독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-21154 게시물

새로고침

Super Micro Raises Revenue and Margin Guidance, Lifting Dell and HPE Premarket Shares

주제 · 超微电脑业绩중대
2026-10-21 18:26 GMT+8

US server manufacturer Super Micro Computer issued a preliminary financial outlook, raising its revenue forecast for the quarter ended September 30 to $18.5 billion–$19.5 billion from the previous guidance of $14.5 billion–$15.5 billion. It also increased its gross margin projection to 12%–12.5% from 10.4%–10.8%. This robust update is viewed as a positive signal for artificial intelligence infrastructure demand.

In response, shares of major AI hardware suppliers Dell Technologies and Hewlett Packard Enterprise rose nearly 5% and 4% respectively in Wednesday premarket trading, while Super Micro's own stock surged more than 12%. Evercore ISI analyst Amit Daryanani noted that Super Micro's margin improvement was "an impressive feat" and suggested it bodes well for Dell and HPE, citing the notable combination of revenue upside and better-than-expected margins.

Despite the boosted sentiment, investors remain skeptical about the sustainability of enterprise AI spending. Barclays downgraded ratings on Dell, HPE, and NetApp last week, citing no "meaningful AI inflection" in enterprise spending and attributing recent revenue strength to higher memory costs rather than accelerated AI adoption. Super Micro is scheduled to report official earnings on November 3, which will provide more details on demand sources.

출처:invezz.com

리서치

Former NCC Chairman Chan Ting-yi Analyzes Anthropic Copyright Case: Court Splits AI Training into Two Independent Acts, Ending the 'Train First, Settle Later' Strategy

주제 · Anthropic版权案중대
2026-10-21 18:35 GMT+8

On June 23, 2025, the United States District Court for the Northern District of California issued a ruling in the case of Bartz v. Anthropic PBC, which concluded with a $1.5 billion copyright damages settlement agreement. Chan Ting-yi, former Chairman of the National Communications Commission (NCC) of Taiwan and Chairperson of the Artificial Intelligence Technology Foundation, believes that this massive settlement is not the most core legal legacy of the case—the true change in the AI industry landscape lies in the 'Behavioral Decomposition Analytical Framework' pioneered by presiding Judge William Alsup.

Before the Bartz case, most technology companies and copyright law scholars viewed AI training as a single, continuous, and indivisible legal act. Chan Ting-yi argues that this holistic characterization was the jurisprudential basis for the 'train first, settle later, always claim fair use' strategy: as long as the final training constituted transformative use, the entire process should be protected under fair use. Judge Alsup rejected this premise, deconstructing the AI model training process into two independent 'manifestations of conduct,' assigning key legal characterizations to each.

The first stage is 'backend model training.' The court determined that when large language models read book texts, their purpose is not to reproduce or leak the content of the works to the public, but to analyze statistical probability patterns, semantic associations, and stylistic features among words to build a language generation model with new expressive capabilities. The judgment characterized this behavior, used purely for machine learning, as 'the process of extracting statistical patterns and semantic relationships for machine learning purposes,' believing it aligns with the core spirit of fair use in U.S. copyright law—highly transformative use.

The second stage is 'frontend data acquisition.' The red line drawn by the court is that no matter how high the transformative nature of the backend is, it cannot whitewash the illegal original sin of the frontend data source. Evidence presented by the plaintiffs during the trial proved that Anthropic had obtained over 7 million unauthorized e-books from pirate shadow libraries such as Library Genesis (LibGen) and Books3. The court held that if AI companies download and use this content while knowing or should have known its source was pirated, the illegality of the data acquisition act itself directly prevents them from claiming fair use.

Chan Ting-yi points out that the establishment of this 'dual-track framework' ends the arrogance of technological determinism, clearly declaring that technological innovation cannot exempt the illegality of acquisition channels. However, the court did not negate all data acquisition behaviors indiscriminately. Anthropic had invested millions of dollars to purchase large quantities of legitimate physical books in major book markets worldwide, handing them over to industrial-grade scanning factories for destructive digitization—using mechanical hydraulic guillotines to remove book spines, separating pages, scanning them into PDFs, and subsequently destroying the original paper copies. The court ruled that this practice of 'format conversion through authorized physical acquisition, combined with the destruction of the original tangible medium to maintain a one-to-one ratio' constitutes lawful transformative use, because it does not increase the total number of circulating copies of works on the market.

출처:TechOrange 科技報橘

리서치

Startup Inception Releases Mercury 3 Diffusion Model, Claims 8x Speed Over GPT-6 Luna

중대
2026-10-21 06:45 GMT+8

US-based AI startup Inception released its flagship model, Mercury 3, on October 19, 2026. The company states that in its internal benchmarks, this diffusion-based language model outperformed OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 in 8 out of 12 tests, achieving generation speeds 4 to 8 times faster than these competitors.

Unlike traditional word-by-word generation, Mercury 3 drafts an entire response in parallel and then refines each position through multiple passes. Inception reported that on Nvidia GPUs, Mercury 3 achieved 1,068 tokens per second, compared to approximately 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna. The model features a context window expanded to 1 million tokens and is described as more than 40 times larger than its predecessor, though specific parameter counts were not disclosed.

No independent third-party organizations have yet verified Mercury 3's performance claims. All data provided is self-reported by the vendor, leaving real-world stability, cost-efficiency, and long-context handling capabilities still to be observed.

출처:SiliconANGLE

리서치

Anthropic CEO Dario Amodei Responds to Safety Criticism in Long Post: Argues for 'Race to the Top' Strategy with Nine Commercial Costs

중대
2026-10-21 10:26 GMT+8

Dario Amodei, CEO of Anthropic, published a long post of approximately 2,000 words on X (formerly Twitter) directly responding to criticism from Nathan Calvin, an AI safety watchdog, regarding the company's safety record. Calvin had previously highlighted several discrepancies between Anthropic's safety commitments and its actual actions. Rather than avoiding these allegations, Amodei put forward a more fundamental argument: Anthropic's safety decisions over the past year—which he listed as nine—were extremely costly commercially, but precisely these costs demonstrate that its 'race to the top' strategy is working.

Amodei's core logic is that only companies at the frontier can set industry standards by making expensive safety trade-offs; such trade-offs by frontier companies have a far greater impact on the industry than those by non-frontier companies. Examples he cited include: the Mythos-class model was ready in February 2026, but Anthropic delayed its public release until June to deploy stronger cybersecurity protections first, which he admitted likely sacrificed months of model leadership advantage; imposing strict classifiers on biological capabilities resulted in false positive rates so high that 'biologists joke about it,' while competitors used this to market their more permissive products; supporting US state-level AI safety legislation like California's SB53, and arguing for stronger provisions than were ultimately passed; restricting the sale of AI services to Chinese-controlled enterprises, including through third-country channels; and announcing this week that the company would dedicate 5% of its equity (and potentially more in the future) to a public benefit fund.

Amodei also made concessions. He admitted that items seven, eight, and nine listed by Calvin constituted 'imprudent wording or policy communication,' citing information lag common in large organizations (he noted Anthropic has over 5,000 employees), and stated, 'As CEO, I take responsibility.' Regarding item five—the revision of the Responsible Scaling Policy (RSP)—he argued that the original September 2023 announcement described the RSP as an 'early version requiring rapid iteration,' so the modification did not constitute a broken promise. However, he acknowledged that item six (the relationship between setting up an FCF outside the RSP) was a complex issue and promised further review.

The most powerful rebuttals came from Amodei's own comment section. Sigal Samuel asked directly: 'What evidence is there that Anthropic’s fundamental strategy—competing at the frontier to drive a race to the top—is actually successful? From the outside, it looks like Anthropic has done more on net to drive a race to the bottom.' Zack Korman pointed out a specific contradiction: Amodei claims the government should have the power to decide who gets access to certain capabilities, which is why he advocates for regulation, 'but we’ve seen the bill Anthropic is pushing, and it doesn’t give the government that power.' Tim Hua believed Anthropic severely overestimated misuse risks. These questions point to the same boundary: all nine costs listed by Amodei are self-reported by the company, and without independent financial data to verify their commercial costs, it remains undeterminable from the outside whether the 'race to the top' truly raised industry standards or merely provided a safety narrative for competition itself. At the end of his post, Amodei acknowledged that when AI-driven acceleration becomes too fast for safety to keep up, the 'pacing' problem must be addressed, stating that Anthropic is thinking in this direction.

출처:X (formerly Twitter)

리서치

Brian MacMurray, President of General Motors Korea R&D Center: The Core Issue with AI Is Not Intelligence but Control

중대
2026-10-21 15:45 GMT+8

If a stranger told you, 'Follow me, I'll take you where you want to go,' most people would refuse. But when getting into a car and opening KakaoMap or Naver Map, people follow the navigation almost without thinking. Brian MacMurray, President of General Motors Korea R&D Center (GMTCK), used this analogy at the FIX2026 forum held in Daegu, South Korea, on October 21, 2026, to illustrate: 'Unknowingly, we have handed control over to a strange entity.'

His core argument is that the real issue in the AI era is not how smart machines are, but whether humans still retain decision-making power. This judgment points to a specific risk mechanism—the 'auto-approve' function of AI agents. When AI executes tasks without asking users item by item, users are ceding approval authority. MacMurray warned: 'Pressing the auto-approve button is handing over your approval authority. If you're not careful, you might even hand over your job.'

MacMurray is not opposed to AI itself. He acknowledges that AI can improve efficiency and help make better decisions, but insists that who determines the direction and whether to accept the results must be judged by humans. 'I have already accepted the fact that AI will make me better, but I will not hand over the authority to decide which direction to go or what results from the machine to accept.'

Regarding the issue of AI hallucinations, he cited this as a reason why humans must remain in the review loop: even if AI provides results faster than humans, the role of finding errors and reconfirming cannot be abandoned. His advice is based on professional knowledge and experience: stop and check when you feel 'something is wrong'—'that is how humans take responsibility.'

Finally, he concluded with an automotive-related image: customers get in the car believing the driver will do everything right and not put them in danger. 'I want to sit in the driver's seat. It doesn't mean I want to drive personally, but I want to retain the authority to decide what to do.'

It should be noted that MacMurray's argument assumes human review is always feasible and necessary. However, in practice, the cognitive load of item-by-item review of AI outputs may exceed what humans can sustainably provide; the convenience of 'auto-approve' exists precisely because manual review is too slow. Which decisions deserve human attention and which can be delegated is more complex than his proposed framework suggests.

출처:zdnet.co.kr

리서치

NVIDIA Releases RFD4-Proteina Open Model for Protein Design, Claims Higher Success Rates Than Baselines

주제 · 英伟达生物计算拓展중대
2026-10-21 18:13 GMT+8

NVIDIA announced at GTC Berlin on October 21, 2026, that it has debuted RFD4-Proteina, a new open-source model developed in collaboration with the University of Washington Medicine Institute for Protein Design. The model combines programmable protein design with search capabilities to help scientists navigate vast biological design spaces more efficiently.

According to research drafts cited in NVIDIA's press release, RFD4-Proteina achieved higher computational design success rates and generation throughput than evaluated baselines across four benchmark suites: protein binding, small-molecule binding, enzyme scaffolding, and DNA binding. These performance metrics are currently self-reported by the vendor and have not been independently reproduced. The model is available through partners including Amazon Bio Discovery, Anthropic’s Claude Science, Hugging Face, and others.

Beyond drug discovery, NVIDIA also announced a collaboration with German medical technology company Siemens Healthineers to develop next-generation adaptive radiotherapy systems. By leveraging Nemotron foundation models and CUDA-accelerated computing, the project aims to use agentic co-planning and real-time simulation to improve treatment precision and address specialist capacity constraints in cancer care.

For AI industry investors, this move signals NVIDIA's strategy to extend its hardware dominance into high-barrier vertical markets like life sciences. By providing specialized open models and platform tools, NVIDIA seeks to lock in long-term compute demand from research institutions and pharmaceutical companies, although the near-term financial contribution of these applications remains uncertain.

출처:NVIDIA Newsroom

리서치

NVIDIA and Cisco Unveil USFP Optics Standard, Claiming Fourfold Bandwidth Density Increase

주제 · 英伟达思科USFP标准중대
2026-10-21 14:36 GMT+8

U.S. chip giant NVIDIA and networking equipment provider Cisco announced on October 21, 2026, that they are collaborating to develop a new optical transceiver form factor called USFP (Ultra-Dense Small Form Factor Pluggable). Designed to support 400 Gb/s per lane connectivity, the standard aims to address the network demands of next-generation "AI factories." According to the companies, USFP offers four times the bandwidth density of existing 200 Gb/s per lane pluggable optics, along with approximately 1.3 times better power efficiency, with capacities reaching up to 12.8 Tb/s per module.

The core innovation lies in addressing physical bottlenecks in scaling AI clusters. As GPU counts surge, traditional network architectures require additional switching layers, increasing latency, cost, and power consumption. The USFP design integrates 16 full-duplex lanes into an OSFP-sized footprint, allowing more computing systems to be connected without adding network complexity. Additionally, its novel two-dimensional array connector features a two-step insertion mechanism where high-speed electrical contacts engage only after the module is fully inserted, reducing wear and improving signal integrity. The modules can dissipate up to 100W without requiring dedicated embedded cooling circuits, simplifying switch designs.

It is important to note that these performance metrics are official claims by NVIDIA and Cisco, not yet verified by independent third-party testing or large-scale deployment. The companies are currently working with connector manufacturers like TE Connectivity and Amphenol to develop components and plan to establish an open industry specification through a community-led multi-source project. For data center operators globally, if this standard succeeds, it could offer a flexible path to higher-bandwidth Ethernet networks, but its actual commercial timeline and ecosystem compatibility remain to be seen.

출처:eeNews Europe

리서치

Reuters: Chinese tech giants order over 6 million downgraded Nvidia AI GPUs, far exceeding next year's total capacity

주제 · 英伟达对华出口管制중대
2026-10-21 04:46 GMT+8

According to Reuters, major Chinese technology companies including Alibaba, ByteDance, and Tencent have placed orders for more than 6 million RTX PRO 5500 graphics cards from Nvidia. At the suggested retail price, the total value of these orders is approximately $35 billion. These GPUs are intended for artificial intelligence inference tasks rather than gaming.

However, Nvidia faces a severe production bottleneck. Citing sources, Reuters reported that Nvidia's maximum output for 2027 will be only 3 million units, covering global demand across all customers. This means that even if all orders are valid, the actual delivery volume to Chinese firms could represent just 5% to 10% of their requested quantities.

This massive demand for downgraded GPUs stems from US export control policies. US President Donald Trump has banned the export of Nvidia's most powerful GB102 and GB112 chips to China. Previously, Nvidia introduced the restricted RTX PRO 6000D specifically for the Chinese market, but production was halted after the Chinese government refused to import "second-tier" chips that had passed through US security checks. Currently, the RTX PRO 5500 remains one of the few models exempt from the ban that can partially meet AI inference needs.

For AI industry investors, the stark gap between the sheer scale of these orders and Nvidia's hard production ceiling highlights how geopolitical friction distorts the global AI infrastructure supply chain, significantly driving up the time cost and uncertainty of acquiring compute power.

출처:heise online

리서치

NVIDIA Launches DSX POD Modular AI Factory Design, Cutting Deployment Time to As Little as 14 Days

주제 · 英伟达DSX POD중대
2026-10-21 17:34 GMT+8

NVIDIA introduced DSX POD at GTC Berlin on October 21, 2026, a reference design for building complete Vera Rubin AI factories. The solution involves partners integrating computing, networking, cabling, and liquid cooling systems within modular containers before delivery, allowing units to connect directly to prepared sites' power and cooling infrastructure.

This model aims to address key bottlenecks in traditional data center construction: high onsite complexity, specialized labor shortages, and component delay risks. NVIDIA states that DSX POD can reduce deployment time at a prepared site from the traditional six months to as little as 14 days. This "ready out of the box" design enables operators to begin generating production tokens and earning revenue sooner, improving the capital efficiency of AI infrastructure investments.

NVIDIA has established a global ecosystem with partners including Anduril, Armada, Bull, Crusoe, DayOne, Dell Technologies, Delta, Eaton, Foxconn, GIGABYTE, HPE, Lenovo, Pegatron, Quanta Cloud Technology, Schneider Electric, Vertiv, and Wistron, planning to deliver these modular AI factories worldwide in early 2027.

출처:NVIDIA Newsroom

리서치

The Chinese Philosophical Tradition Does Not View Intelligence as the Essence of Humanity, Redefining the Threat of AI

중대
2026-10-20 16:36 GMT+8

A premise underlying Western anxiety about AI is rarely scrutinized: that intelligence is the essence of what makes us human. We have even named our species after this concept—Homo sapiens, Latin for 'wise man.' If intelligence defines humanity, then when machines catch up to or surpass us in intelligence, the threat becomes existential: we would be competing against an 'artificial version of ourselves.' In his article, author Arnaud Bertrand notes that this premise explains much of Western AI anxiety, but it is not the only possible starting point.

China possesses an equally ancient and sophisticated intellectual tradition that has never placed intelligence at the core of 'what makes us human.' In this tradition, what defines humanity is 'ren' (仁)—a concept encompassing both 'benevolence' and 'humanity,' rooted in the 'xin' (心), a term referring simultaneously to emotion and reason. The heart/mind is not the brain; it is where we feel, think, discern right from wrong, and care for others.

Philosopher Zhao Tingyang of the Chinese Academy of Social Sciences therefore offers a judgment almost impossible to encounter in Western AI discourse: attempting to align AI with human values 'is likely to be a suicidal mistake.' Liu Yongmou, a philosopher of technology at Renmin University of China, flips the problem around: 'Humanity’s biggest problem is the machinization of humans, not the humanization of machines.' Zhu Songchun, director of the Beijing Institute for General Artificial Intelligence, aims not to make machines more like humans, but to 'establish a heart for machines.'

This divergence traces back over two thousand years. In Politics, Aristotle imagined a world where shuttles wove automatically and lyres played themselves, concerned with how social order would change—specifically, whether slaves would still be necessary. In Zhuangzi, there is a nearly identical scenario: an old gardener refuses to use labor-saving irrigation machinery, arguing that 'those who use machines must possess a machine-heart' (机心)—using machines makes people calculating and shrewd, stripping away simplicity. He says, 'I am not unaware [of the machine], but I am ashamed to use it.' Two labor-saving devices, two entirely different questions: Aristotle asked what automation does to social order, while Zhuangzi asked what it does to the human interior.

Bertrand also cites another warning from Zhao Tingyang: the bad consequences of AI can at least be fought, but 'humanity may die of all the good things AI creates.' There is also a discussion regarding 'uselessness': if humans become 'useless' due to AI, this fear presupposes that human value equals utility—a premise demolished long ago by the tree in Zhuangzi that survived for 2,300 years precisely because it was useless.

It should be noted that Bertrand is an interpreter within the Western context; his selection of material and framework itself carries the self-awareness of being a 'frog in a well'—he acknowledges that Western AI discourse is far from the only perspective. The original text is behind a paywall; the free portion presents core arguments and key quotations, but the full argumentation behind Zhao Tingyang’s 'suicidal mistake' judgment and the specific connotations of Liu Yongmou’s 'machinization of humans' may require reading the full article to fully understand. These thinkers’ viewpoints are attributed arguments, not verified facts; they change the way questions are asked, rather than providing definitive answers.

출처:arnaudbertrand.substack.com

리서치

Japan's METI Plans Additional ¥400 Billion Investment in Wafer Foundry Rapidus

주제 · Rapidus融资与量产중대
2026-10-20 10:35 GMT+8

According to Kyodo News, Japan's Ministry of Economy, Trade and Industry (METI) stated on October 19 that it plans to seek an additional ¥400 billion investment in public-private partnership wafer foundry Rapidus through the "industrial investment" scheme during fiscal years 2027-2029. If approved, the Japanese government's cumulative investment in Rapidus would reach ¥650 billion.

Previously, METI had invested a total of ¥250 billion in Rapidus twice through the Information-technology Promotion Agency (IPA). Rapidus aims to begin mass production of 2-nanometer chips in the second half of fiscal year 2027. In addition to direct equity investment, the Japanese government also plans to provide subsidies under the name of commissioned R&D, amounting to approximately ¥630 billion in FY2026 and ¥300 billion in FY2027.

출처:TechNews 科技新報

리서치

Europe's tech funding hits €27.8B in Q3 2026, with AI and cloud taking nearly half

중대
2026-10-21 17:50 GMT+8

According to Tech.eu's Q3 2026 European tech funding report, European tech companies raised €27.8 billion across 732 deals this quarter. This represents a 32.4% increase from €21 billion in Q3 2025 and is more than double the €11 billion recorded in the same period in 2024, highlighting a strong rise in capital deployment over the past two years.

Funding activity was highly concentrated, particularly with a sharp surge in September. Data shows five deals exceeded €1 billion, involving four companies: Mistral AI (€3 billion), Nscale (two rounds worth approximately €2.95 billion and €2.59 billion), Helsing (€1.58 billion), and Quantum Systems (€1.05 billion). Notably, Nscale raised about €6.3 billion during Q3 alone through three transactions.

Geographically, the UK maintained its lead, raising almost €12 billion across 188 deals, accounting for around 43% of total European investment. Germany and France followed with approximately €4.9 billion and €4.0 billion respectively. Together, these three countries attracted nearly three-quarters of all capital raised in the quarter.

By sector, artificial intelligence and cloud computing were the primary recipients of capital. AI companies raised €6.7 billion, while cloud companies secured €6.4 billion. Combined, these two sectors attracted over €13 billion, representing about 47% of all capital raised during the quarter. This indicates that European venture capital is accelerating towards AI infrastructure and applications, though it also reflects a trend of concentration among leading large-model and cloud platform vendors.

출처:Tech.eu

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

However, the court's characterization of Anthropic's internal 'central database' was completely different. Many AI vendors permanently retain large volumes of downloaded or scanned works in servers to facilitate engineers adjusting parameters, performing data cleaning, and reproducing training results. Anthropic lawyers argued that these databases were merely 'incidental temporary storage acts' for model training. The judge dismissed this defense with the phrase 'It kept them anyway,' ruling that the long-term storage of unauthorized works after training was completed was no longer pure machine learning, but effectively established an 'unauthorized digital library,' allowing internal personnel to retrieve and access original works at any time. This replaced the commercial licensing demand and circulation of originals in the normal market, constituting a serious risk of market substitution. Chan Ting-yi believes this determination directly cuts off AI companies' attempts to use 'transient copying' as a legal safe harbor.

In Chan Ting-yi's view, the practical impact of the behavioral decomposition framework on the industry is reflected at the negotiating table. The Bartz case ruling shows that the court's scrutiny of 'improper sources' is extremely strict. To eliminate the devastating risk of models being forcibly taken offline or companies facing compensation, major players like OpenAI completely abandoned their early hardline stance of defending fair use in court against media outlets, shifting entirely to defensive upfront licensing. Rights holders began demanding clauses requiring the 'deletion of original cached data immediately after training' in negotiations, or charging high premiums for rights regarding long-term retention, retrieval, and fine-tuning. Third-party audits also became necessary contract clauses—rights holders have the right to dispatch independent institutions to audit AI companies' databases.

At the end of her article, Chan Ting-yi proposes three trend predictions: First, collective management organizations modeled after the music industry will become mainstream within two years, with the publishing sector establishing standardized, tiered pricing licensing pools; Second, legitimate physical book whitewashing pipelines and synthetic data will evolve along dual tracks, but synthetic data faces technical bottlenecks regarding model collapse; Third, copyright holders will adopt proactive rights defense technologies such as Nightshade and Glaze, embedding pixel or semantic perturbations into the underlying code of digital publications, causing training data forcibly scraped to directly disrupt model parameter associations. These are predictions deduced by the author based on the legal framework and have not yet been verified by the market or technology.