East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- Samsung’s decision to lock up 70% of its memory production capacity through 2031 in long-term contracts leaves Nvidia facing prolonged supply constraints that will force the GPU giant to diversify its high-bandwidth memory sourcing.
- Japan’s record ¥8.89 trillion defense budget request accelerates the integration of AI command-and-control systems and autonomous drones, shifting the country’s defense stance from passive deterrence to active, software-driven preparedness.
- Nvidia’s $3.4 billion investment in Taiwan’s MediaTek extends its reach beyond data center GPUs into edge-computing silicon, positioning the duo to dominate the next generation of AI-enabled automotive and consumer hardware.
This Issue’s Analysis
The Signal
Samsung Locks 70% of Memory Capacity to 2031, but Nvidia Still Faces Shortages
Samsung Electronics has reportedly allocated 70% of its memory production capacity through 2031 to long-term agreements (LTAs) with major tech companies like Nvidia, Microsoft, and
Semiconductors & Hardware
Nvidia’s $279B Supply Commitments: Secure Capacities for the Rubin Platform’s 70% Growth Target
NVIDIA reported Q2 FY2027 revenue of $96.22 billion, up 106% year-over-year, with data center revenue reaching $89 billion. The company anticipates Q3 revenue of $108 billion and,
Policy & Regulation
Japan’s ¥8.89T Defense Budget: Ministry of Defense to Fund AI Command Systems and Drones
Japan’s Ministry of Defense has requested a record ¥8.89 trillion (approximately $60 billion USD) for the 2027 fiscal year, aiming to integrate AI into Self-Defense Forces command
Semiconductors & Hardware
Nvidia’s ₩4.7T MediaTek Bet: Why the GPU Giant Is Backing Taiwanese SoC Design
NVIDIA is reportedly investing 4.7 trillion Korean won (approximately $3.4 billion USD) into MediaTek, a Taiwanese chip designer. This strategic investment aims to expand NVIDIA’s
🧩 Pattern This Issue
- Korea/Taiwan: Samsung secures 70% of memory capacity through 2031
- Korea/Taiwan: South Korean exports surge 68.7% on strong chip demand
- Korea/Taiwan: Nvidia invests 4.7 trillion won in MediaTek chip design
Nvidia’s aggressive long-term capacity lock-ups and strategic Asian design partnerships are systematically monopolizing the hardware supply chain, leaving Western competitors structurally locked out of the physical infrastructure required to scale next-generation AI.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 AI & Machine Learning
Chinese MiniMax H3 Max Video AI Now Available for Real-Time Generation via fal.ai, Enabling ‘Endless AI Broadcasts’
Chinese AI firm MiniMax released its H3 video generation model, which US-based fal.ai has optimized as H3 Max for faster output. fal.ai offers a free trial site and API for H3 Max, capable of generating 480p, 5-second videos with audio in about 3 seconds, significantly faster than real-time. fal.ai also launched ‘fal.live,’ a livestream site demonstrating continuous AI-generated programming that evolves based on viewer input.
Why it matters: The capability to generate video faster than real-time, and continuously, shifts the paradigm for AI content creation from static generation to dynamic, interactive media. While Western AI discourse often centers on model size and benchmark scores, the practical deployment and speed enhancements, as seen with fal.ai’s H3 Max optimization, are what enable new commercial use cases like the ‘endless AI broadcast’ demonstrated by fal.live.
For Western readers: Western media and entertainment companies, especially those in advertising and gaming, should evaluate MiniMax H3 Max and similar real-time generation tools from Asia, as these are moving quickly from concept to deployable solutions that can create immersive, viewer-driven content economically.
🗾 AI & Machine Learning
Anthropic’s AAR: AI Develops Superior Safety Methods in Six Hours, Outperforming Human Researchers
Anthropic’s research team has unveiled ‘Automated Alignment Researcher (AAR),’ an AI agent built on Claude Opus 4.8, which autonomously conducts AI safety research. AAR can identify and mitigate various alignment failures, developing methods that significantly reduce issues like deception and hallucination. In comparisons, AAR achieved superior results in approximately six hours, outperforming 28 human researchers with an average of 2.5 years of experience given eight hours for the same task.
Why it matters: The Japanese framing of this story, while reporting on a US-based AI company, subtly emphasizes the ‘automation of research’ and the ‘acceleration of future progress’ rather than just the benchmark scores. This aligns with Japan’s broader industrial policy focus on using AI to augment R&D and address labor shortages, positioning AI as a tool for efficiency and self-improvement, which is a more practical-minded framing than the typical Western obsession with frontier model capabilities.
For Western readers: Western AI developers and policymakers should recognize that AI safety advancements may increasingly come from autonomous AI systems rather than purely human-driven research, requiring a shift in how R&D is structured and regulated.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 Semiconductors & Hardware
Samsung Electronics Aims to Break Memory Limits with 3D Stacking HBM
Samsung Electronics is focusing on advanced 3D stacking technology for High Bandwidth Memory (HBM) to overcome current memory capacity and performance limitations. This approach involves vertically integrating memory dies, enabling greater data throughput and density, which is critical for next-generation AI and high-performance computing applications.
Why it matters: The race in HBM is not just about raw capacity, but about the manufacturing techniques that deliver it at scale and yield. Samsung’s emphasis on 3D stacking highlights a key battleground in process technology where efficiency and reliability are paramount for meeting the insatiable demand from AI applications. This is a head-on challenge to SK hynix, which currently leads in HBM market share, especially with its HBM3 and HBM3E products.
For Western readers: If you are a Western AI chip designer or data center operator, your future roadmaps will depend on the successful and timely mass production of these advanced HBM types, dictating when and at what scale you can deploy next-generation AI infrastructure. Watch for actual product announcements, not just technology development updates.
Semiconductors & Hardware
South Korea’s August Exports Surge 68.7% to US$98.25 Billion on Strong Chip Demand
📊 Featured Chart
Source: Ministry of Trade, Industry and Resources
South Korea’s exports in August jumped 68.7% year-on-year to US$98.25 billion, largely driven by a 209% surge in semiconductor shipments, reaching $46.65 billion. This marks the third consecutive month that semiconductor exports have exceeded $40 billion, fueled by hyperscaler investments in AI infrastructure.
Why it matters: This report confirms the ongoing strength of the memory semiconductor cycle, driven specifically by AI infrastructure buildouts. The significant growth in semiconductor exports to both China and the US points to persistent demand for South Korean components in critical AI hardware, regardless of origin, and indicates that supply constraints, not demand, are the primary limiting factor for chip producers.
For Western readers: Western technology companies should factor in continued strong demand and potentially tight supply for high-bandwidth memory (HBM) and other advanced memory products as hyperscalers globally continue to scale AI infrastructure. Plan for lead times accordingly and expect HBM pricing to remain robust.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 AI & Machine Learning
Anthropic Secures $35 Billion Cloud Deal, Nvidia Supplies Chips and Holds Lease
AI developer Anthropic has signed a $35 billion cloud computing contract with Nvidia-backed cloud service provider Lambda. This complex deal involves data centers built by Hut 8, where Nvidia holds the lease, highlighting Nvidia’s growing role in enabling AI companies like Anthropic to access expensive computing resources. Anthropic previously secured a $45 billion contract with Nscale, another Nvidia-backed cloud provider, indicating significant demand for AI computing capacity.
Why it matters: Nvidia is leveraging its market dominance in AI chips to become a kingmaker in the cloud computing space, especially for high-growth AI startups that might otherwise struggle to finance multi-billion dollar compute needs. This move also shows Nvidia competing with traditional cloud providers by backing and enabling ‘neocloud’ companies, effectively controlling access to GPU compute from multiple angles. It represents a subtle, but significant, shift in how AI infrastructure is financed and deployed.
For Western readers: Western businesses in AI development should recognize that Nvidia is building an ecosystem that extends beyond hardware sales into infrastructure and financing, potentially reducing the leverage of traditional hyperscale cloud providers. Prepare for a future where Nvidia dictates more of the terms for accessing cutting-edge AI compute, regardless of which specific cloud provider you contract with.
🇹🇼 Semiconductors & Hardware
TSMC’s Comprehensive Roadmap for System-Level Miniaturization in the AI Era
📊 Featured Chart
Source: TSMC, SEMICON Taiwan 2026
At SEMICON Taiwan 2026, TSMC’s Dr. C.Y. Chen outlined the company’s full-dimension technology roadmap, shifting focus from single-chip SoC to system-integrated miniaturization. This strategy uses advanced silicon processes, 3D stacking (SoIC), silicon photonics (COUPE), and CoWoS platforms to overcome power, transmission, and cooling challenges driven by escalating AI computing demands.
TSMC plans to advance its N2 process, launch A16 and A14 nodes, scale SoIC to 4.5 micron in 2029, and enhance HBM with N12 and N3 base dies. The company also announced volume production of its first-generation COUPE and the largest CoWoS packaging platform, aiming for further expansion and power/thermal management through integrated capacitors, voltage regulators, and liquid cooling solutions.
Why it matters: This announcement from TSMC isn’t just about future nodes; it details concrete steps to manage the power and thermal challenges that threaten to bottleneck AI at the system level. The emphasis on 3D integration, silicon photonics, and advanced cooling shows where the physical limits are being pushed. It confirms that the performance race has moved beyond just transistor counts to how efficiently entire chiplets and memory can communicate and dissipate heat within a single package.
For Western readers: Western AI chip designers need to closely align their roadmaps with TSMC’s packaging and integration advancements, especially regarding CoWoS scaling and cooling, as these will dictate achievable performance and package power budgets. Assume that TSMC’s capacity for advanced packaging will continue to be the primary constraint for high-performance AI chip deployment into the early 2030s.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 Robotics & Automation
Tsinghua AIR and Domain Transform Introduce Zeva: Embodied In-Context Causal Learning for Self-Evolving WAMs
Tsinghua University’s AIR (Institute for Artificial Intelligence Research) and Domain Transform have unveiled Zeva, the first embodied ‘Whole-Arm Manipulation‘ (WAM) model to achieve In-Context Causal Learning (ICCL). Zeva enables robotic agents to continuously learn from their own interaction experiences without updating model weights, significantly improving task success rates in both simulated and real-world chemical lab environments.
Why it matters: The core insight of Zeva — that models can ‘self-evolve’ by learning causal contexts rather than requiring explicit weight updates — changes the calculus for robot deployment. This approach reduces the dependency on continuous re-training and large datasets, offering a path to more autonomous and adaptable robotic systems, particularly for delicate manipulation tasks where precision and real-time adaptation are crucial.
For Western readers: Western researchers and developers in embodied AI should evaluate Zeva’s In-Context Causal Learning framework as a potential alternative to traditional fine-tuning or test-time training, especially for applications requiring rapid adaptation and one-shot learning from human demonstrations in dynamic environments.
🇨🇳 AI & Machine Learning
MiniMax’s H3 Max Model Achieves Real-Time AI Video Generation, Enabling New Commercialization Paths
Chinese AI company MiniMax, in collaboration with AI inference specialist Fal, has launched H3 Max, a video generation model capable of creating 5-second videos in under 3 seconds. This speed enables real-time AI-powered livestreaming and short video applications, showcased by developers creating AI-generated live channels and interactive video apps.
Why it matters: The speed breakthrough from MiniMax and Fal’s H3 Max model is more than just an incremental improvement; it changes the underlying economics of video production by eliminating the wait time. This opens up new commercial avenues for interactive content and live broadcasting that were previously unfeasible, positioning MiniMax to capitalize on a market where ‘faster’ becomes ‘fundamentally different’.
For Western readers: Western content platforms and creators should recognize that real-time AI video generation is moving beyond niche applications and could soon underpin scalable, dynamic content, reducing reliance on traditional production pipelines and potentially shifting audience engagement models.
Policy & Regulation
Huawei’s H1 Revenue Up 9.6% as R&D Spending Reaches RMB121.4 Billion Amid Profit Decline
📊 Featured Chart
Source: Huawei H1 2026 report
Huawei Investment & Holding reported a 9.55% year-over-year increase in first-half revenue to RMB467.82 billion. Despite this, net profit fell significantly from RMB37.05 billion to RMB23.43 billion, while research and development spending surged by approximately 25% to RMB121.38 billion, representing 25.9% of total revenue.
Why it matters: Huawei’s decision to sharply increase R&D spending despite declining profits reflects a strategic imperative driven by geopolitical pressures rather than pure commercial optimization. This signals a sustained effort to develop domestic alternatives for critical technologies, especially semiconductors and networking components, directly impacting global supply chain resilience and the competitive landscape for non-Chinese vendors.
For Western readers: Western companies relying on global technology supply chains, particularly in telecommunications equipment and semiconductors, should assume Huawei’s capability in key component areas will continue to advance, necessitating a re-evaluation of long-term competitive threats and potential market shifts.
🔺 The Prism
Where US and East Asian technology interests intersect
AI & Machine Learning
Automated Multi-Agent LLM Framework for Multilingual Climate-Health Literature Analysis
Chinese researchers have developed a multi-agent large language model (LLM) framework designed to automate the analysis of multilingual scientific literature, specifically targeting the interdisciplinary climate-health field. The framework, which includes specialized agents for evaluation, extraction, and review, achieved an F1 score of 0.92 in core information extraction when validated on a Chinese-English corpus of over 32,000 papers.
Why it matters: China consistently frames its AI advancements in terms of national interest and domestic capability. This research, while presented as a general scientific tool, fits squarely into Beijing’s strategy to build ‘digital infrastructure’ for core national priorities, and to ensure that the tools and data platforms used for critical analysis remain within its control. The F1 score of 0.92 on a substantial bilingual dataset indicates a functional system, not just a theoretical one.
For Western readers: Western governments and research institutions should recognize that China is building AI frameworks not just for commercial applications but as foundational tools for scientific and strategic analysis, potentially creating data analysis ecosystems that operate independently of Western-developed platforms and methodologies.
AI & Machine Learning
Evaluating Large Language Models on China’s GAOKAO Benchmark
Chinese researchers introduced GAOKAO-Bench, a new benchmark using questions from China’s national college entrance exam, GAOKAO, to evaluate LLMs. The study found that models like GPT-4 achieved competitive scores on the exam but showed performance disparities across subjects, with a moderate consistency between human and model grading for subjective questions.
Why it matters: The development of GAOKAO-Bench highlights a strategic move within the Chinese AI community to create culturally specific benchmarks, potentially pushing Chinese LLM development towards better performance on localized tasks crucial for their domestic market. This aligns with China’s broader goal of fostering ‘data sovereignty’ and developing AI tailored to national needs, rather than solely relying on Western-developed benchmarks.
For Western readers: Western AI developers and businesses targeting the Chinese market should recognize that evaluation criteria for LLMs in China are increasingly localized, meaning a competitive product must perform well on benchmarks like GAOKAO-Bench, not just global English-language tests.
Semiconductors & Hardware
Kioxia and Sandisk Plan $31B Japan Investment to Expand NAND Capacity
Kioxia and Sandisk are planning over $31 billion (5 trillion yen) in joint investments in Japan through 2032 to expand advanced NAND flash memory production at Kioxia’s Yokkaichi and Kitakami plants. This aims to ensure stable supply amidst increasing demand driven by AI and data-intensive applications, subject to Japanese government support.
Why it matters: This spending commitment shows that Japan’s METI isn’t just funding foundries; it’s also backing critical memory capacity with established players. The Kioxia-Sandisk partnership has a proven track record of joint investment, which gives this announcement more weight than a typical consortium announcement. The stated goal of supporting ‘an AI-driven society’ is code for ensuring domestic access to essential components for data centers and advanced computing initiatives.
For Western readers: Western companies relying on NAND flash for AI infrastructure should view this investment as a positive step towards supply stability in the long term, but understand that government support will steer some of this capacity towards Japan’s strategic priorities.