East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- Mitsui Kinzoku’s plan to establish a new production base in Malaysia for its VSP copper foil directly addresses the severe material bottlenecks facing high-frequency AI chip substrates.
- The divergent approaches to next-generation HBM base dies presented by Samsung and SK Hynix at Hot Chips 2026 show that memory makers are split on whether to internalize or outsource advanced logic manufacturing.
- Softcreate becoming the first Japanese partner to earn Microsoft’s Secure AI Productivity specialization reflects a growing push by domestic enterprises to adopt sovereign, compliance-first deployments of generative AI tools.
This Issue’s Analysis
The Signal
Samsung and SK Hynix Split on HBM4, Forcing a Design Choice for AI Chipmakers
At Hot Chips 2026, Samsung Electronics emphasized ‘active base die’ technology, transforming the HBM base die from a passive connector into an active logic chip using 4nm logic pro
Semiconductors & Hardware
TSMC’s Power Demand Will Hit 28.8B kWh in 2025, Consuming 10% of Taiwan’s Electricity
TSMC’s latest sustainability report shows its global power consumption is projected to reach 28.77 billion kilowatt-hours (kWh) in 2025, a 12.6% increase from 2024. This figure rep
Semiconductors & Hardware
SK hynix’s HBM4 Packaging Roadmap: Doubling TSVs to Meet Western AI Accelerator Demand
SK hynix detailed its advanced packaging roadmap for High-Bandwidth Memory (HBM) at Hot Chips 2026, focusing on HBM4 and beyond. The company plans to double TSV counts, increase I/
Semiconductors & Hardware
Zhongcheng Hualong’s 10K-Card Supernode: How China Scales AI Inference Without Advanced U.S. Chips
Zhongcheng Hualong (中诚华隆) has officially launched its new HL200 AI inference chip and a ‘supernode’ intelligent computing cluster solution in Beijing. This marks a shift from singl
🧩 Pattern This Issue
- Korea/Taiwan: Samsung and SK Hynix split on next-gen HBM architecture at Hot Chips 2026
- Japan: Mitsui Kinzoku expands high-end copper foil production for AI servers in Malaysia
- Korea/Taiwan: TSMC projected power demand hits 28.8 billion kWh amid AI scaling
The physical bottlenecks of AI scaling are shifting from pure compute silicon to the basic physical constraints of power transmission, packaging, and advanced materials, handing critical leverage to East Asian industrial suppliers that control these underlying hardware links.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 Enterprise & Cloud
Softcreate Becomes First in Japan to Achieve Microsoft’s Secure AI Productivity Specialization, Enhancing AI Implementation Support
Softcreate announced it is the first company in Japan to acquire Microsoft’s “Secure AI Productivity” Specialization, a key recognition within the Microsoft AI Cloud Partner Program. This specialization certifies Softcreate’s ability to build secure operational foundations for enterprise AI adoption, encompassing ID and device management, threat protection, and information safeguarding.
Why it matters: Microsoft’s specialization program for AI security is not just about technical certifications; it’s a channel strategy to accelerate enterprise AI adoption with trusted partners. For Softcreate, being the first to achieve this in Japan helps them capture early demand from Japanese companies, which prioritize security and managed implementation support over rapid, unvetted deployment, especially when integrating new technologies like generative AI into their established systems. The focus on zero-trust principles reflects deep-seated caution in the Japanese market.
For Western readers: Western enterprises deploying Microsoft AI services in Japan should recognize that local partners like Softcreate are framing their value proposition around comprehensive security and implementation support, rather than just raw AI capabilities; adjust your engagement strategy to align with this local emphasis.
🗾 Semiconductors & Hardware
Toyo Technica and Partners Launch Joint Research to Advance Quantum Sensing for Social Implementation
Toyo Technica, the Quantum Science and Technology Agency (QST), and Type-I Technologies have initiated a joint research project focused on quantum sensing technology. The collaboration aims to commercialize quantum sensors and devices using diamond NV centers to accelerate their practical application across various industries. This initiative addresses current challenges in NV diamond supply and quality, leveraging each partner’s expertise in quantum sensing, diamond NV center technology, and measurement.
Why it matters: This initiative represents a concrete step in Japan’s national quantum strategy, moving beyond theoretical research to focus on the practical challenges of commercialization, such as ensuring a reliable supply of high-quality NV diamonds. The emphasis on ‘social implementation’ (社会実装) in the title itself shows a clear focus on deploying the technology in industries like automotive and semiconductor analysis, rather than just basic science.
For Western readers: Western companies in high-precision measurement, EV manufacturing, and semiconductor failure analysis should watch for the emergence of advanced diamond quantum sensors from Japan, as they could offer superior capabilities for diagnostics and process control.
🗾 Semiconductors & Hardware
Mitsui Kinzoku to Boost Copper Foil Production for AI Infrastructure, New Malaysia Base Planned
📊 Featured Chart
Production targets
Mitsui Kinzoku is significantly increasing its production capacity for specialized copper foils, VSP and MicroThin, used in high-frequency substrates and optical modules for AI infrastructure. The company plans to continuously expand capacity from 2028 onward, including acquiring new land near its Malaysian factory for a new production base after 2031, with investments totaling over 37 billion yen.
Why it matters: This move strengthens Mitsui Kinzoku’s position as a vital enabler for high-performance AI hardware, particularly for high-frequency data transmission and advanced packaging. It reflects a strategic acceleration of existing plans, indicating that demand from AI infrastructure is outpacing even recent aggressive projections, and that the company is responding to explicit signals from major tech players.
For Western readers: Western companies relying on advanced copper foils for AI server components or optical modules should anticipate a more stable supply from Mitsui Kinzoku by 2028-2030, but also recognize the upstream dependency on specialized Japanese materials for high-performance computing.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 AI & Machine Learning
NVIDIA Overcomes AI Model Routing Bottleneck by Successfully Transmitting KV Cache via Linear Operations
NVIDIA has announced a breakthrough in AI model inference by successfully transmitting the Key-Value (KV) cache using linear operations, addressing a critical bottleneck in large language model (LLM) routing. This new method aims to improve the efficiency and speed of processing long sequences and complex AI tasks by reducing the computational load associated with KV cache management.
Why it matters: NVIDIA’s solution allows for more efficient management of the KV cache, which directly translates into faster and more cost-effective inference for LLMs. This is important for AI developers and enterprises that are currently struggling with the computational demands of deploying large-scale AI models, particularly for applications requiring long context windows.
For Western readers: Western businesses heavily reliant on NVIDIA GPUs for AI inference, especially those deploying LLMs, should anticipate reduced operational costs and increased throughput for their AI workloads, potentially enabling more ambitious applications. This could affect architectural decisions in cloud AI infrastructure and enterprise AI solution design.
🇰🇷 Startups & Funding
NVIDIA Discussing Multi-Billion Dollar Investment in Perplexity, Valuing Company at ₩41 Trillion
NVIDIA is reportedly in talks to invest several billion dollars in the generative AI search startup Perplexity AI, which could value the company at approximately ₩41 trillion (around $30 billion). This investment would be part of a larger funding round aimed at accelerating Perplexity’s growth and competitive positioning in the AI search market.
Why it matters: NVIDIA investing in an AI search application like Perplexity shows the company is not just selling chips, but actively shaping the generative AI application layer and trying to ensure that layer is built on its CUDA stack. This isn’t just about revenue from chip sales; it’s about controlling the software and service ecosystem that makes those chips indispensable. For Western readers, it’s a clear signal that NVIDIA sees value in vertical integration through strategic investment.
For Western readers: If you are developing or funding generative AI applications, assume NVIDIA will continue to strategically invest in its ecosystem partners, increasing competitive pressure on those outside its orbit and potentially setting de facto standards for the application stack. This means aligning with NVIDIA’s platform is becoming increasingly important for scaling.
🇰🇷 Semiconductors & Hardware
LS Electric to Supply ₩230.9 Billion in Power Facilities to US Big Tech Data Centers
LS Electric, a South Korean electrical equipment manufacturer, secured a contract worth 230.9 billion Korean won (approximately $167 million USD) to supply power facilities to data centers operated by a major US technology company. The agreement, disclosed by LS Electric, is for facilities to be delivered over a period of 18 months, concluding by February 2028.
Why it matters: The specific customer is not named, but the scale of this deal indicates a major US technology firm is making substantial investments in data center expansion, likely driven by AI infrastructure needs. For LS Electric, it confirms their competitiveness in providing critical power solutions globally, moving beyond traditional industrial clients to hyperscale data centers.
For Western readers: Western data center operators and AI infrastructure planners should recognize that the global supply chain for critical power equipment is tightening as AI-driven demand accelerates, requiring diversified sourcing strategies and longer lead times for complex deployments.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 AI & Machine Learning
Nvidia Announces AVO Architecture Enabling AI Models to Handle Long-Duration Tasks
Nvidia introduced its Agentic Variation Operators (AVO) architecture, an AI agent system designed to enable models to autonomously execute complex, multi-step tasks over extended periods. AVO enhances AI models with persistent memory and a supervision mechanism, allowing them to take actions, observe results, and refine strategies continuously. This system was initially developed for software engineering and GPU core optimization, demonstrating autonomous operation for up to seven days in one experiment.
Why it matters: Nvidia’s focus on an agent architecture rather than just another model release highlights a pragmatic shift in AI development toward operational robustness and real-world task completion. This perspective suggests that the critical bottleneck for practical AI adoption isn’t just raw computational power, but the ability of AI systems to sustain complex operations reliably over time, a nuance often overlooked in Western discussions heavily centered on model-centric benchmarks.
For Western readers: Western businesses investing in AI solutions should re-evaluate their strategies to prioritize not just model selection, but also the agentic frameworks and operational feedback loops that will govern how those models perform in continuous, real-world tasks, as this is where actual value is being created.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 AI & Machine Learning
Alibaba’s Wan3.0 Video Large Model Officially Launched, Industry Praises ‘Stable, Realistic, Textured’
📊 Featured Chart
Limited-time 70% off pricing, 2026-08-24 to 2026-09-23
Alibaba officially launched its Wan3.0 video generation large model today, capable of generating 30-second videos and supporting various document formats as input. The model aims for high realism and stability, with early enterprise users describing it as “stable, realistic, and textured,” particularly for maintaining consistency in characters and scenes over longer durations.
Why it matters: Alibaba is positioning Wan3.0 not just as a creative tool, but as a production-grade asset for industries like short-form video, film, and advertising. Its focus on generating continuous, stable content across 30-second sequences and supporting multiple input formats speaks to a pragmatic drive for integration into existing enterprise workflows, rather than solely chasing benchmark scores.
For Western readers: Western businesses evaluating generative AI for video production should recognize that Chinese providers like Alibaba are rapidly advancing in delivering models that prioritize industrial-scale content consistency and cost-efficiency over raw photorealism in single frames. This shifts the competitive landscape toward practical enterprise integration, not just model capability.
🇨🇳 Cross-Regional Analysis
Apple Foldable iPhone Nears Launch, New iPhone Models Expected to Rise in Price; Xiaomi to Announce Xuanjie Chip Progress
Apple is reportedly preparing to launch its first foldable iPhone and considering price increases for some iPhone models next month, while also adjusting its Siri and Vision Pro teams, affecting approximately 200 positions. Concurrently, Xiaomi is set to provide an update on its ‘Xuanjie’ (玄戒) chip family today, following the shipment of over one million Xuanjie O1 chips across three devices. Separately, DeepSeek adjusted its API pricing to offer off-peak rates all day on weekends, and OpenAI temporarily lowered GPT-5.6 Sol API prices by 20-33%.
Why it matters: Apple’s reported price increases for new iPhones, alongside a foldable model launch, indicate a strategy to maintain premium positioning and recoup R&D costs in a competitive market. The reduction in its Vision Pro gaming and content teams, despite claims of continued commitment, suggests a recalibration of investment strategy, shifting more responsibility for content development to external partners. Xiaomi’s chip developments, particularly after shipping over a million units of its first generation, demonstrate tangible progress in domestic component substitution, posing a long-term challenge to foreign suppliers in China’s vast electronics market. Meanwhile, the AI pricing adjustments by DeepSeek and OpenAI show the intensifying battle for developer mindshare and market share in the LLM space, where cost-efficiency is becoming a key differentiator.
For Western readers: Western hardware manufacturers should anticipate increased competition from vertically integrated Chinese players like Xiaomi, who are making verifiable progress in chip development. Western developers and enterprises leveraging AI models should factor in the ongoing price reductions from providers like DeepSeek and OpenAI, which can significantly alter the cost dynamics of integrating advanced AI capabilities. Companies looking to partner with Apple on Vision Pro content should note the shift towards external content creation opportunities rather than relying on Apple’s internal development.
🔺 The Prism
Where US and East Asian technology interests intersect
Semiconductors & Hardware
Nvidia’s Inference Pivot Reaches Rebellions in Korea
Nvidia is shifting its focus from training to inference in the AI chip market, a move impacting South Korean semiconductor companies that have heavily invested in HBM production for training GPUs. This pivot is causing concern among some Korean firms as it could diminish demand for high-end HBM if the inference market leans towards less memory-intensive solutions or different architectures.
Why it matters: For South Korean memory makers, this isn’t just a market trend; it’s a direct threat to their core business model, which has banked on HBM becoming the indispensable component for all AI workloads. Nvidia’s move suggests HBM might not be as universally critical for inference as it is for training, forcing Korean firms to reassess their investment strategies and potentially diversify.
For Western readers: Western investors and system builders should adjust their assumptions about the long-term demand curve for HBM, recognizing that Nvidia’s strategic shift could dampen future growth for memory manufacturers and favor alternative chip architectures for inference.
Workforce & Culture
Samsung Electronics faces internal divide over huge bonus gap
Samsung Electronics is experiencing internal conflict due to a significant disparity in bonuses between its highly profitable semiconductor division and its struggling smartphone unit, leading to employee unrest and calls for unity from Chairman Lee Jae-yong.
Why it matters: The internal unrest at Samsung over compensation structures exposes a fundamental tension within diversified tech giants when one division vastly outperforms others. While the chip division’s success funds the company, the smartphone unit’s employees feel undervalued, creating a retention risk if top talent in the less profitable but still strategic divisions decides to seek opportunities elsewhere.
For Western readers: Western companies competing with Samsung in semiconductors or smartphones should monitor whether this internal divide affects Samsung’s ability to attract and retain talent in key product areas, which could create opportunities to poach dissatisfied engineers or managers.