East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- Samsung Electronics’ $235 million co-investment in Dutch AI chip startup Euclid positions the South Korean giant to challenge Nvidia’s dominance in high-bandwidth memory integration for next-generation edge devices.
- The three-way alliance among Alphabet’s Waymo, Japanese ride-hailing leader GO, and Japan Taxi aims to deploy level-4 autonomous vehicles in Tokyo by 2027, bypassing Japan’s acute taxi driver shortage by integrating US self-driving software with local fleet operations.
- Nvidia’s direct investment talks with South Korean AI chip designer Rebellions represent a strategic move to co-opt potential low-power inference rivals before they can scale within the East Asian hardware ecosystem.
This Issue’s Analysis
The Signal
Waymo’s Tokyo Expansion: Why a Taxi Shortage Forced Japan to Approve Autonomous Rides by 2027
Japanese taxi-hailing app GO, US autonomous driving company Waymo, and Japan Taxi announced a strategic partnership to launch fully driverless (Level 4) taxi services in Tokyo by 2
Semiconductors & Hardware
Nvidia’s Rebellions Talks: A Bid to Control Korea’s AI Chip and Memory Value Chain
NVIDIA is reportedly in early-stage discussions with Rebellions, a Korean startup specializing in inference NPUs, for potential technology partnership, investment, or even acquisit
Semiconductors & Hardware
MediaTek’s 2nm Dimensity 9600 Pro Challenges U.S. Dominance in Premium Edge AI
MediaTek officially announced its Dimensity 9600 Pro and 9600 M flagship 5G Agentic AI chips. The 9600 Pro, built on TSMC’s 2nm process, features an all-big-core CPU architecture a
AI & Machine Learning
Elon Musk’s Grok Roadmap: Why xAI Backs Dario Amodei’s Push for an AI Slowdown
Elon Musk, CEO of SpaceXAI, shared updates on his company’s Grok AI model via X, stating that Grok 4.8’s training would complete this week, Grok 4.9 would be ‘Astra/Fable class,’ a
🧩 Pattern This Issue
- Korea: Samsung leads 320B won co-investment in Dutch startup Euclid
- Taiwan: MediaTek reveals 2nm Dimensity 9600 Pro agentic AI chip
- Korea/Taiwan: ASML expects accelerated EUV orders driven by AI demand
East Asian hardware giants are locking down advanced 2nm execution paths and specialized chip designs, which shifts the bottleneck of the AI race from training-set sizes to high-yield silicon manufacturing.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 Enterprise & Cloud
Google Detects Shift in Attacker Behavior: From ‘Stealing AI’ to ‘Operating AI’ for Attacks
Google’s latest threat intelligence report indicates a significant evolution in cyberattack methods, with malicious actors now using multiple AI agents to autonomously execute complex attack stages, rather than just assisting human attackers. These AI systems can automate tasks from vulnerability scanning to credential harvesting, drastically accelerating operations and reducing detection windows. One observed incident involved an AI-driven campaign collecting thousands of credentials within six hours of breaching a cloud environment.
Why it matters: The shift from human-assisted AI to autonomous AI agents for attacks fundamentally changes the calculus for cyber defense. Enterprises need to recognize that their defensive AI and human response times will be pitted against autonomous attack systems, making traditional detection-and-response frameworks less effective.
For Western readers: Western companies relying on cloud environments and extensive OSS for AI development must reassess their supply chain security and incident response capabilities, assuming AI-driven attacks will occur faster and with less human oversight than previously anticipated.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 Semiconductors & Hardware
Samsung Electronics Co-Invests 320 Billion Won in Dutch AI Chip Startup Euclid
Samsung Electronics is jointly investing 320 billion Korean won (approximately $235 million USD) in Euclid, a Dutch AI chip startup specializing in next-generation AI processors. This investment aims to strengthen Samsung’s position in the AI semiconductor market and diversify its portfolio beyond memory chips. The move follows Samsung’s previous acquisition of AI chip design companies and indicates a strategic shift towards integrated AI solutions.
Why it matters: Samsung’s investment in Euclid is not just about funding a startup; it’s a direct move to secure intellectual property and accelerate its AI chip development, reducing reliance on external AI accelerators. This reflects a broader trend among major tech players to vertically integrate AI capabilities, ensuring they control more of the core technology stack rather than just manufacturing components.
For Western readers: Western AI chip developers should anticipate increased competition from integrated East Asian players like Samsung, which are now directly investing in core AI processing technology rather than solely focusing on memory solutions. If you compete in the AI chip space, be aware Samsung is building out its own capabilities and not just relying on foundry services.
🇰🇷 AI & Machine Learning
SKT Hosts GSMA RCS Meeting on AI Agent Pre-Payment SMS Verification
SK Telecom hosted a meeting of the GSMA’s Rich Communication Services (RCS) Business Group in Seoul, focusing on integrating AI agents with RCS for secure transactions. A key discussion point was a new feature enabling AI agents to request user confirmation via SMS before making payments, aiming to enhance security and user trust in AI-driven services.
Why it matters: SK Telecom is pushing for telco-led AI agent services that prioritize security through explicit user consent via SMS. This approach aims to differentiate carrier offerings from potentially less secure, third-party AI payment integrations, focusing on building user trust in a space where data security concerns are paramount.
For Western readers: Western telecom operators and AI service providers should consider SMS-based payment verification as a standard for AI agent transactions, particularly in regulated markets, as it establishes a clear path for accountability and user control.
🇰🇷 AI & Machine Learning
Konyang University Hospital to Build Dedicated AI Data Center, Investing Up to 20 Billion Won
Konyang Education Foundation, operating Konyang University and Hospital, plans to invest up to 20 billion Won (approximately $15 million USD) to establish an AI-specific data center. This initiative, deploying 20-25 NVIDIA B300 GPUs over three years, aims to accelerate advanced precision medicine research and integrate AI into medical education and clinical applications. It marks only the second instance in South Korea of a university and hospital jointly securing their own AI data center.
Why it matters: The investment by Konyang University Hospital into a dedicated AI data center, specifically leveraging NVIDIA B300 GPUs, demonstrates a tangible commitment to integrating AI into healthcare beyond pilot programs. This is not just about research; it’s about building the infrastructure to move AI models from academic studies into actual clinical validation and patient care, addressing the gap between theoretical models and practical deployment.
For Western readers: Western healthcare providers and AI developers should recognize that South Korea is building direct, dedicated AI infrastructure within hospitals, which can accelerate real-world clinical application and validation faster than purely academic or cloud-based initiatives. This approach emphasizes data sovereignty and direct control over patient data, potentially setting a precedent for similar models in other regulated markets.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 AI & Machine Learning
Apple Releases Six Major OS Updates, Introducing Apple Intelligence and Siri AI
Apple has officially rolled out updates for iOS 27, iPadOS 27, macOS 27, watchOS 27, visionOS 27, and tvOS 27. These updates bring performance enhancements, interface improvements, and significant security patches across all platforms. Crucially, the updates introduce the new Apple Intelligence system, which integrates generative AI capabilities into Siri and core applications like Photos, Safari, Mail, and Calendar, powered by third-generation Apple Foundation Models developed in collaboration with Google’s Gemini technology.
Why it matters: Apple’s deep integration of generative AI into its operating systems and applications means that users will experience these capabilities as fundamental parts of the Apple ecosystem, rather than optional add-ons. The collaboration with Google on the underlying Foundation Models points to the immense complexity and resource requirements of developing competitive LLMs, even for a company of Apple’s scale, and it changes the competitive dynamic for other AI providers who were not part of that deal.
For Western readers: Western businesses in the AI and app development space should assume a higher baseline for AI functionality in Apple devices, and developers should prepare to leverage Apple Intelligence APIs to create seamless, AI-enhanced experiences within their own apps, or risk being outflanked by Apple’s built-in capabilities.
🇹🇼 Semiconductors & Hardware
ASML’s Dominance Extends to 2030s Amid Strong EUV Orders and Accelerated High NA Adoption
ASML’s CFO Roger Dassen stated that the AI boom has altered client sentiment, with nearly all 2027 EUV capacity booked and substantial 2028 orders, indicating robust demand for advanced manufacturing. Customers like TSMC, Samsung, and SK hynix are accelerating their adoption of next-generation High NA equipment, with Intel already processing over one million wafers using the technology.
Why it matters: The rapid uptake of High NA EUV by key players like Samsung for memory chips by 2028, earlier than market expectations, suggests that the process shrinks needed for next-gen AI accelerators are proceeding faster than many had predicted. Memory makers adopting advanced lithography for smaller die sizes means more efficient memory for AI, and the article notes this adoption for memory chips is easier, suggesting a faster path to volume production for this critical component of the AI supply chain.
For Western readers: Western chip designers and cloud providers should expect the density and efficiency of advanced memory and logic chips to improve on an accelerated timeline, impacting their hardware roadmaps and cost structures.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 AI & Machine Learning
Seven PhD Students Train a 7B LLM from Scratch in Three Months, Releasing All Code, Data, and Training Logs
📊 Featured Chart
Average score from 9 core participants
Seven PhD students from Beijing Zhongguancun Institute successfully trained a 7-billion parameter Large Language Model (ZGCM-1) from scratch in just three months, using hundreds of AI agents to manage data processing, experiments, and evaluation. They open-sourced all training data, model weights, code, checkpoints, and logs, highlighting an ‘AI for AI‘ (AI4AI) development paradigm.
Why it matters: The ‘AI for AI’ approach used by this small team points to a future where foundational model development costs, currently dominated by large-scale human engineering teams, could be dramatically reduced. This changes the economics of who can develop competitive models, favoring agility and efficient tooling over sheer headcount and capital.
For Western readers: Western researchers and startups aiming to compete in the LLM space should prioritize the development and adoption of AI-agent-driven workflows to reduce the resource gap with well-funded incumbents, or risk being outpaced by more agile East Asian teams using these methods.
AI & Machine Learning
DeepSeek Open-Sources Harness Agent Runtime With Everything-Is-a-Plugin Design
Chinese AI company DeepSeek has open-sourced its Harness Agent Runtime framework, designed to facilitate the development of AI agents capable of performing complex tasks through an “everything-is-a-plugin” architecture. The framework offers features like concurrent tool execution, a comprehensive tool registry, and support for multiple LLM interfaces, aiming to reduce development costs and improve efficiency.
Why it matters: DeepSeek’s open-sourcing of Harness Agent Runtime accelerates the shift towards more autonomous AI systems by making sophisticated agent design accessible. This initiative, rather than focusing on the LLM itself, pushes the frontier on how these models interact with external tools and execute multi-step tasks, which is a critical area for enterprise AI adoption.
For Western readers: Western businesses building AI applications that require agents to interact with diverse tools should evaluate DeepSeek’s Harness framework for its architectural approach and potential efficiency gains, particularly if their development strategy includes integrating open-source components from varied global sources.
🇨🇳 Other
Siemens Xcelerator Stargazer Ecosystem Conference in Shenzhen to Discuss Industrial AI and Ecosystem Co-creation
Siemens will host its 2026 Xcelerator Stargazer Ecosystem Conference on September 21st in Shenzhen, focusing on the future of industrial AI and collaborative ecosystem development. The conference aims to address challenges in deploying industrial AI, from adapting models to production lines to scaling successful projects, by fostering cooperation among its over 600 ecosystem partners.
Why it matters: Siemens’ Xcelerator platform and its “Stargazer” ecosystem represent a strategic effort to embed its industrial AI solutions deeply into the Chinese manufacturing sector. By focusing on ecosystem collaboration, Siemens aims to overcome the common challenges of industrial AI deployment in China, such as lack of fit for local production lines and scaling issues, effectively localizing its offerings and strengthening its market position.
For Western readers: Western industrial software and automation companies should observe how Siemens leverages its partner ecosystem in China to convert technical advantages into commercial opportunities, as this model could inform their own strategies for market entry and scaling in complex, state-influenced markets.
🔺 The Prism
Where US and East Asian technology interests intersect
Semiconductors & Hardware
BOOST: Concurrent Access to Host Memory and HBM to Accelerate LLM Inference
Researchers have developed BOOST, a new runtime system that significantly improves Large Language Model (LLM) inference throughput by enabling concurrent access to both high-bandwidth memory (HBM) and host memory. This method leverages previously underutilized host memory bandwidth, bypassing the limitations of traditional hierarchical memory access and prefetching strategies.
Why it matters: The GPU memory system is a key choke point for AI. By optimizing how LLMs access memory, BOOST can extract more performance from current hardware without requiring new silicon. This directly benefits the East Asian firms supplying HBM and designing accelerator chips, as it extends the useful life and performance ceiling of their existing products and pipelines, potentially delaying the need for radical architectural shifts or expensive HBM overhauls.
For Western readers: Western AI accelerator designers and cloud providers should evaluate integration of runtime systems like BOOST to maximize throughput from existing hardware, particularly from HBM supplied by Korean manufacturers. This could mean deferring some hardware upgrade cycles while still improving per-chip inference performance.
AI & Machine Learning
STRIDE: Accelerating Multi-Turn Agentic On-Policy Distillation for LLMs
📊 Featured Chart
Source: arXiv:2609.14636
Researchers have developed STRIDE, a method that significantly accelerates on-policy distillation (OPD) for large language models (LLMs) in multi-turn agentic settings. This technique optimizes the process of transferring capabilities from large ‘teacher’ models to smaller ‘student’ models by adaptively stopping rollouts and restarting generation at critical points, achieving speedups up to 5.10x while maintaining or exceeding performance.
Why it matters: This innovation from East Asian researchers directly addresses a key bottleneck for companies building and deploying AI agents: the immense computational cost of training. Faster distillation allows companies to develop and iterate on agentic AI applications more quickly, which is critical for staying competitive in areas like customer service, financial analysis, and industrial automation where multi-turn interactions are common.
For Western readers: Western AI companies should be aware that their East Asian counterparts are actively pursuing and publishing advancements in LLM efficiency, which could lead to a faster deployment cycle for agentic AI applications in these markets. Focus on integrating similar distillation acceleration techniques into your own development pipelines to match this pace.
AI & Machine Learning
TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models
Researchers have introduced TimeThink, a synthetic data generation and training framework that significantly improves how timeseries multimodal large language models (TS-MLLMs) handle compositional reasoning, particularly for out-of-distribution temporal patterns. By creating atomic and composite question-answer pairs with objective ground truth and employing a reinforcement learning with verifiable rewards (RLVR) strategy, TimeThink enables models to learn underlying logic rather than just imitate reasoning traces.
Why it matters: The core challenge TimeThink addresses—the failure of existing TS-MLLMs to capture dynamic temporal patterns and provide explicit, verifiable reasoning—is a major hurdle for deploying AI in critical East Asian sectors like healthcare and industrial analytics. Japan, for example, is pushing for AI in diagnostics and elder care, where implicit reasoning is a non-starter. This framework provides a path towards more reliable and explainable AI in these high-stakes applications.
For Western readers: Western businesses developing AI solutions for time-series data, especially in healthcare or finance, should evaluate the TimeThink framework for improving model robustness and interpretability, as its adoption in East Asia could set a new standard for AI reliability.