East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- SoftBank Group CFO Yoshimitsu Goto’s defense of the sector’s capital expenditure amid a 40% drop in global AI startup funding to $149.5 billion in Q2 2026 demonstrates that capital is concentrating away from speculative software and toward physical infrastructure.
- NEC’s deployment of 17 specialized AI agents to run an entirely unmanned internal department serves as a real-world test bed for corporate automation, moving Japan’s AI strategy past defensive policy and into direct operational execution.
- The $6.3 billion joint investment by Sony and TSMC in a second Kumamoto fabrication plant secures Japan’s domestic supply of advanced image sensors and logic chips, cementing the country’s position as the primary hardware anchor for global AI and edge-computing devices.
Core Move
CXMT’s Rise: China’s Memory Champion Challenging Global Leaders
ChangXin Memory Technologies plans an $8.6 billion Shanghai IPO. This move shows that China has bypassed Washington’s export controls in the memory sector. It did this by replacing advanced tools with massive factory scale. Western analysts often focus on the company’s lack of advanced EUV machines. Yet, CXMT holds a 7.7% share of the global DRAM market. This share proves that older DUV tools can still build a strong, state-funded rival.
China is using its classic industrial playbook for computer chips. It wants to turn the technology into a cheap commodity. Then, it will scale up fast with state funds and underprice foreign rivals. Chinese media does not view CXMT as just a normal business. Instead, they see it as a key asset that protects China’s hardware supply chain from foreign shocks.
This fast growth shows a big mistake in Western export policy. That policy assumes that blocking new tools stops a chipmaker. In fact, memory production relies heavily on high volume and high yields. CXMT is perfecting its work by focusing on older DRAM. It is also building a strong group of local suppliers. This plan mirrors how Chinese solar and EV firms won global markets years ago. Those firms mastered high-volume work before they made fancier goods.
Yet, CXMT faces a major threat to its long-term survival. Its rapid growth could create a huge global DRAM oversupply. This glut would drop prices so low that even state funds cannot cover the high costs. If global prices fall, CXMT will be highly vulnerable. It lacks high-margin High Bandwidth Memory products, which require the banned EUV tools.
You can watch two clear signs over the next year to see if CXMT can succeed. First, look to see if the firm wins a major DRAM contract from a top Chinese server maker like Inspur. Second, check if its post-IPO spending files show better yields on its 17-nanometer “Qiangxin No. 1” chip. It needs to push those yields toward the 90% mark.
🗾 Japan Radar
What Japanese media is reporting that Western outlets miss
Japan is securing the physical hardware and corporate automation pipelines to survive global capital flight from pure software play AI.
Semiconductors & Hardware
Sony and TSMC to Invest $6.3bn in Advanced Kumamoto Image Sensor Plant
Sony Group and Taiwan Semiconductor Manufacturing Co. (TSMC) have agreed to invest $6.3 billion in a new joint venture in Kumamoto, Japan. The facility will focus on the development and mass production of next-generation image sensors, with operations targeted to begin as early as 2029.
Why it matters: Sony secures a cutting-edge domestic manufacturing partner to defend its 50%-plus market share in image sensors against Samsung. By anchoring TSMC deeper into Kumamoto, Japan’s Ministry of Economy, Trade and Industry (METI) successfully turns a regional agricultural hub into a highly resilient, advanced semiconductor cluster that is insulated from Taiwan Strait geopolitical risks.
For Western readers: Western smartphone and automotive OEMs should expect the next generation of high-end vision and autonomous sensing hardware to remain deeply dependent on Japanese fabrication, making Japan the critical gatekeeper for optical silicon supply chains by 2029.
🗾 AI & Machine Learning
Google’s Jeff Dean Establishes New Company Discovery Loop to Automate Large-Scale AI Experiments
Jeff Dean, Google DeepMind and Google Research’s Chief Scientist, has departed Google after 27 years to co-found Discovery Loop, a Public Benefit Corporation. The new company, also co-founded by Google Senior Fellows Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, aims to automate large-scale experimentation in machine learning research and engineering. Google CEO Sundar Pichai confirmed Google’s involvement as an initial investor and cloud partner, indicating continued collaboration.
Why it matters: Dean and his co-founders are among the most influential figures in modern AI infrastructure and research. Their departure and Google’s decision to invest in and partner with Discovery Loop reflects both the increasing fragmentation of top-tier AI talent and Google’s pragmatic approach to retaining influence over innovations from its former top researchers.
For Western readers: Western AI and tech companies should recognize that the competition for top AI talent is intensifying, with key figures increasingly opting for independent ventures or rival labs, often with financial backing from their former employers; consider how this changes your own talent retention strategies.
🗾 Semiconductors & Hardware
AI Data Centers ‘Overwhelmingly in Short Supply’ – SoftBank Group CFO Goto Denies ‘Bubble’ Claims
📊 Featured Chart
SoftBank Group Q1 FY2027 (Apr 1 – Jun 30, 2026)
SoftBank Group (SBG) CFO Yoshimitsu Goto stated that AI data centers are facing an “overwhelming supply shortage,” dismissing concerns of a bubble. He emphasized the necessity of their global data center expansion projects, including 5GW in France, planned centers in Texas for OpenAI, and a 10GW facility in Ohio.
Why it matters: SoftBank’s scale of investment into AI infrastructure, particularly the multi-gigawatt data centers, is a hard bet on the compute-intensive future of AI. The scale of these projects means they’re not just buying chips; they’re creating the foundational digital real estate for the next generation of AI services, aiming to control a critical bottleneck in the AI supply chain. This is less about building a ‘datacenter business’ in the traditional sense and more about enabling their core AI investments, ensuring their portfolio companies have the compute they need.
For Western readers: Western businesses relying on cloud compute for advanced AI development should recognize that capacity will remain a critical, expensive constraint for years, despite large investments. SoftBank’s move, especially with explicit OpenAI ties, suggests that securing compute will increasingly dictate who can scale AI initiatives. Don’t assume generalized cloud capacity will be sufficient; targeted, high-power AI data centers are a distinct and limited resource.
🗾
AI Startup Funding Drops 40% in Q2 2026 as Capital Concentrates in Mega-Rounds
Global AI startup funding fell 40% quarter-on-quarter to $149.5 billion in Q2 2026, driven by a sharp concentration of capital where mega-rounds exceeding $100 million accounted for nearly 90% of the total. Despite the overall funding drop, investor enthusiasm remains high as the quarter produced 37 new unicorns—the highest in four years—led by high-profile firms like Anthropic, Jeff Bezos’s Project Prometheus, and China’s DeepSeek.
Why it matters: This extreme concentration of capital creates a winner-take-all dynamic, starving mid-tier AI startups of growth capital while giving a select few giants the massive war chests needed to monopolize top-tier talent and advanced semiconductor compute resources.
For Western readers: Western investors and executives should abandon the assumption of US dominance in the AI unicorn pipeline, as Asian competitors—particularly Chinese firms like DeepSeek—are capturing a rapidly growing share of high-valuation milestones despite ongoing geopolitical tensions.
🗾
NEC Launches Unmanned Department Staffed by 17 AI Agents to Drive Corporate Automation
NEC has established a new, entirely unmanned “Corporate AI Workforce Department” staffed by 17 specialized AI agents acting as managers and staff to automate company-wide administrative tasks. Launched on November 1, the department operates like a human organization—complete with internal executive meetings where AI managers evaluate subordinates—to handle complex workflows starting with executive-level management analysis and sales data processing. Humans will not be replaced, but will instead oversee the department’s operations, monitor risks, and review the recorded decision-making processes of the AI agents.
Why it matters: This initiative shifts the corporate AI paradigm from individual co-pilots to autonomous, multi-agent organizational structures that self-govern and scale. By establishing a formalized “AI department,” NEC creates a blueprinted methodology for enterprise-wide automation that reduces the technical friction other business units face when trying to adopt AI on their own.
For Western readers: Western executives should expect Japanese IT giants to emerge as major exporters of turnkey, autonomous organizational software designed to run entire corporate sub-divisions without human staff.
🇨🇳 China Watch
China’s technology moves, framed for Western readers
Beijing is rapidly constructing a low-cost, domestically insulated AI stack to bypass Western sanctions and dominate global software deployment.
AI & Machine Learning2 STORIES
DeepSeek Drives Chinese LLMs to Global Token Dominance via Extreme Cost Efficiency
Chinese large language models have secured the highest volume of global token utilization for fifteen consecutive weeks, propelled by aggressive price cuts and software-level optimizations. At the forefront is DeepSeek, whose newly released V4-Flash model has claimed the top spot in active token volume by utilizing advanced cache-hit systems and engineering workarounds to bypass hardware constraints. Together, these developments show how Chinese AI firms are winning a global volume war through ultra-low-cost APIs that challenge the economics of traditional frontier models.
Why it matters: In the East Asian business ecosystem, this price-to-performance breakthrough accelerates local enterprise AI adoption by lowering the financial barrier to entry, forcing regional competitors to pivot their strategies from hardware-centric scaling to extreme algorithmic and operational efficiency.
For Western readers: Western tech leaders must abandon the assumption that US export controls will stall Chinese AI progress; instead, they must actively prepare for a market where Chinese software-level optimizations make Western frontier models economically uncompetitive for mass-market applications.
AI & Machine Learning
Tencent Elevates Desktop AI Agent WorkBuddy to Top Strategic Priority
Tencent has designated its WorkBuddy desktop AI office agent as a top strategic priority, placing it on an internal scale comparable to its flagship products QQ and WeChat. The tech giant is aggressively backing the product with massive advertising campaigns across Beijing and Shenzhen, alongside significant allocations of computing and organizational resources.
Why it matters: Tencent is skipping the consumer AI chatbot hype cycle to focus squarely on enterprise workflow automation where actual economic value can be captured. By integrating directly with local files and multi-step task execution, WorkBuddy aims to become the operating system for Chinese office productivity, entrenching Tencent at the center of corporate IT budgets.
For Western readers: Western enterprise software providers active in Asia must assume that generic AI copilots will face brutal competition from deeply integrated, localized agents like WorkBuddy, making standalone SaaS offerings increasingly difficult to sell in China without deep local workflow integration.
Policy & Regulation
Apple Walks Back Qwen AI Integration Guide for Mainland China
Apple has removed a Chinese-language Mac support guide that explicitly mentioned integrating Alibaba’s Qwen model into Apple Intelligence for users in mainland China. Customer service representatives stated that no such feature has been officially launched, backtracking on documentation that briefly went live on August 8.
Why it matters: This sudden retraction exposes the friction Apple faces in securing regulatory clearance and final agreements with local Chinese AI vendors. Securing a partner like Alibaba is critical for Apple to arrest its market share decline in China, but state security reviews of foreign-domestic integrations mean these rollouts are prone to administrative delays and communication breakdowns.
For Western readers: Do not assume Apple will have a functional, competitive AI offering in the Chinese market anytime soon; expect a fragmented device experience where Chinese domestic brands continue to hold a structural advantage in native AI features due to pre-approved local LLM integrations.
Enterprise & Cloud
China’s First 100,000-Card Domestic AI Super-Cluster Goes Live, and the National Compute Grid Takes Shape
China has launched its first ultra-large AI computing cluster powered entirely by over 100,000 domestically produced AI chips, marking a major milestone in its self-reliance drive. This massive super-cluster is integrated into the national ‘East-West Computing Resource‘ strategy, designed to systematically rebalance computing power from China’s coastal tech hubs to energy-rich western provinces.
Why it matters: This deployment proves that domestic Chinese silicon can scale to massive cluster sizes, bypasses the software bottleneck of Nvidia’s CUDA ecosystem through unified compiler layers, and provides state-backed enterprises with a viable, sanction-proof pipeline for training large language models.
For Western readers: Western tech leaders must abandon the assumption that US chip export bans will permanently stall China’s frontier LLM development, as Beijing is successfully substituting single-chip performance deficits with massive, state-coordinated cluster scale.
🔺 The Triangle
Where US, Japan, and China technology interests intersect
Asia’s expanding manufacturing capacity and specialized packaging ventures are decoupling hardware execution from Western frontier model development.
Semiconductors & Hardware
TSMC July 2026 Revenue Surges 44.7% as Kumamoto Sensor Plant with Sony is Target for 2029
TSMC reported July 2026 revenue of approximately $14.5 billion, representing a 44.7% year-on-year surge and a 5.6% monthly increase from June. Cumulative revenue for the first seven months of 2026 reached $89 billion, a 37% jump compared to the same period in 2025. Additionally, TSMC and Sony announced a joint $6.4 billion investment to construct a specialized sensor manufacturing plant in Kumamoto, Japan, slated to begin production in 2029.
Why it matters: While the market focuses on TSMC’s blistering monthly revenue growth as an indicator of AI chip demand, the real strategic story is the $6.4 billion Kumamoto sensor plant with Sony. This project cements a long-term supply pipeline for advanced image sensors and specialized silicon right on Japanese soil, giving Japan’s automotive and industrial sectors a protected domestic supply of critical edge-AI components.
For Western readers: Western hardware buyers should expect Taiwan-Japan co-production to dominate the high-end image sensor and edge-silicon markets by 2029, making supply chain integration with Japanese ecosystems essential for next-generation robotics and automotive designs.
Semiconductors & Hardware
Taiwan’s Powertech Partners with Broadcom in $5.66B Singapore Packaging Venture, Landing AMD as First Client
Taiwanese packaging specialist Powertech Technology is investing $400 million for a 30 percent stake in a joint venture with Broadcom called XCIP Technologies in Singapore. The facility will deploy fan-out panel-level packaging (FOPLP) to manufacture massive AI processors, with AMD signed on as the anchor customer for shipments beginning in the second half of 2027.
Why it matters: While TSMC holds a near-monopoly on high-end AI packaging with CoWoS, this venture establishes a viable, high-volume alternative in Singapore, boosting production efficiency by yielding roughly 45 chips per panel compared to just eight or nine on traditional silicon wafers. By anchoring the project with AMD and Broadcom, Powertech is positioning Singapore as a premier neutral hub for advanced packaging, bypasses the bottleneck of Taiwanese fabrication plants, and reduces reliance on TSMC’s packaging pipeline.
For Western readers: If you design or source high-performance AI accelerators, qualify panel-level packaging design rules now, as AMD’s commitment to this 2027 pipeline indicates the industry is actively shifting away from wafer-level packaging to resolve chronic backend supply constraints.
Semiconductors & Hardware
Global Semiconductor Sales Surge as China and Asia-Pacific Lead Month-on-Month Growth
📊 Featured Chart
Source: Semiconductor Industry Association
Global semiconductor sales reached $403.3 billion in Q2 2026, driven by robust demand in China and the broader Asia-Pacific region. China posted a 112.8% year-on-year increase and led month-on-month growth at 10.4%, cementing its position as a primary demand driver despite ongoing trade restrictions.
Why it matters: China’s double-digit month-on-month growth proves that domestic procurement remains relentless, signaling that US export controls are failing to suppress overall Chinese chip consumption. Instead, these measures are forcing a massive build-out of legacy and near-frontier capacity that keeps East Asian supply chains running at hot utilization rates.
For Western readers: Do not assume export curbs are starving the Chinese tech sector of silicon; plan for Chinese buyers to continue outbidding Western peers for global mature-node capacity.
AI & Machine Learning
China’s Alibaba Releases Wan3.0 AI Video Model in Public Beta with 30-Second Clips
Alibaba Cloud has launched the public beta of Wan3.0, an AI video generation model capable of producing continuous 30-second clips, doubling the length of its predecessor. The model accepts multimodal inputs including PDFs, presentations, and audio, and features specialized capabilities to maintain high visual continuity and realistic facial expressions.
Why it matters: By integrating document inputs like PDFs and PowerPoints directly into a 30-second video generator, Alibaba is targeting immediate enterprise automation and marketing workflows rather than just creative filmmaking. This provides Chinese businesses with a highly integrated, domestic-hosted alternative to Western creative tools, accelerating the adoption of generative AI in local corporate operations.
For Western readers: Western enterprise software and marketing platforms should expect Alibaba Cloud’s integrated multimodal-to-video workflow to set a new baseline for business productivity tools, forcing Western SaaS providers to integrate similar end-to-end document-to-video capabilities to remain competitive in multinational deployments.
🧩 Pattern This Issue
- Japan: SoftBank and TSMC spend $12B+ building physical AI hardware foundations
- China: Domestic 100k-card super-clusters scale despite Western export restrictions on advanced silicon
- China: DeepSeek and Tencent pivot to ultra-cheap tokens and desktop agents
While Western venture capital pulls back from software-centric AI rounds, East Asian players are aggressively bifurcating the market by building massive, low-cost physical infrastructure on one end and hyper-monetized, highly efficient local agents on the other, leaving Western software-only players squeezed on both infrastructure costs and localized execution.
AsiaAI.FYI ·
Written by Dick Weisinger ·
Subscribe