East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- The US Commerce Department’s move to block Chinese firms from using cloud-based remote access to bypass export controls will force Chinese AI developers to rely more heavily on domestic hardware, accelerating the growth of local alternatives like Huawei’s Ascend chips.
- Kioxia and Western Digital’s ¥5 trillion expansion of their joint-venture flash memory plants in Japan secures a massive domestic supply of NAND flash, insulating global AI storage supply chains from geopolitical disruptions in the Taiwan Strait.
- Changxin Memory Technology’s leap to a 1.8 trillion yuan net profit in the first half of 2026 demonstrates that China’s state-backed memory sector is rapidly scaling production to capture the surging domestic demand for AI hardware.
This Issue’s Analysis
The Signal
U.S. Cloud Export Controls: Closing the Remote AI Chip Loophole for Chinese Tech
The US Department of Commerce is planning new regulations to prevent Chinese companies from circumventing AI chip export controls by using cloud services or remote access to advanc
Semiconductors & Hardware
Kioxia and Western Digital Plan a ¥5T NAND Expansion to Meet AI Storage Demand
Joint venture partners Kioxia and Western Digital’s SanDisk have announced a massive ¥5 trillion ($31 billion) investment through 2032 to boost advanced NAND flash memory productio
Semiconductors & Hardware
CXMT Reaches 77.6B Yuan H1 Profit as AI Demand Rebounds China’s DRAM Leader
Changxin Memory Technology (CXMT), a leading Chinese semiconductor memory producer, announced a net profit of 77.6 billion yuan ($1.8 trillion) for the first half of 2026, a signif
Semiconductors & Hardware
Nvidia’s Revolving Finance Risk: Why $105B in Partner Guarantees Outpaces Quarterly Revenue
Nvidia reported strong Q2 FY2027 earnings, exceeding market expectations with $96.2 billion in revenue and a 106% year-over-year increase, driven by a 117% surge in data center rev
🧩 Pattern This Issue
- Policy: US closes remote cloud access loophole for Chinese AI developers
- China: CXMT books record profits on domestic AI-driven memory demand
- Taiwan/Korea: Nvidia faces Taiwan power grid constraints and rising memory costs
As Washington squeezes China’s access to compute via cloud loopholes, Beijing is successfully pivoting to domestic hardware self-sufficiency, while Nvidia’s supply chain faces physical limits from Taiwan’s energy grid and rising Korean memory costs.
Also This Issue
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 AI & Machine Learning
Why Nvidia Might Acquire Hugging Face: From Open Model Download to Distribution
AI Times Korea discusses speculation that Nvidia could acquire Hugging Face, suggesting such a move would shift Hugging Face’s role from a platform for downloading open-source AI models to a robust distribution channel. This acquisition would deepen Nvidia’s integration into the AI software ecosystem, leveraging Hugging Face’s developer community and model repository.
Why it matters: Nvidia’s dominance in AI hardware is established, but control over software and model distribution is the next frontier. Acquiring Hugging Face would give Nvidia direct influence over the open-source AI community and the flow of models, providing a strategic defense against competitors developing their own full-stack AI solutions or alternative hardware architectures. It also means Nvidia would control a key layer of the AI stack currently seen as neutral.
For Western readers: Western AI developers and companies relying on Hugging Face for open-source models should recognize that Nvidia’s potential ownership could shift the platform’s neutrality and introduce new commercial incentives or integration points with Nvidia’s ecosystem, potentially altering how models are accessed and deployed.
🇰🇷 Semiconductors & Hardware
Surge in Memory Prices Impacts NVIDIA’s Performance Outlook
NVIDIA is reportedly lowering its revenue projections for its fiscal third quarter due to a sharp increase in HBM (High Bandwidth Memory) prices. This situation is compelling NVIDIA to adjust its product pricing and potentially delay some shipments to manage costs.
Why it matters: NVIDIA’s decision to revise its outlook due to HBM costs directly contradicts the prevailing narrative that AI chip demand is impervious to pricing pressures. This indicates that even a market leader like NVIDIA faces constraints from its supply chain, affecting both its margins and delivery schedules.
For Western readers: If you are an investor in NVIDIA or an enterprise client planning large AI infrastructure deployments, assume a tighter supply and higher pricing environment for HBM-intensive chips will persist, potentially delaying rollouts or increasing CapEx. Focus on the actual availability of HBM capacity, not just headline demand.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 AI & Machine Learning
Google Launches Gemini 3.5 Transcribe Model for Speech-to-Text, Speaker Diarization, and Oral Fluency Processing
Google has introduced Gemini 3.5 Transcribe, a speech-to-text model available in public preview via the Gemini API and Gemini Enterprise Agent Platform. It features advanced capabilities like automatically removing filler words, handling speaker corrections, and organizing punctuation, numbers, and paragraphs into readable text.
Why it matters: Google’s move with Gemini 3.5 Transcribe signals a continued investment in improving AI’s utility for enterprise applications, particularly in Asia where accurate transcription of complex spoken languages and dialects can unlock significant business value. The focus on cleaning up oral speech and distinguishing multiple speakers addresses key pain points for companies dealing with large volumes of audio data, from customer service calls to international business meetings.
For Western readers: Western businesses operating in East Asia, especially those with multilingual teams or customer bases, should evaluate Gemini 3.5 Transcribe’s capabilities for Mandarin Chinese (Simplified) and Cantonese (Traditional) to enhance communication and data analysis, assuming Google expands its language support for specific regional variants like Taiwanese Mandarin.
🇹🇼 Policy & Regulation
NVIDIA Wants Power, Not Substations? Northern Taiwan Grid Nears Overload as Public Misunderstands Energy Mix
📊 Featured Chart
Source: TechNews, based on current electricity mix
NVIDIA’s new ‘Star Cluster’ R&D and manufacturing hub in Beitou-Shilin Technology Park faces local opposition to a critical substation project, highlighting Taiwan’s intensifying electricity grid challenges. A recent survey reveals that while the public supports AI and semiconductor industry expansion, there’s a significant lack of understanding regarding power generation sources and the need for new infrastructure.
Why it matters: Taiwan’s energy supply is critical to the global semiconductor supply chain, and local resistance to infrastructure directly threatens the expansion plans of companies like NVIDIA and TSMC. The public’s misinformed opposition to substations could create a severe bottleneck, preventing new AI and semiconductor facilities from securing the necessary power to operate, directly impacting global tech production capacity.
For Western readers: Western companies relying on Taiwanese manufacturing should anticipate potential delays and increased costs due to energy infrastructure bottlenecks; this issue could tighten capacity for CoWoS packaging and other power-intensive processes. Watch for how the Taiwan government resolves local opposition to grid projects, as it will signal the viability of future high-tech investments there.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
AI & Machine Learning
Tencent Open-Sources Hy4 LLM with 770B Parameters and 1M-Token Context
📊 Featured Chart
Source: Tencent internal blind evaluation
Tencent Hunyuan has open-sourced a preview of its Hy4 large language model, featuring 770 billion total parameters and a 1 million token context window. The model, designed for productivity tasks, is accessible through Tencent’s various applications and via API, with a limited free trial period.
Why it matters: Tencent’s decision to open-source a model of this scale, even as a preview, indicates a shift towards a more collaborative, platform-driven strategy to capture developer mindshare in China. The company’s internal evaluation claims of efficiency gains in its own systems are notable, suggesting this isn’t just a public-facing play but a sign of genuine integration and optimization within Tencent’s tech stack.
For Western readers: Western businesses operating in China, or those with Chinese partners, should expect to see Tencent’s Hy4 ecosystem gain traction in enterprise applications, potentially becoming a de facto standard for certain productivity and development tasks within the Chinese market. It will be important to understand how its API and enterprise offerings integrate into local supply chains.
🇨🇳 AI & Machine Learning
Coding No Longer Exclusive to Programmers! Alibaba’s Qoder Shows Impressive Advances
Alibaba’s Qoder, an AI agent for coding, has released a new desktop version featuring a ‘desktop pet’ interface that allows users to generate code and manage projects through natural language conversations. The updated Qoder can understand vague verbal requirements, plan development tasks, write code, run tests, and self-correct, operating beyond traditional Integrated Development Environments (IDEs). It has already served 6 million users and 100,000 enterprise clients globally in the past year.
Why it matters: Alibaba is pushing the boundaries of AI-driven code generation beyond the developer community, moving it to the desktop where non-programmers can interact with it conversationally. This reflects a strategic pivot from merely assisting developers to enabling entirely new classes of users to build digital tools, positioning Qoder not just as a coding agent but as a ‘digital execution’ layer for various tasks.
For Western readers: Western businesses should recognize that Chinese AI players like Alibaba are not just competing on model benchmarks but on user accessibility and integration across diverse workflows; assume this will lead to a broader, faster diffusion of AI-driven tool creation within Chinese enterprises compared to Western counterparts still focused on developer-centric tools.
AI & Machine Learning
Ant Group Launches Finance-Tuned Ling-3.0-flash-Fin AI Model, Plans Open-Source Release
Ant Group has introduced Ling-3.0-flash-Fin, a large language model specifically optimized for financial tasks, leveraging the existing Ling-3.0-flash architecture. The model, with 124 billion total parameters and 5.1 billion active parameters, is designed for financial document analysis, investment analysis, and banking applications. Ant Group plans to open-source the model weights next week, offering a one-month free API period for professionals.
Why it matters: Ant Group’s move to open-source a finance-tuned model signals a strategy to accelerate adoption and foster an ecosystem around its AI capabilities within China’s financial industry. This could enable smaller financial institutions to integrate advanced AI without heavy upfront development costs, potentially reinforcing Ant Group’s platform dominance in adjacent financial services through its model’s ubiquity.
For Western readers: Western financial technology firms and data providers should recognize this as a direct challenge to their market share in China, as a potent, open-source domestic alternative will likely be favored by local regulators and customers.
🔺 The Prism
Where US and East Asian technology interests intersect
Semiconductors & Hardware
Kioxia Plans $31.4 Billion NAND Investment, SK hynix Considers Deeper Ties
Kioxia, with Western Digital’s SanDisk, plans to invest $31.4 billion over six years into new NAND flash capacity, with $11.3 billion allocated to a new fab in Kitakami, Japan, seeking significant government subsidies. Concurrently, SK hynix is evaluating closer collaboration with Kioxia, hinting at potential strategic moves beyond its current indirect stake through a Bain Capital investment vehicle.
Why it matters: Kioxia’s investment, backed by Japanese government subsidies, aims to secure Japan’s position in advanced NAND manufacturing, leveraging the familiar ‘national champion’ consortium model to shore up domestic tech supply chains. SK hynix’s explicit interest in ‘co-developing’ the NAND market with Kioxia signals a potential strategic alignment that could shift the competitive balance in the global memory sector, moving beyond its existing indirect financial holding.
For Western readers: Western companies relying on NAND supply should track Kioxia’s Kitakami fab buildout and the progress of Japanese government subsidies, as this capacity could stabilize or increase future supply. They should also watch for any concrete announcements from SK hynix regarding its relationship with Kioxia, as a deeper alliance could impact pricing power and technology roadmaps in the NAND market.
Cross-Regional Analysis
SoftBank’s US Gas Power Plant Faces Mixed Local Reaction Amid AI Data Center Buildout
SoftBank Group‘s plans for the largest natural gas-fired power plant in the U.S., located in Waverly, Ohio, have received a mixed reception from local residents at a public meeting. This plant is intended to power a major AI data center, highlighting SoftBank’s strategy to secure energy infrastructure for its growing AI investments.
Why it matters: This initiative underscores how crucial energy supply has become for AI infrastructure, driving Japanese firms like SoftBank to directly invest in power generation in key overseas markets. It shows that the buildout of AI capacity is less about pure compute and more about the entire integrated supply chain, from power to cooling.
For Western readers: Western businesses involved in AI infrastructure development, or those planning large-scale data center operations, should factor direct energy production or long-term energy contracts into their strategic planning, as the scramble for reliable power becomes a bottleneck.
AI & Machine Learning
Automata from Agent Traces: Failure and Next-Step Prediction in LLM Agents
Researchers, including Seonglae Cho, published a paper introducing a method to extract compact finite-state machines (FSMs) from large language model (LLM) agent traces. These FSMs provide a structural substrate for auditing and runtime monitoring, enabling more accurate next-step prediction and early failure detection in complex multi-step tasks. The FSMs demonstrate high fitness in replaying data and are computationally efficient to build.
Why it matters: The core problem this paper tackles — making LLM agent behavior predictable and auditable — is a major barrier to enterprise adoption, not just a research curiosity. East Asian industrial firms, particularly in manufacturing and logistics, are keen to automate complex workflows with AI agents, but they demand reliability and clear accountability that current black-box LLMs often lack. This method offers a path to meeting those industrial requirements.
For Western readers: Western enterprises deploying LLM agents, especially in regulated industries, should assess how this FSM-based approach could be integrated into their existing MLOps and safety frameworks to enhance observability and reduce operational risk, as the need for explainable AI becomes more pressing globally.