East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- Sony’s capture of 50% of the global CMOS image sensor market positions the company to dominate the next hardware bottleneck as autonomous vehicles and edge-AI devices require increasingly sophisticated vision chips.
- Micron’s plan to double its high-bandwidth memory production capacity by 2026 threatens the historical duopoly of South Korea’s SK Hynix and Samsung in the Nvidia-dominated AI hardware supply chain.
- Google’s partnership with Marvell Technology to develop custom silicon beyond its TPU accelerators establishes a new playbook for hyperscalers looking to bypass standard chip merchant silicon and reduce long-term capital expenditure.
This Issue’s Analysis
The Signal
Sony’s 50% Image Sensor Dominance: The New Chokepoint in Edge AI and Automotive Supply Chains
Sony’s CMOS image sensor (CIS) market share has reached approximately 50% globally, according to a recent Yole Group survey, matching Sony’s internal projections. The company’s sem
Semiconductors & Hardware
Micron’s 100K Wafer HBM Target: Doubling Capacity by 2026 to Challenge Samsung
Micron plans to double its High Bandwidth Memory (HBM) production capacity by the end of 2026, targeting a monthly output of approximately 100,000 wafers. This aggressive expansion
Semiconductors & Hardware
Taiwan’s Supply Chain Is Bracing for a Record Q4 Export Surge Driven by Next-Gen AI Platforms
UBI Funds predicts strong performance for US, Taiwan, and Korean stock markets in Q4 2026, driven by NVIDIA’s GB200 and next-generation chip platforms entering mass production and
Semiconductors & Hardware
Kintron Technology Expands Glass Substrates to Solve the AI Packaging and CPO Bottleneck
Taiwanese firm Kintron Technology is expanding its Through Glass Via (TGV) substrate and glass micro-processing technologies for AI advanced packaging and co-packaged optics (CPO).
🧩 Pattern This Issue
- Korea/Taiwan: Micron doubling HBM capacity challenges Korean memory dominance
- Taiwan: Kintron glass substrates advance AI packaging and co-packaged optics
- Korea/Taiwan: NVIDIA GB200 shipments fuel Taiwan and South Korea supply chains
The escalation of hardware bottlenecks is shifting AI value creation from model software to physical manufacturing in Taiwan and South Korea, leaving Western firms increasingly dependent on East Asian advanced packaging and memory capacity.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 Semiconductors & Hardware
Google and Marvell Partnership Signals Shift in AI Architecture to Custom Silicon Beyond Accelerators
Google has partnered with Marvell Technology for custom silicon, a deal extending beyond Google’s Tensor Processing Units (TPUs) to include AI inference accelerators, storage controllers, network interface controllers, memory interface controllers, and near-memory compute. This collaboration involves Google potentially acquiring up to 59 million Marvell shares at $206.58 per share, contingent on Marvell achieving up to $120 billion in cumulative sales by fiscal year 2033 from custom products supplied to Google.
Why it matters: Google’s move with Marvell suggests a broader industrial shift where hyperscalers are not just building custom AI accelerators but custom silicon across the entire data path within their data centers. This vertical integration indicates that traditional hardware vendors might see their market share eroded as customers like Google design more of their own components for greater efficiency and control.
For Western readers: Western hardware suppliers should anticipate increased competition from customer-designed silicon, particularly in high-volume components for AI data centers, and plan for a future where major cloud providers are their own largest component customers.
🗾 AI & Machine Learning
Google Releases Gemini 3.8 Flash and Cyber Defense-Specialized Gemini 3.8 Flash Cyber, Prices Unchanged
Google has launched Gemini 3.8 Flash, three weeks after 3.7 Flash, touting significant improvements in reasoning and coding. Concurrently, it introduced Gemini 3.8 Flash Cyber, a version specialized for cyber defense, available exclusively to vetted ‘trusted defenders’ via its new Fairwind Program, with pricing for 3.8 Flash remaining at 3.7’s introductory rates through 2026.
Why it matters: Google is pursuing an interesting strategy by releasing a rapid succession of ‘Flash’ models for general use while simultaneously restricting access to a specialized ‘Cyber’ version. This indicates they are trying to capture both speed/cost-sensitive developers and high-stakes enterprise/government users, and they understand that the latter group will need a very different level of control and security around the models.
For Western readers: Western businesses, particularly those in critical infrastructure or cybersecurity, should investigate the Fairwind Program as a potential source for advanced AI defense tools, rather than assuming all powerful AI models will be generally available. This restricted access model signals a tiered approach to AI deployment that will become more common.
🗾 Policy & Regulation
Trump Administration Supports OpenAI in Copyright Lawsuits, Argues AI Training is Fair Use
The US Department of Justice, under the Trump administration, filed an amicus brief on September 1st with the Southern District of New York federal court, supporting OpenAI in multiple copyright infringement lawsuits, including those brought by The New York Times. The brief argues that the reproduction of copyrighted works for LLM training constitutes ‘transformative’ fair use, as the models learn statistical patterns rather than reproducing expressions for their original purpose.
Why it matters: The Trump administration’s legal stance is not merely an opinion; it’s a clear signal from the US executive branch favoring accelerated AI development over content producers’ direct licensing demands. This position implies a strategic prioritization of national AI competitiveness, specifically against rivals unconstrained by similar licensing requirements, over the immediate commercial interests of legacy media companies. It also suggests that the current US administration views AI training as a distinct, transformative use, rather than a direct substitute for original content, which could set a precedent for future legal interpretations globally.
For Western readers: Western businesses involved in AI development should consider this a strong indication that the US government intends to provide a favorable regulatory environment for AI training, reducing the immediate risk of widespread, mandatory licensing fees for data ingestion. For content providers, this means the path to compensation for AI use of their data will likely be through negotiation for output-based value rather than upfront licensing for training data.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 AI & Machine Learning
CrowdWorks to Verify AI Agent Safety with Data Building Expertise
CrowdWorks, a South Korean AI data specialist, announced its entry into the AI agent safety verification market. Leveraging its extensive experience in constructing diverse datasets for AI models, the company aims to establish robust methodologies and infrastructure to assess the reliability and ethical compliance of AI agents.
Why it matters: CrowdWorks is taking a logical step, extending its data expertise into the nascent field of AI agent validation. This isn’t just a new service; it’s an effort to define best practices for a foundational AI component that the industry has yet to standardize. Establishing a rigorous verification process for agents is crucial for their broader adoption, especially in regulated or sensitive applications, and the company that sets the standard here could win significant market share.
For Western readers: Western AI developers and integrators should track the methodologies and standards emerging from companies like CrowdWorks, as these will likely influence global best practices for AI agent safety and compliance, potentially impacting their own development and deployment costs.
🇰🇷 AI & Machine Learning
Global VC Founder Praises South Korean Government AI Investment, Citing Samsung and SK Hynix for Leap Forward
📊 Featured Chart
Source: Artificial Analysis AI Index
Nathan Benaich, founder of AI-focused VC AirStreet Capital, lauded South Korea’s 10 trillion won AI investment and its goal to become a top-three AI powerhouse. He highlighted Samsung Electronics and SK hynix’s semiconductor capabilities, skilled developers, and a rapid adopter user base as key strengths for the nation’s AI ambitions. Benaich also noted the strong performance of Motif Technology’s ‘Motif3’ foundation model, which surpassed models from major Korean conglomerates in recent benchmarks.
Why it matters: Benaich’s comments provide a significant third-party validation for South Korea’s strategy of combining government investment with its existing semiconductor manufacturing prowess. This isn’t just a government announcement; it’s a prominent VC founder specifically calling out the strength of the Korean semiconductor giants as a foundation for broader AI growth, which suggests serious traction for the national AI agenda. The performance of Motif Technology’s model is also a concrete data point showing that Korean AI startups can achieve competitive results on a relatively lean budget.
For Western readers: Western businesses considering AI partnerships or investments in East Asia should specifically evaluate South Korea’s emerging AI ecosystem, especially those seeking to integrate AI with hardware or advanced manufacturing. Do not dismiss smaller Korean AI firms; the performance of Motif Technology shows they can produce competitive models, potentially offering more cost-effective or specialized solutions than larger, more visible players. If you’re looking for AI talent or models that are highly attuned to industrial applications, South Korea should be on your radar.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 Policy & Regulation
US Government Supports OpenAI in Copyright Dispute with The New York Times, Citing Fair Use
The US Department of Justice (DOJ) filed a ‘Statement of Interest’ in a consolidated copyright infringement lawsuit against OpenAI, explicitly backing OpenAI’s use of copyrighted articles for AI model training under the principle of fair use. This marks the first time the US federal government’s stance on this issue has been publicly stated. The lawsuit, initiated by The New York Times in December 2023, accuses OpenAI and Microsoft of unauthorized use of its content to train AI products like ChatGPT, with other media outlets and authors joining the consolidated case.
Why it matters: The DOJ’s intervention on the side of OpenAI clearly indicates the US government views a robust domestic AI industry as a national interest, framing a broad interpretation of fair use as essential for innovation and competition. This position directly challenges the media industry’s claim that AI training constitutes market dilution and will likely embolden AI developers while putting pressure on content owners to negotiate new licensing frameworks.
For Western readers: Western businesses in the AI sector should proceed with greater confidence in their data acquisition strategies for training, assuming government backing for a ‘transformative use’ argument, but content creators should prepare for a challenging legal environment where their claims of market harm from AI training are likely to be disputed by federal agencies.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 Robotics & Automation
One Model Dominates Across Scenarios! ItSihang AWE3.7 Shows Powerful Generalization Capabilities
Chinese robotics company ItSihang (它石智航) has showcased its AWE3.7 (AI World Engine) general embodied AI model, demonstrating its ability to perform complex tasks across diverse environments. Over ten days, the company released a series of videos showing the model controlling robots in industrial production lines, logistics, and household service scenarios, using a single ’embodied brain’ system for all tasks. This aims to prove the model’s ‘universal generalization‘ capability, moving beyond single-task competency to enable skill transfer and reuse across different environments and robot bodies.
Why it matters: The company’s framing of a single foundational model handling complex, diverse tasks from industrial assembly to home care directly challenges the traditional robotics paradigm of specialized models. If ItSihang can translate these demonstrations into robust, scalable deployments, it indicates a stronger domestic player emerging in the embodied AI sector, shifting competition from specialized hardware to generalized intelligence platforms.
For Western readers: Western robotics and AI companies should recognize that Chinese firms are rapidly advancing their general-purpose embodied AI models, aiming to bypass the need for extensive task-specific training. This could accelerate the deployment of autonomous systems in diverse sectors, and potentially erode the competitive advantage of Western firms still focused on highly specialized robotic solutions.
🇨🇳 Robotics & Automation
Mysterious Embodied AI Team Releases New Demo Videos, Revealing Self-Evolving Model and Technical Route
A previously anonymous Chinese embodied AI team, internally codenamed “MVP” (Make Veritable People), has released several new unedited demo videos showcasing advanced robot capabilities, including nuanced physics understanding, human-like spatial reasoning, and collaborative multi-robot task completion. The team claims the model achieves 80% zero-shot success rates and emphasizes its self-evolutionary ability, which they believe brings human-level cognition to robots.
Why it matters: The claims of 80% zero-shot success and the ability to unify physical and human logic are significant; if true, they represent a leap past the ‘imitation-only’ limitations of many existing models. This isn’t just a slight improvement; it’s a claim about a fundamentally different way robots interact with the physical world, moving beyond rote learning to what looks like genuine understanding and adaptation.
For Western readers: Western robotics and AI firms should closely analyze these demos, particularly the claims around zero-shot generalization and dynamic understanding, as they suggest Chinese research may be pioneering new architectural approaches that circumvent the brute-force data reliance seen in some Western models.
🇨🇳 Cross-Regional Analysis
Xiaomi Announces Foldable Phone and EV Launch on September 7, Clashing with Huawei and Apple Events
Xiaomi will host its autumn flagship new product launch event on September 7, releasing its 18 Fold foldable smartphone, Pad 9 Pro Max tablet, and the Pengcheng N70 Pro, N70 Max, and N90 Max EV models. Both the foldable phone and tablet will feature Xiaomi’s self-developed Xuanjie O3 AI flagship processor. This event directly clashes with Huawei’s HarmonyOS 7 and Mate XT 2 launch on the same day, and precedes Apple’s special event on September 10.
Why it matters: Xiaomi’s simultaneous launch of a foldable phone, tablet with a self-developed AI chip, and three new EV models underscores its strategic push into high-margin segments and its commitment to vertical integration. The direct clash with Huawei’s launch day, rather than avoiding it, indicates a more aggressive competitive posture, particularly in the domestic Chinese market, where both companies are vying for mindshare and market leadership.
For Western readers: Western businesses should recognize that Chinese tech giants are not shying away from direct competition with each other or global players like Apple on major launch days, which signals increased intensity in product cycles and marketing spend, especially in the premium smartphone and emerging EV sectors.
🔺 The Prism
Where US and East Asian technology interests intersect
Startups & Funding
Nvidia Relationship Boosts SoftBank SB Energy’s $50bn IPO Despite Lack of Data Center Revenue
SoftBank Group’s SB Energy unit is targeting a $50 billion valuation for its upcoming U.S. IPO, a figure largely propped up by its strategic relationship with Nvidia and anticipated lease commitments from OpenAI. Despite the high valuation, the company’s IPO filings show no current revenue from its main data center business, suggesting its worth is tied to future AI infrastructure demand. This move highlights SoftBank’s pivot toward providing the foundational energy and data center infrastructure for the burgeoning AI industry.
Why it matters: SoftBank is trying to monetize the AI energy and data center buildout, but they are doing it by relying heavily on the implied future demand from Nvidia and OpenAI, rather than on existing revenue. This is a bet on the AI boom sustaining and the power requirements continuing to scale, positioning SoftBank as a landlord and utility provider for the industry rather than a direct AI technology developer.
For Western readers: Western investors should recognize that current AI infrastructure valuations, particularly in the energy and data center space, are increasingly speculative and tied to future commitments from key AI players like Nvidia and OpenAI rather than established revenue streams. Scrutinize the underlying revenue and long-term contracts of any AI infrastructure play.
AI & Machine Learning
CNeo-Bench: Diagnosing Large Language Models on Chinese Neologisms
Researchers have introduced CNeo-Bench, a new benchmark of 4,759 Chinese neologisms, to evaluate Large Language Models (LLMs). The study found that most of the 18 LLMs tested, including major Chinese and global models, perform poorly on defining and manipulating these unique linguistic expressions, often scoring below 40% on definition generation.
Why it matters: The systematic failure of current LLMs to accurately handle Chinese neologisms means that products leveraging these models will struggle in use cases requiring nuanced understanding of real-time Chinese language, limiting their applicability in areas like social media analysis, content moderation, and culturally relevant conversational AI.
For Western readers: Western AI companies aiming for the Chinese market should understand that merely translating models or using standard Chinese datasets is insufficient; a deeper, mechanism-level understanding of dynamic Chinese linguistics is required for true local competence.
Enterprise & Cloud
Blue Yonder Still Chasing Profitability Five Years After Panasonic Acquisition
Five years after Panasonic Holdings acquired U.S.-based supply chain management software provider Blue Yonder for a substantial sum, the company has yet to achieve the profitability expected from the acquisition. Despite using a series of acquisitions to expand its capabilities under Panasonic’s ownership, Blue Yonder continues to struggle in a market increasingly influenced by advanced AI.
Why it matters: This case illustrates the execution risks involved in Japanese corporations’ attempts to acquire Western software assets for digital transformation. While the strategic intent to move into higher-margin software and recurring revenue is sound, successful integration and the ability to adapt to rapid technological shifts, particularly in AI, remain significant hurdles. It also highlights Panasonic’s struggle to realize its vision of becoming a comprehensive B2B solutions provider.
For Western readers: Western investors should temper expectations for rapid, profitable integration when Japanese conglomerates acquire foreign software firms, especially in fast-moving fields like AI-driven supply chain management where the target’s value proposition can quickly erode or shift. Focus on clear revenue and synergy metrics, not just strategic intent.