East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- NVIDIA’s $3.5 billion joint investment with Taiwan’s MediaTek to develop edge-AI platforms bypasses traditional foundry bottlenecks by embedding custom silicon directly into automotive and smart-device supply chains.
- The debut of UBTECH’s automated assembly line in China, boasting a production rate of one humanoid robot every 10 minutes, shifts the robotics race from laboratory-scale prototyping to high-yield industrial manufacturing.
- Anthropic co-creator David Soria Parra’s keynote at AGNTCon+MCPCon in Tokyo targets Japan’s enterprise software market as the primary testing ground for the Model Context Protocol, aiming to unify fragmented legacy databases across the country’s major conglomerates.
This Issue’s Analysis
The Signal
Google’s Gemini 3.8 Live: Japanese Enterprises Test Real-Time Agentic Automation
Google has launched two new real-time voice interaction models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, available since September 15. Gemini 3.8 Live focuses on cost
Robotics & Automation
China’s First Humanoid Robot Factory Is Producing One Unit Every 10 Minutes
UBTECH Robotics announced the official launch of the world’s first large-scale humanoid robot smart factory, covering 14,000 square meters. The factory is designed to produce one h
Semiconductors & Hardware
Nvidia Invests $3.5B in MediaTek to Build Edge-to-Cloud AI Hardware
NVIDIA and MediaTek are expanding their long-standing partnership to build AI platforms from edge to cloud, including AI infrastructure, local AI computing, and automotive sectors.
Enterprise & Cloud
AI Server Power Surge: Why Grid Delivery Bottlenecks Will Constraint US Data Centers from 2028
TrendForce estimates global data center electricity demand will reach 161GW by 2026, a 31% annual increase, driven primarily by AI servers. The proportion of AI servers in total da
🧩 Pattern This Issue
- Taiwan: MediaTek adopts TSMC 2nm node for next-gen edge AI silicon
- Korea/Taiwan: Nvidia partner Kool-S builds packaging tools targeting Rubin Ultra GPU
- Japan: Nvidia and MediaTek expand $3.5B partnership for edge AI
The center of gravity for AI deployment is shifting to the physical edge, where Taiwan and Korea’s control over next-generation 2nm silicon and advanced packaging tools makes East Asian manufacturing capacity the ultimate gatekeeper of the next hardware cycle.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 AI & Machine Learning
Anthropic’s MCP Co-Creator Visits Japan, Keynote Addresses Future MCP Focus Areas at AGNTCon+MCPCon Japan 2026
David Soria Parra, a co-creator of Anthropic’s Model Context Protocol (MCP), delivered a keynote at AGNTCon+MCPCon Japan 2026 in Tokyo. He discussed MCP’s rapid growth, its role in enabling AI agents to handle long-horizon tasks, and its future evolution towards agent messaging and richer semantics.
Why it matters: The rapid adoption of MCP, with billions of tool calls on Claude and hundreds of millions of SDK downloads monthly, indicates a swift standardization of AI agent communication. This suggests that the ecosystem is coalescing around common protocols for complex multi-step AI tasks, moving beyond isolated model-tool interactions to more integrated agent systems.
For Western readers: Western developers and enterprise strategists should recognize that the ‘connectivity’ layer for AI agents, exemplified by MCP, is quickly becoming standardized. Companies that integrate and contribute to such open protocols will be better positioned to leverage advanced AI agent capabilities across diverse enterprise applications, especially as AI models take on longer, more complex tasks.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 AI & Machine Learning
Naver Applies ‘AI Safety System’ to Real Services, Publishes Safety Report
Naver has implemented its self-developed AI safety system across its commercial services and released its first AI safety report. This move aims to establish transparent governance for ethical and safe AI development and deployment, particularly for its large language model, HyperCLOVA X.
Why it matters: Naver is trying to differentiate itself in the global AI race not just on technical capability, but on a ‘responsible AI’ brand. This is a common strategy among East Asian tech companies, often framed as a way to build domestic trust and navigate potential regulatory headwinds while still competing with Western models.
For Western readers: Western businesses developing or deploying AI should recognize that ‘AI safety’ is increasingly being weaponized as a competitive and policy tool in Asia, not just a technical or ethical concern. This is about market positioning as much as it is about eliminating bias.
🇰🇷 Enterprise & Cloud
Salesforce and NVIDIA Unveil ‘Co-pilot’ CRM Inference Model to Reduce Big Tech Reliance
Salesforce and NVIDIA have jointly introduced ‘Co-pilot,’ an AI inference model specifically designed for Customer Relationship Management (CRM) applications. This collaboration aims to provide businesses with a tailored AI solution, reducing their dependency on broader, general-purpose AI models often offered by major technology companies.
Why it matters: The introduction of Co-pilot by Salesforce and NVIDIA is significant because it allows enterprise customers to run specialized AI inference closer to their data within CRM platforms, potentially offering better performance and data sovereignty compared to generic large language models. This move directly addresses enterprise concerns about data control and the cost associated with general-purpose AI services.
For Western readers: If you are evaluating AI solutions for enterprise applications, prioritize specialized, domain-specific models integrated into platforms like Salesforce over general-purpose LLMs from hyperscalers for improved data governance and application-specific performance.
🇰🇷 Semiconductors & Hardware
Kool-S Develops Ultra-Large Package Bonding Technology Targeting NVIDIA’s Rubin Ultra
Korean semiconductor equipment manufacturer Kool-S has developed a new bonding technology for ultra-large chip packages, specifically designed to meet the advanced requirements of NVIDIA’s upcoming Rubin Ultra AI accelerators. This technology focuses on improving the efficiency and yield of large-scale packaging, which is crucial for integrating high-bandwidth memory (HBM) with powerful GPUs.
Why it matters: The continuous push for larger and more complex AI chip packages by companies like NVIDIA places immense pressure on the backend packaging and assembly ecosystem. Kool-S’s new technology shows Korean equipment makers are keeping pace with leading-edge requirements, positioning them as essential partners for advanced semiconductor manufacturing, which is crucial for maintaining chip performance gains. This is a practical, behind-the-scenes engineering win rather than a splashy product launch.
For Western readers: If you are an AI chip designer, understand that the advanced packaging supply chain for future generations of AI accelerators like Rubin Ultra is being built out now, and Korean equipment suppliers are key players in enabling these next-gen designs. Western foundries or IDMs should assess how this technology could impact their own roadmap for high-density integration.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 Semiconductors & Hardware
MediaTek Unveils First TSMC 2nm Phone Chip, Expands AI Footprint Across Edge and Data Centers
MediaTek introduced its Dimensity 9600 Pro and 9600M mobile chips, with the Pro version being its first utilizing TSMC’s 2nm process for high-end smartphones and the M version on 3nm. The company is also developing an AI accelerator for a major US cloud provider, slated for mass production in Q4 2026, marking a dual strategy for AI in both mobile and data center segments. The 9600 Pro features a new AI processor, showing a 51% performance increase for on-device generative AI pre-processing.
Why it matters: MediaTek’s move to TSMC’s 2nm node for a premium phone chip isn’t just about speed; it’s about cementing its place in the top tier of mobile silicon, where profit margins are higher. The push into data center AI accelerators, especially for a large US cloud provider, indicates a strategic pivot to capture a piece of the server-side AI market, which is currently dominated by a few players. This diversification is a smart hedge against the cyclical nature of the mobile market.
For Western readers: Western mobile OEMs should recognize that MediaTek is serious about competing for their premium phone business, not just the mid-range. For Western cloud providers, MediaTek could emerge as a credible alternative for custom AI accelerators, potentially disrupting the current duopoly and offering more choice in a tight supply chain. Keep an eye on MediaTek’s Q4 2026 AI accelerator production numbers to gauge the scale of this new venture.
🇹🇼 AI & Machine Learning
GPT Most Vulnerable, Claude Most Resilient: Study Reveals 7 LLMs Cannot Withstand Long-Term Misinformation Bombardment
A new study from the University of Arizona, published in Scientific Reports, found that most leading AI models eventually compromise with misinformation when subjected to repeated pressure in long conversations. The research tested GPT-3.5, GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama-3-70B, and DeepSeek-R1, using 100 fabricated statements repeated 50 times to evaluate fallibility, persuadability, and correctability.
Why it matters: The findings show that even leading AI models, including those widely deployed in enterprise settings like GPT-4o and Gemini 1.5 Pro, can be systematically misled over extended interactions. This exposes a significant gap in current AI risk assessments, which typically focus on single-turn accuracy, ignoring the cumulative effect of persistent misinformation in real-world use cases. This is not just about ‘fake news’; it’s about the reliability of any system that accumulates knowledge from ongoing dialogue.
For Western readers: Western enterprises deploying AI in high-stakes environments must reassess their risk models to account for conversational ‘drift’ and misinformation susceptibility, moving beyond single-query benchmarks. Implement independent fact-checking mechanisms and context-clearing for new tasks, especially in critical domains like finance, healthcare, or defense.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 Robotics & Automation
Why Siemens is Central to Factory Robotics Even Though It Doesn’t Build Robots
Siemens’ Xcelerator platform is facilitating the integration of diverse robotics solutions into manufacturing facilities in China, addressing the complexities of multi-vendor environments and existing infrastructure. The platform connects partners providing specific robotic capabilities, such as vision systems, collaborative robots, and mobile robot coordination, with Siemens’ industrial software and automation expertise. This approach enables factories to adopt advanced robotics without extensive line modifications, demonstrated by an application at Siemens Numerical Control (Nanjing) Co. Ltd. where a Galaxy General humanoid robot is used for PCB handling.
Why it matters: Siemens Xcelerator’s strategy in China illustrates how industrial software and automation platforms are becoming the critical layer for integrating disparate robotics hardware and AI algorithms into existing factory operations. It demonstrates that the bottleneck for advanced manufacturing isn’t always hardware, but often the orchestration of complex, multi-vendor systems. This is an integration challenge as much as an AI or hardware one, and platforms like Xcelerator position Siemens to capture value from this layer, even as the robot hardware market becomes more fragmented.
For Western readers: Western manufacturers and robotics companies should recognize that platforms enabling seamless integration of multi-vendor solutions will be key to scaling industrial automation. Rather than focusing solely on robot performance or AI models, businesses should prioritize how their solutions fit into open ecosystems and address the practicalities of deployment in brownfield factories, especially in competitive markets like China.
🇨🇳 AI & Machine Learning
Handing Memory Management to the CPU Could Speed Up Large Models
Chinese AI service providers are facing performance bottlenecks as large language models (LLMs), particularly AI agents, demand increasingly larger KV Caches for long-context inference. QbitAI reports that the growing size of KV Cache, which stores intermediate attention states, consumes significant GPU memory and leads to costly re-computation (Prefill) when data is evicted, directly impacting Time To First Token (TTFT). The article proposes a tiered memory strategy that offloads KV Cache from expensive GPU HBM to CPU DDR memory and SSDs, leveraging CPU for memory management to free up GPUs for token generation.
Why it matters: The operational cost and scalability of LLM inference is a major concern globally, and this article points to a practical solution: optimizing memory management across different hardware tiers. China’s AI ecosystem, with its massive user base and drive for cost-efficient deployments, often reveals production-level challenges and solutions that eventually become relevant worldwide. This approach shifts the burden of memory off GPUs, allowing more efficient use of expensive compute resources.
For Western readers: Western AI infrastructure providers and enterprise users should anticipate similar challenges as their own LLM applications, especially agents, scale. They should evaluate hybrid memory architectures that intelligently distribute KV Cache across GPU, CPU, and other storage tiers to improve cost-efficiency and reduce TTFT for long-context applications.
AI & Machine Learning
Huawei opens beta testing for XiaoYi Work AI assistant across phones, tablets and PCs
Huawei has launched beta testing for XiaoYi Work, an AI work assistant designed for office tasks, coding, and creative work. The assistant operates across Huawei’s HarmonyOS ecosystem, including phones, tablets, and PCs, allowing users to assign and monitor tasks seamlessly between devices.
Why it matters: Huawei’s focus on a multi-device, integrated AI assistant under HarmonyOS indicates its strategy to build a comprehensive, closed-loop software ecosystem. This mirrors earlier efforts by Western tech giants but is critical for China to develop resilient domestic alternatives against ongoing US technology restrictions.
For Western readers: Western enterprise software and cloud providers should expect Huawei to aggressively push XiaoYi Work within China, potentially creating a significant domestic alternative that prioritizes data sovereignty and tight integration with other Huawei services, making it harder for foreign solutions to gain traction.
🔺 The Prism
Where US and East Asian technology interests intersect
Semiconductors & Hardware
MediaTek says new AI phone chip reduces memory use amid supply crunch
MediaTek has launched its new Dimensity 9600 Pro smartphone chip, designed to optimize AI computing performance while reducing memory usage. This design choice is a direct response to the ongoing global supply crunch and rising costs of memory chips, impacting the broader East Asian electronics supply chain.
Why it matters: MediaTek is reacting to the reality of the memory market, where tight supply and rising prices are not temporary. Designing a chip to use less memory is a hard engineering choice that directly impacts bill-of-materials costs for their customers, predominantly Chinese and other Asian phone makers, giving them a tangible advantage in a constrained environment. This reflects a shift from maximizing raw performance to optimizing for supply chain resilience and cost efficiency, which is a major concern for Asian electronics assemblers.
For Western readers: Western smartphone brands, particularly those sourcing components from the same East Asian supply chain, should expect similar memory cost pressures and potentially slower adoption of memory-intensive AI features in devices that don’t command premium prices.
AI & Machine Learning
The Universe of Universes: Benefit Yield Functions, Implosion Thresholds, and Infrastructure-Aware Optimization in Multi-LLM Systems
Researchers have proposed the “Universe of Universes” (UoU) framework for optimizing multi-LLM systems, introducing concepts like the Benefit Yield Function (BYF) and an “implosion threshold.” This framework aims to formalize how adding more LLMs to an ensemble affects performance, identifying the point where further additions degrade rather than improve aggregate output.
Why it matters: The concept of an ‘implosion threshold’ directly challenges the prevailing assumption in some parts of the East Asian AI industry that simply scaling up or combining more LLMs will always lead to better results. This research suggests there are hard limits to performance gains, which has implications for resource allocation and architectural decisions among companies like Tencent and SoftBank that are building large-scale AI infrastructure.
For Western readers: Western businesses should understand that East Asian AI developers will need to shift from a ‘more is better’ mindset to a more optimized approach for multi-LLM systems, meaning that the demand for high-performance computing might not scale linearly with the number of models deployed.
AI & Machine Learning
TASPO: Reconciling Process Supervision with Outcome-Based Credit in Agentic Policy Optimization
Researchers, including Jingxiao Yang and colleagues, have introduced a new method called TASPO (Trajectory-Aware Supervision for Policy Optimization) for improving the training of language-model agents. TASPO addresses the ‘supervision-credit gap‘ in reinforcement learning by converting privileged training information into outcome-grounded action credit, leading to more stable policy optimization and better generalization.
Why it matters: While this is a research paper from arXiv, not a commercial product, the methodology presented here offers a concrete step forward in making AI agents more robust and capable. Better agentic policy optimization reduces the ‘trial and error’ inefficiency in training, which directly impacts the compute resources and time required to develop advanced AI models. For East Asian AI developers, this means the potential for more efficient use of expensive compute infrastructure and faster iteration cycles in model development.
For Western readers: Western AI researchers and developers should integrate TASPO’s methodology, or similar approaches for outcome-grounded credit assignment, into their agent training pipelines to improve efficiency and model performance, particularly for long-horizon tasks.