compounds like everything else. One company discovers its problems while they are still small; the other finds out from its customers, competitors, or share price.

This mechanism can be observed at three levels. First, at the customer level: Dallas car dealer Carl Sewell refused to assume he knew what his customers wanted, instead actively seeking out customers who had left to ask why they bought a Mercedes or Lexus instead of a Cadillac—most companies study existing customers, rarely those who have churned, and the truth usually lies with the latter. Second, at the employee level: When Tom Hennesy took over Tractor Supply, he asked store managers only one question, “What’s wrong?” After listening for a month or two, he said, “We have two problems: we’re confused, and we don’t have any inventory to sell.” The solution followed, and the company subsequently became one of America’s finest compounding machines. Third, in the behavioral patterns of new executives: David Marquet, upon taking command of the nuclear submarine Santa Fe, skipped the standard two-week file review procedure and instead spent time talking to people on the ship; within two years, the submarine went from the worst-performing in the fleet to the best.

Garrett ultimately translates these observations into executable due diligence questions for investors: How did the CEO spend their first three months in the role? When was the last time an executive or board member heard directly from a dissatisfied customer? What happened the last time someone told the CEO they were wrong? Where do improvement ideas come from? Can management point to a significant decision changed because of feedback from employees or customers? These answers rarely appear in annual reports; they emerge in conversations with customers, employees, and management, often requiring years of accumulation. The author points out that the real challenge may not lie in knowing what questions to ask, but in being willing to hear the answers.

출처:Investment Masters Class

리서치

Gartner: Global AI Protection Market to Grow 2.7x to $7.7 Billion by 2028

중대
2026-10-28 16:00 GMT+8

According to a report by South Korean tech outlet ETNews, market research firm Gartner predicts that the global artificial intelligence (AI) protection market will grow approximately 2.7 times over the next two years. The market is valued at $2.835 billion this year and is expected to increase by 68.7% to $4.783 billion next year, reaching about $7.7 billion by 2028.

As enterprises widely adopt generative AI, demand for security solutions addressing AI-specific vulnerabilities and cyber threats has surged. Among key segments, 'AI Usage Control'—which manages which AI services employees use and what data they input—is projected to have the highest growth rate, increasing 73% from $433 million to $749 million next year. 'AI Gateways' and 'AI Governance Platforms' are forecast to grow by 70.9% and 68%, respectively.

'AI Application Security' remains the largest segment, expected to reach $851 million next year with a 67.5% increase. Gartner cited supply chain attacks on external software and open-source components as key drivers, recommending that companies integrate dedicated AI tools into existing security frameworks. The firm also anticipates increased acquisitions of startups by major security vendors to broaden product portfolios.

출처:미래를 보는 창 - 전자신문

리서치

South Korea Announces 1,500 Trillion Won AI Investment Plan Focusing on Semiconductors, Data Centers, and Physical AI

주제 · 韩国AI投资计划중대
2026-10-28 16:00 GMT+8

Lee Jin-soo, Director-General for AI Policy Planning at the Ministry of Science and ICT (MSIT), outlined South Korea's strategy to become a global "Top 3" AI power during a speech at the "K-Open Source AX" event in Seoul on October 28. Facing overwhelming computational advantages held by the United States and China, the government identified three key areas for narrowing the gap: expanding computing infrastructure, strengthening competitiveness of proprietary AI models, and resolving disparities in AI adoption.

To achieve these goals, the government plans to drive a total of 1,500 trillion won in public-private investments across three "Mega Projects": semiconductors, AI Data Centers (AIDC), and Physical AI. According to details from a July announcement, over 800 trillion won will be allocated to semiconductors, approximately 550 trillion won to domestic and international AIDC construction, and 20 trillion won specifically to R&D and industrialization of Physical AI. Lee described the synergy using an analogy: semiconductors are the brain, data centers are the heart, and Physical AI is the body, together forming future growth engines.

Beyond hardware, South Korea emphasizes building "full-stack" capabilities. Citing NVIDIA CEO Jensen Huang's "Five-Layer Cake" theory (energy, semiconductors, infrastructure, models, applications), Lee noted that South Korea is one of the few nations possessing indigenous technology and companies at every layer. However, he acknowledged that the top-layer AI application market remains in its early stages. Consequently, the government aims to promote personal AI agents through the "AI for All" project and deepen vertical-specific model applications in sectors like law, chemistry, and manufacturing to reduce reliance on foreign services.

On the international front, South Korea is actively integrating into global AI governance. During the San Francisco AI Summit in July, President Lee Jae-myung signed six AI infrastructure cooperation agreements totaling $950 billion (approx. 1,275 trillion won). Additionally, the country plans to establish a "Global AI Hub" involving nine international organizations and five multilateral development banks to enhance its voice in setting international standards.

출처:미래를 보는 창 - 전자신문

리서치

Security Firm Cyfirma Reports Suspected North Korean Hackers Scanning Global PHP Hosts for PHP2Shell Exploitation, Prioritizing Taiwan

주제 · PHP2Shell漏洞중대
2026-10-28 17:29 GMT+8

Security firm Cyfirma warned that immediately after the PHP development team announced and patched the major PHP2Shell vulnerability on October 21, suspected North Korean hacker groups linked to Lazarus began reconnaissance and scanning of global web hosts vulnerable to PHP2Shell, filtering for targets that had not yet completed patching.

Cyfirma discovered two Korean-speaking hackers coordinating attack activities and sharing infrastructure. They built a list of 216 hosts suspected of being unpatched or only partially mitigated, executing low-frequency, batched vulnerability scans through 21 compromised nodes to avoid detection by minimizing dense connections. By October 24, hackers had scanned approximately one-third of the targets, attempting to deploy PHP-based Webshells and contaminate data.

The targeted countries and regions include Taiwan, the United States, Japan, Thailand, Indonesia, South Korea, Australia, India, Singapore, Saudi Arabia, and the European Union. The targeted industries span government agencies, banks and financial institutions, manufacturing, logistics, shipping, airports and aviation, and online trading platforms.

출처:iThome

리서치

ByteDance's Doubao Work Upgrades: Adds Desktop Operation Recording and Background Task Execution on Locked Macs

주제 · 豆包工作Mac锁屏功能간단한 요점
2026-10-28 15:00 GMT+8

According to Leifeng Wang, ByteDance's AI assistant "Doubao Work" has recently upgraded its computer operation capabilities. The update introduces two main features: allowing users to record desktop operations as reusable personal skills, and supporting the continued execution of computer-operating tasks even when the Mac screen is locked.

For skill recording, after selecting "Record Operations to Generate Skill" in the chat interface and completing authorization, users can operate their computer as usual while adding voice instructions for specific requirements (such as file naming rules). The system records mouse clicks, text input, and window content, organizing the workflow into a skill. Next time a similar task is needed, users can simply invoke the skill and specify the new target, eliminating the need to re-explain steps.

Regarding task execution, enabling the "Allow Doubao to Work When Screen is Locked" setting permits Doubao Work to continue running while the Mac is locked. Users can issue commands to the locked Mac via their phone and view results remotely. This requires the Mac to remain powered on, connected to the internet, and with the client running. This feature aims to help users delegate repetitive operations to AI, freeing up time for judgment and creative work.

출처:leiphone.com

리서치

Standard Chartered Q3 Net Profit Jumps 46%, Raises 2026 RoTE Target to 14%

주제 · 渣打银行Q3业绩중대
2026-10-28 12:28 GMT+8

Standard Chartered reported a 46% year-on-year increase in third-quarter net profit to US$1.5 billion, beating analyst forecasts of US$1.2 billion. The bank's shares closed 7.6% higher at HK$219.80 on Wednesday following the release.

The performance was primarily driven by robust growth in the wealth management unit and a significant reduction in credit impairment charges. Fee and other non-interest income rose 19% to US$2.82 billion, while Hong Kong, the bank's largest single market, saw underlying pre-tax profit surge 91% to US$944 million. Bad debt charges fell 52% year-on-year to US$90 million.

CFO Manus Costello noted the addition of 76,000 new-to-bank affluent clients during the quarter. CEO Bill Winters announced an upgrade to the 2026 outlook, stating that operating income growth is now expected to exceed the previous 5-7% range. Consequently, the bank has raised its Return on Tangible Equity (RoTE) target from 12% to approximately 14% for 2026, reflecting confidence in capturing structural shifts in global trade and wealth creation.

출처:South China Morning Post

리서치

Perplexity Researcher Adel Karimi Reports First Successful Exploitation of PHP2Shell Vulnerabilities in Honeypots

주제 · PHP2Shell漏洞중대
2026-10-28 17:10 GMT+8

Last week, the PHP development team announced and patched major security vulnerabilities collectively known as PHP2Shell (CVE-2026-109618, CVE-2026-109619). Security firms Wiz and Calif, who discovered these flaws, had previously noted that without AI assistance, researchers would have struggled to find vulnerabilities with actual destructive power, and most reported active exploitation attempts were limited to reconnaissance. However, recent observations in self-hosted honeypot traps indicate that attackers have successfully exploited the vulnerability.

Perplexity researcher Adel Karimi stated that between 19:26 and 19:31 UTC on October 24, he detected successful exploitation of PHP2Shell through six sensors in honeypot traps. Attackers executed 35 shell processes in total, launching /bin/sh via specific Apache processes in a honeypot running PHP 8.4.26 with Apache mod_php. This marks the first time anyone has successfully exploited PHP2Shell to achieve arbitrary code execution.

Regarding attack sources, the activity originated from two IP addresses: 15.204.59[.]70 and 162.35.99[.]94. One IP address triggered only three shell executions, while the other exhibited more diverse activities, including the use of tools such as PHP CLI, curl, and wget, followed by dynamic identity checks. Although both IP addresses shared templates and body hash characteristics, they did not share a control mechanism. Adel indicated that it remains unclear whether these actions were carried out by the same group or initiated by different attackers.

출처:iThome

리서치

Hong Kong Exports Surge 64.9% in September, Highest Growth in Over 40 Years Driven by AI Products and Tariff Front-Loading

중대
2026-10-28 16:53 GMT+8

Preliminary data released by the Hong Kong Census and Statistics Department on Wednesday showed that total export values surged 64.9% year-on-year in September, reaching HK$762.6 billion (US$97.8 billion). This represents the highest monthly growth rate in more than four decades, surpassing the previous peak recorded in January 1984. The figure follows a 53% increase in August, marking the 20th consecutive month of double-digit growth.

Artificial intelligence-related products were a primary driver of this momentum. Exports of electrical machinery and parts soared 66.2% to HK$390.8 billion, while office machines and automatic data processing equipment expanded by 183.7% to HK$135.4 billion. A government spokesman noted that while trade performance continues to benefit from the wave of AI investment, risks such as Middle East conflicts, trade protectionism, and the rapid expansion of global AI investment warrant close monitoring.

Beyond AI demand, uncertainty surrounding US tariff policies has led to significant front-loading of orders. Lee Shu-kam, director of the Business, Economic and Public Policy Research Centre at Hong Kong Shue Yan University, attributed the surge to companies rushing to buy AI products before policy clarity emerges. He estimated that this strength would likely persist until tariff policies become clear later next year, though subsequent months may not match September's record pace.

Exports to Asia rose 67.5% year-on-year, with particularly sharp increases to Taiwan (159.3%), Malaysia (131.5%), and Thailand (118%). Total imports grew 57.8% to HK$808.9 billion, resulting in a trade deficit of HK$46.3 billion. For the first nine months of 2026, overall exports and imports increased by 45.2% and 45% respectively, accumulating a trade deficit of HK$417.7 billion.

출처:South China Morning Post

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

It should be noted that the strength of this argument lies in the richness of its attributions and the clarity of its mechanism, rather than quantitative validation. Garrett provides no control groups, regression analysis, or measurable proxy indicators to isolate the independent contribution of listening relative to other management advantages. A company that listens well may simultaneously possess better hiring, stronger culture, and superior products; listening is not necessarily the sole causal variable. However, as a due diligence framework—that is, observing whether information reaches decision-makers before distortion—it provides a set of specific test points more concrete than reading strategic statements. Its value lies not in proving causality, but in narrowing the scope of judgment.