East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- Malaysia’s decision to integrate Huawei’s Ascend chips into its national digital infrastructure demonstrates how Southeast Asian nations are actively bypassing US sanctions to secure cheaper, readily available Chinese hardware.
- Japan’s 50.8% year-on-year surge in semiconductor sales for July 2026 establishes the country as the primary beneficiary of the global hardware supply chain shift away from Taiwan.
- Siemens Digital Industries Software’s integration of its Fuse EDA AI Agent platform into TSMC’s advanced packaging workflow moves chip design verification from human-in-the-loop engineering to autonomous, overnight execution.
This Issue’s Analysis
The Signal
Siemens’ Fuse EDA AI Agent: Japan’s Autonomous Verification Bet to Save the Rapidus 2nm Timeline
Siemens Digital Industries Software announced an expansion of its ‘Fuse EDA AI Agent’ platform, integrating NVIDIA AI technology to enable autonomous and long-running verification
Semiconductors & Hardware
Japan’s Chip Sales Surge 50.8% as Global Semiconductor Demand Breaks Annual Records in Seven Months
The Semiconductor Industry Association (SIA) announced that global semiconductor sales for July 2026 reached $146.8 billion, a 135.1% increase year-over-year. The total sales from
Semiconductors & Hardware
Winbond’s Chip Testing Outlook: Why Capacity Bottlenecks—Not a Bubble—Limit AI Supply
During SEMICON Taiwan 2026, Winbond VP Chen Shao-kun stated that the AI demand is not a bubble but rather an issue of insufficient supply chain capacity. He highlighted continued c
Semiconductors & Hardware
Huawei and Xiaomi Dual Launches: China’s Next-Gen Foldables and EV Tech Hit the Market Together
Huawei held its autumn product launch, unveiling HarmonyOS 7, the triple-fold Mate XT 2, and other devices. Concurrently, Xiaomi launched its Pengcheng range-extended SUV, the 18 F
🧩 Pattern This Issue
- Japan: Siemens deploys Fuse EDA AI agents to automate round-the-clock chip verification
- Korea: Synopsys integrates Ansys software to unify design across complex HBM systems
- Taiwan: Winbond points to severe physical supply chain shortages over AI bubbles
The bottleneck in AI scaling is shifting from model architectures to the physical complexity of hardware manufacturing, forcing the EDA and memory sectors to deploy autonomous AI agents and integrated design suites just to keep up with next-generation silicon production demands.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 AI & Machine Learning
OpenAI Announces GPT-6 Astra: How Much Has Codex’s Coding Evolved From Previous Generations?
📊 Featured Chart
GPT-5.6 Sol values: 65.7, 78.5, 37.3, 72.7, 55.9
OpenAI has announced GPT-6 Astra, its next-generation large language model, demonstrating significant improvements in autonomous PC operation, coding, and cybersecurity capabilities. The model achieved a 99.9% score on the ARC-AGI-3 abstract reasoning benchmark and a 100% score on ExploitBench, which assesses the ability to create exploit code from known vulnerabilities. GPT-6 Astra will be rolled out to various ChatGPT plans, OpenAI API, Microsoft Azure, and Amazon Bedrock in the coming days.
Why it matters: The jump in GPT-6 Astra’s autonomous PC operation (OSWorld 2.0) and coding proficiency (Terminal-Bench 4.0, DeepSWE v1.1) means that OpenAI is moving beyond purely text-based interfaces to models that can genuinely interact with existing enterprise software environments. The 100% score on ExploitBench, a benchmark for generating attack code, is particularly striking and will shift the cybersecurity risk landscape, forcing enterprises to re-evaluate their defense strategies against AI-powered attacks.
For Western readers: Western businesses should assume that AI-powered autonomous agents capable of interacting with standard enterprise software and developing exploit code will be widely available soon. This necessitates an immediate review of internal automation strategies, cybersecurity defenses against AI-generated attacks, and the implications for software development workflows where AI can now contribute more deeply.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 Policy & Regulation
Malaysia Proceeds with Huawei AI Chip Adoption Despite US Warnings
Malaysia is moving forward with plans to integrate Huawei’s AI chips into its national digital infrastructure, disregarding explicit warnings from the United States regarding security concerns. This decision positions Malaysia as a significant partner for Huawei in Southeast Asia, particularly in the realm of AI and data center development.
Why it matters: Malaysia’s decision provides a significant win for Huawei, demonstrating its continued ability to secure major infrastructure contracts outside of direct US influence, particularly in a region critical for global supply chains. For China, it reinforces a narrative of technological self-reliance and expanding geopolitical reach, pushing back against US containment strategies.
For Western readers: If you are a Western government or company involved in technology diplomacy, recognize that US warnings alone are insufficient to deter nations with strong economic ties to China from adopting Chinese technology, especially when it aligns with their national development goals. Western chip and AI infrastructure providers should expect increased competition from Huawei in Southeast Asia and other emerging markets.
🇰🇷 AI & Machine Learning
AimIntelligence: AGI-level ‘Astra’ Still Has Security Gaps; Next Focus is Robotics, Autonomous Driving
AimIntelligence, a Korean AI security company, argues that even advanced AGI models like their ‘Astra’ platform still possess significant security vulnerabilities. The company plans to expand its focus beyond large language models (LLMs) to address security challenges in rapidly developing areas such as robotics and autonomous driving.
Why it matters: The security of AI systems is not a peripheral concern; it is fundamental to their deployment, particularly when AI moves into safety-critical applications like autonomous vehicles. The fact that a specialized AI security firm like AimIntelligence identifies persistent vulnerabilities even in AGI-level models indicates that current security paradigms are insufficient and that the industry needs to mature its approach rapidly.
For Western readers: Western businesses developing or deploying AI in critical infrastructure, robotics, or autonomous systems should assume current AI models, regardless of their perceived intelligence level, have significant security attack surfaces, and integrate specialized AI security measures from the earliest design phases.
🇰🇷 Semiconductors & Hardware
Synopsys Integrates Ansys to Target HBM Market, Unifying Design from Chip to System
Synopsys held a press conference in Seoul to announce its strategy of targeting the high-bandwidth memory (HBM) and AI semiconductor markets by integrating Ansys’s simulation software with its own EDA and IP solutions. This new ‘Multiphysics Fusion‘ solution aims to provide comprehensive engineering tools that cover semiconductor design to full system simulation, enabling physical-based decision-making throughout the product development lifecycle.
Why it matters: Korean memory manufacturers are leading the charge in HBM development, and this announcement from Synopsys, delivered directly in Seoul, shows a clear strategic focus on servicing that market. The integration with Ansys is not merely about incremental feature additions; it targets the core engineering challenge of HBM and AI chips, which require deep, system-level understanding of physical phenomena to achieve performance gains and faster release cycles.
For Western readers: Western semiconductor design firms and their fabless partners should evaluate how this integrated Synopsys-Ansys offering addresses the increasing complexity of HBM and AI chip co-design, particularly as it relates to system-level performance and thermal management, which are major bottlenecks for next-generation hardware.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 Semiconductors & Hardware
ASML Breaks Ground on Second Large Industrial Park in Eindhoven to Meet Expansion Needs, Expected Completion 2029
ASML held a groundbreaking ceremony for its second large industrial park, BIC North, in Eindhoven, Netherlands, on September 8, 2026. This multi-stage development, spanning 350,000 square meters, will create up to 20,000 jobs and significantly expand production capacity for ASML’s long-term growth. The first phase, featuring a new “flow factory” manufacturing model for TWINSCAN lithography systems, is expected to open in 2029 with at least 3,000 employees.
Why it matters: ASML’s announcement of a new industrial park, with a novel ‘flow factory’ model, suggests a deliberate move to de-risk and accelerate its production capacity. While Western reporting might focus on the geopolitical implications of ASML’s centrality, Taiwanese coverage emphasizes the practical impact on the supply chain, seeing this as a direct response to the persistent global demand for advanced chips that power AI and high-performance computing, rather than just a strategic maneuver.
For Western readers: Western chipmakers and AI developers should adjust their supply chain forecasts to reflect ASML’s planned capacity increase, potentially easing the bottleneck for advanced lithography tools by 2029 and enabling faster scaling of cutting-edge chip production.
🇹🇼 Enterprise & Cloud
PostgreSQL Logical Decoding Vulnerability Hidden for 12 Years, Replication Accounts Can Load Arbitrary Code
Multiple high-severity cybersecurity vulnerabilities have been disclosed and exploited across various platforms and services, including a 12-year-old PostgreSQL vulnerability that allows replication accounts to load arbitrary code. Blockstream’s Liquid Network was attacked, leading to a loss of approximately 4,000 Bitcoin, though most funds were returned. JetBrains’ Cadence cloud service and Coder’s module Registry were also compromised due to exploitation of known vulnerabilities.
Why it matters: The continuous discovery of critical vulnerabilities, some dormant for over a decade, in foundational software like PostgreSQL or widely adopted development tools such as JetBrains TeamCity, directly affects the stability and trustworthiness of the entire software supply chain. For East Asian companies, especially those in manufacturing and advanced technology, reliance on these tools means that such exploits introduce significant operational risks and potential for IP theft or disruption.
For Western readers: Western businesses using any of the affected platforms — PostgreSQL, JetBrains TeamCity, Coder, N-able N-central, Citrix NetScaler, or PaperCut — should immediately verify patching status, review logs for indicators of compromise, and rotate credentials to mitigate active threats and prevent potential supply chain attacks originating from these vulnerabilities.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 AI & Machine Learning
Wang Yunhe’s Startup, Jiyuan Lyudong, Launches Its First Agent-Native Model NeoHorse
Jiyuan Lyudong (基元律动), a startup founded by Wang Yunhe, former director of Huawei’s Noah’s Ark Lab and head of the Pangu large model, has released its first Agent-Native model, NeoHorse, in 4B and 9B versions. The model, supported by compute from Wuwenzhixin (无问芯穹) and research from Tsinghua and Peking Universities, focuses on agent capabilities like tool calling, error detection, and path adjustment. The 4B model, after ‘Agentic Post-Training,’ performs comparably to or slightly better than a 9B foundational model across various benchmarks.
Why it matters: Jiyuan Lyudong’s move to train its own models, rather than just orchestrate others, is a telling sign of the maturation of China’s AI ecosystem. They are transforming operational data from multi-agent systems into training data, effectively creating models ‘native’ to agentic workflows. This suggests a strategic effort to capture more value up the AI stack, embedding execution intelligence directly into the model itself.
For Western readers: Western AI developers and platform providers should recognize that Chinese firms are not just fine-tuning existing models but actively developing specialized, agent-native architectures based on real-world operational data. Expect increasing competition from China in specialized AI agents that are highly optimized for specific task execution rather than just general language fluency.
🇨🇳 Startups & Funding
DeepCtrls Secures Major Investment from CATL, Aramco Ventures, and Others to Accelerate Physical AI Infrastructure Development
Chinese ‘physical AI‘ firm DeepCtrls (深度智控) has completed a new multi-hundred-million yuan B+ funding round, led by CATL, with participation from Aramco Ventures, Taiping Innovation Investments, GF Xinde, and Fosun RZ Capital. This round follows two other recent funding rounds, signaling strong investor belief in AI’s expansion from the digital to the physical world, particularly in energy and computing infrastructure.
Why it matters: The concept of ‘physical AI’ — AI that understands and controls real-world physics — is where the practical value for industrial systems lies, especially for energy efficiency and complex manufacturing. This funding for DeepCtrls, particularly from a major energy player like CATL, indicates a pivot from pure model development towards tangible industrial applications that directly impact operational costs and resource utilization. What matters is that actual control systems are being developed, not just better chatbots.
For Western readers: Western firms in industrial AI, energy management, and robotics should recognize that Chinese players are moving aggressively to integrate AI into physical infrastructure, creating systems that autonomously optimize real-world operations. This shifts the competitive landscape from model performance benchmarks to real-world deployment and control efficacy, so watch for how quickly these ‘physical AI’ systems achieve scale and measurable efficiencies in major industrial deployments.
🔺 The Prism
Where US and East Asian technology interests intersect
Policy & Regulation
China’s Huawei Heads to Trial in US Over Sanctions and Trade Secrets
Chinese telecommunications giant Huawei is facing a trial in a New York court over allegations of evading U.S. sanctions and stealing intellectual property, more than eight years after initial charges. This legal battle represents a critical ongoing point of tension in the broader US-China technology and trade relationship, directly impacting a key Chinese national technology champion.
Why it matters: This trial, even if delayed, puts the long-running US campaign against Huawei back into the public eye, reinforcing the narrative of Chinese firms engaging in unfair practices. While Huawei has largely adapted its supply chain to reduce reliance on US components, a guilty verdict could still trigger further restrictions or penalties, particularly on its remaining international business.
For Western readers: Western companies should understand that the legal battle against Huawei is a persistent facet of US-China tech competition, and even an old case can rekindle pressure on firms with any remaining ties to Huawei or similar Chinese entities.
Startups & Funding
SoftBank Group Issues 1 Trillion Yen Retail Bond with 4.75% Coupon
SoftBank Group has set a 4.75% coupon rate for its 1 trillion yen ($6.4 billion) retail bond offering, marking the highest rate for the investment group’s straight bonds in 17 years. This massive offering is tied with NTT Finance’s issuance as the largest by a Japanese company.
Why it matters: This bond issuance represents a significant fundraising maneuver by SoftBank, securing substantial capital from domestic retail investors in Japan. While it fuels SoftBank’s ambitious AI investment agenda, it also signals the increasing reliance of large Japanese conglomerates on local capital markets to fund global technology plays, especially as borrowing costs rise.
For Western readers: Western investors should recognize that SoftBank’s ability to raise 1 trillion yen domestically indicates a deep pool of Japanese retail capital available for its investment vehicles, which could provide a buffer against fluctuations in international markets for future fundraising efforts.
AI & Machine Learning
AI Oversight: Answer Access Improves Conclusion Checking, Not Reasoning Verification
📊 Featured Chart
Source: arXiv:2609.00264, n=8 monitors
Research published on arXiv challenges common AI oversight methods, showing that providing human or LLM monitors with a trusted reference answer primarily helps them check conclusion consistency, rather than genuinely verifying the underlying reasoning chain. The study used 237 solutions to physics questions from frontier models, finding that access to the correct answer significantly increased the flagging of wrong-answer traces but decreased detection of ‘critical traces’ where the answer was correct despite flawed reasoning.
Why it matters: The paper highlights a critical blind spot in how AI models are often evaluated, especially for reasoning tasks. If human evaluators or monitoring LLMs are given the ‘right answer,’ they default to checking if the model reached that answer, rather than scrutinizing the process. This creates a risk that East Asian AI companies, eager to demonstrate high accuracy, might inadvertently mask underlying logical flaws in their models by using evaluation methods that prioritize outcome over process.
For Western readers: Western businesses building or deploying AI systems should re-evaluate their monitoring and auditing frameworks, recognizing that a correct output does not guarantee sound reasoning. When assessing East Asian AI partners or products, prioritize transparency in their evaluation methodologies and push for methods that independently verify reasoning steps rather than just final answers, especially in high-stakes applications.