East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered often.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- TSMC’s planned $100 billion investment expansion in the United States secures the physical capacity needed to meet soaring global AI hardware demand, but it also exposes the Western supply chain’s total reliance on Taiwanese manufacturing execution to bring these advanced nodes online.
- Nvidia’s expanded partnerships in Japan, ranging from industrial robotics leaders to sovereign cloud providers, lock in the company’s hardware ecosystem across Japanese heavy industry before domestic METI-backed alternatives can establish a foothold.
- Xiaomi’s deployment of its own humanoid robots on its Beijing smart factory assembly lines moves the company ahead of Western consumer electronics rivals by using its existing manufacturing footprint as a real-world testing ground for embodied AI.
Core Move
Claude, Codex, and ChatGPT Work: Usage Limits Reset Again Amidst ‘Reset Battle’
Anthropic and OpenAI are quickly resetting usage limits for Claude, Codex, and ChatGPT Work. This change shows that the AI market is entering a phase of intense, slow battle. Firms now value fast user growth and developer attention more than making money from each user. This is not a new product cycle. It is a direct fight for market share through low prices and easy access. This shift affects how big companies can safely use these tools.
These limit resets are not about new technology. They are bold business moves to win customers. They prove that the race to dominate the market is still very active. Japanese media outlets like ITmedia AI+ call this a reset battle. They focus on who is winning the war for new users right now. This view shows that long-term profit depends on high volume and keeping developers on one platform.
Japanese firms are usually slow to adopt new tech. For them, these unstable policies bring new risks. They want to build on a single big Western model. Yet, they must now worry about sudden changes to business terms. This situation is like the early browser wars between Netscape and Internet Explorer. Those firms fought for market share by bundling features. They often gave away software for free to control the market.
Japan is now seeing its own version of that war. This time, the fight is over the basic layers of enterprise AI. The big danger for firms is not just changing costs. The main threat is the weak foundation created by relying on unstable vendors. A provider might change its rules twice in a single week.
It is wrong to assume that current AI prices and usage rules are stable. Providers do not know what the market will pay yet. They are using their own customers as weapons against their rivals. This fight will speed up the growth of safer business tools. These tools will likely use open-source models or smaller, specialized Japanese AI services.
You should watch for three clear signs next. First, see how fast Japanese system integrators like Fujitsu or NEC launch managed AI services. These services can protect buyers from vendor changes. Second, watch the growth of custom open-source models like Llama in big firms. Third, track public statements from major Japanese companies like Toyota or Mitsubishi about how they buy AI. Watch if they start to favor using many vendors or choosing local options.
🗾 Japan Radar
What Japanese media is reporting that Western outlets miss
Nvidia and TSMC are anchoring Japan’s sovereign AI ambitions to physical infrastructure rather than frontier software development.
🗾 AI & Machine Learning
Claude, Codex, and ChatGPT Work: Usage Limits Reset Again Amidst ‘Reset Battle’
Anthropic announced a reset of its 5-hour and weekly usage limits for its AI service, Claude, on July 16th. Shortly after, OpenAI executives reported similar resets for their desktop AI service, ChatGPT Work, and AI coding tool, Codex, suggesting a competitive ‘reset battle’ between the two companies. This marks the second time in a week, following similar resets on July 9th, indicating intensifying competition in the AI service market. The speed of these resets, happening twice in a week, shows that these companies are reacting in near real-time to each other’s moves. It reflects intense pressure to maintain user engagement and attract developers, which often means sacrificing short-term revenue per user for market share in the long run. The local framing emphasizes the ‘battle’ aspect, focusing on who is winning the immediate user-acquisition war.
For Western readers: Western enterprises should assume that major AI model providers will continue to adjust usage policies rapidly; factor this unpredictability into long-term architectural planning and avoid single-vendor lock-in where possible.
🗾 AI & Machine Learning
Gemini Spark Released in Japan, Starting with Ultra — A 24/7 ‘Personal AI Agent’
Google has launched its personal AI agent, Gemini Spark, in Japan, initially as a beta for Gemini Ultra subscribers (14,500 yen/month and up). The service, powered by Gemini 3.5 and the Google Antigravity platform, integrates with Google Workspace tools and partner apps to autonomously perform tasks like information gathering, data organization, and scheduling 24/7 in the cloud. Google hinted at future expansion to Pro users (2,900 yen/month). The introduction of a deeply integrated, always-on AI agent like Gemini Spark in Japan shows that major Western tech players are not just translating models, but building out full-stack, ecosystem-level AI services. The ability for Spark to connect with custom Model Context Protocols (MCP) and third-party apps means Google is making a direct play for the ‘operating system’ layer of the AI stack, potentially giving it an edge in an market where many local firms are still focused on building foundational models.
For Western readers: Western enterprise AI strategists should recognize that Google’s agent-based approach, combining its own models with an agent development platform and deep app integration, sets a new bar for AI productization in major markets. Assume this type of fully autonomous, integrated AI will become the standard user expectation within 18-24 months, pushing beyond current chat-interface-only models.
AI & Machine Learning2 STORIES
Nvidia Deepens Ties in Japan, Securing Industrial and Sovereign AI Partnerships
Nvidia CEO Jensen Huang’s visit to Japan has solidified major collaborations, expanding the company’s partnership with Toyota into smart cities and factory automation while backing the SoftBank-led ‘sovereign AI’ developer Noetra. Together, these moves integrate Nvidia’s hardware and software ecosystem deeply into Japan’s industrial infrastructure and government-backed AI initiatives.
Why it matters: Nvidia is not just selling chips here; they are embedding their software stack and ecosystem into Toyota’s operations, which is a key industrial keiretsu. This approach builds dependency and ensures their hardware and software become the default, rather than just another component, making it harder for competitors to displace them later. This strategy extends beyond automotive into manufacturing and urban infrastructure.
For Western readers: Western AI and industrial technology companies should recognize that winning in Japan requires deep, long-term ecosystem partnerships with major players like Toyota, rather than just product sales, to secure market share and influence in other East Asian markets.
Semiconductors & Hardware
TSMC plans additional $100bn US investment for AI chip demand
TSMC announced an additional $100 billion investment in the U.S. for new cutting-edge chipmaking and advanced packaging facilities in Arizona, driven by soaring AI demand. This move follows the company’s full-year capital expenditure forecast increase to $64 billion and a projected 40% revenue growth. The strategic expansion aims to meet increasing global AI chip requirements, further solidifying TSMC’s role in the advanced semiconductor supply chain. While Western coverage often frames this as a win for American chip independence, the local context highlights Taiwan’s strategy to balance geopolitical risks with economic opportunity, using ‘AI island’ rhetoric to justify global expansion. The real leverage for Taipei here is in sustaining its critical role, not just exporting it. The Japanese and Chinese read this as a further solidification of US supply chain influence, and a challenge to their own domestic chip ambitions.
For Western readers: If you are a Western investor or buyer, understand that TSMC’s U.S. expansion will not immediately alleviate the reliance on Taiwanese manufacturing for the most advanced nodes, but it does signal a longer-term shift in the geographic footprint of packaging.
🇨🇳 China Watch
China’s technology moves, framed for Western readers
China bypasses commercial pressure by open-sourcing foundational infrastructure to secure long-term architectural dominance in safety, robotics, and agentic workflows.
Robotics & Automation2 STORIES
Xiaomi Advances Embodied AI with Factory Deployment and Open-Source Model
Chinese consumer tech giant Xiaomi has successfully deployed its self-developed humanoid robots onto its electric vehicle production lines while simultaneously open-sourcing Robotics-U0, a massive 38-billion-parameter embodied generative AI model. Together, these milestones demonstrate Xiaomi’s rapid vertical integration of physical robotics and advanced AI to automate complex, real-world manufacturing tasks.
Why it matters: Xiaomi’s deployment of its own robots in its car factory is less about a global robotics breakthrough and more about industrial policy: China wants to build out its domestic robotics supply chain and decrease reliance on foreign automation. This is a clear step towards that goal, integrating a Chinese consumer tech giant’s AI into an industrial setting.
For Western readers: Western automotive manufacturers should expect Xiaomi and other Chinese EV players to achieve greater production autonomy and potentially lower labor costs through advanced domestic robotics, challenging existing manufacturing benchmarks.
AI & Machine Learning
Ant Group Open-Sources AI Safety Model, Details Multimodal Guardrails
📊 Featured Chart
Ant Group model sizes
Ant Group’s AI Safety Lab has open-sourced SingGuard-NSFA, a safety guardrail model specifically for autonomous agents, while also detailing its multimodal safety model, SingGuard. These models aim to detect various AI risks, including prompt injection and data theft, across 133 languages and hundreds of scenarios. The open-sourcing makes advanced AI safety tools developed in China accessible to a wider developer community. Chinese tech companies are investing heavily in AI safety, not just for compliance with Beijing’s tightening regulations, but also to build trust in their platforms as they push into more sensitive applications. This move by Ant Group demonstrates their technical capabilities in a critical, often understated, aspect of AI development — building dependable guardrails for autonomous systems. It is also an effort to shape the discourse around AI safety and establish their standards, rather than solely adopting Western frameworks.
For Western readers: Western businesses developing or deploying AI agents should note that robust, open-source safety models are emerging from Chinese firms, potentially setting a competitive benchmark for reliability and risk mitigation that extends beyond mere performance metrics.
AI & Machine Learning
Zhipu AI Founder Outlines ‘Touch High’ AGI Research Plan, Prioritizing Long-Term Goals Over Immediate Commercialization
Chinese foundation model developer Zhipu AI announced an internal ‘Touch High’ plan, shifting its focus from short-term commercialization to long-term AGI research. Founder Tang Jie outlined four technical priorities, including autonomous agent systems and extreme safety governance. Tang Jie calling out ‘mechanistic interpretability‘ as a major safety direction is notable. Western coverage often frames China’s AI progress as purely about scale and speed, but this shows Zhipu AI thinking about deeper, more fundamental safety challenges that are critical for AGI development. This isn’t just a marketing pivot; it reflects a genuine technical problem set.
For Western readers: Western businesses assessing Chinese AI capabilities should adjust their models to factor in a longer R&D horizon for companies like Zhipu, meaning their core competitive advantage may shift from rapid application development to fundamental model breakthroughs over time.
AI & Machine Learning
Microsoft and Renmin University Open-Source ‘Flint’ for AI Chart Generation
Microsoft has partnered with Renmin University in China to open-source ‘Flint,’ a framework designed to resolve conflicts in AI-generated charts. This collaboration aims to enhance the reliability and consistency of visual data representation from large language models, addressing a critical challenge in AI’s practical application. The focus on ‘Flint’ and resolving chart generation clashes speaks to an underlying truth: current LLMs often struggle with factual accuracy and consistent data representation, especially for structured output. This initiative, while framed as a technical solution, is also a subtle move by Microsoft to embed its tools deeper into Chinese academic and developer ecosystems, a long-standing strategy.
For Western readers: Western businesses building AI tools that generate data visualizations should anticipate similar solutions emerging rapidly from China, which will likely be optimized for Mandarin data and specific local chart conventions. Do not assume Western-led open-source frameworks will be the default for these applications in East Asia.
🔺 The Triangle
Where US, Japan, and China technology interests intersect
East Asia is bypassing the software bottleneck, leveraging physical infrastructure, advanced packaging, and novel materials to control AI’s hardware backbone.
Semiconductors & Hardware
TSMC’s Q2 Revenue Up 34% YoY, Plans Another $100 Billion Investment in Arizona Fabs
📊 Featured Chart
Advanced technologies (7nm and above) accounted for 77% of total wafer revenue
TSMC reported Q2 2026 revenue of $40.20 billion, a 33.7% year-over-year increase, driven by strong demand for its advanced 2nm and 3nm process technologies. The company announced a further $100 billion investment in its Arizona facilities, bringing the total commitment there to $265 billion for four more fabs and a packaging plant. TSMC’s financial performance shows that despite geopolitical pressures, the demand for cutting-edge logic chips remains insatiable, and Taiwan continues to be the only real source for them. The Arizona expansion confirms the trend of diversification, but the sheer scale of the investment makes it clear that the US and its allies are in it for the long haul to build out their own advanced manufacturing base, even if it is still reliant on Taiwanese know-how.
For Western readers: If you are a Western AI chip designer or system integrator, assume your leading-edge chip supply will continue to primarily originate from Taiwan for the foreseeable future, despite US onshore fab buildouts. The Arizona fabs will take time to ramp, and the expertise remains concentrated in Taiwan.
Semiconductors & Hardware
SILITH, UMC Reach Silicon Photonics Mass Production Milestone for AI Optical Interconnects
Taiwan’s United Microelectronics Corp. (UMC) and SILITH Technology have achieved mass production of silicon photonics wafers for AI and hyperscale data center networks at UMC’s Singapore facility. This milestone enables high-volume manufacturing of SILITH’s 1.6T silicon photonics platform, addressing the increasing demand for high-speed optical interconnects in AI infrastructure. UMC’s entry into silicon photonics mass production signals that the foundational technology for AI data center interconnects is maturing into a scalable manufacturing process, not just a lab curiosity. This is not just about a new product; it is about UMC extending its foundry dominance beyond traditional CMOS into a new critical component area for AI compute. The timeline — 18 months from development to production readiness — is unusually fast for this level of technology, suggesting strong execution and demand pulling the product through.
For Western readers: Western cloud and AI infrastructure providers relying on high-speed optical interconnects should assume UMC will become a significant manufacturing source for these components, reducing dependency on niche suppliers and potentially stabilizing future supply chains. This collaboration solidifies a Taiwanese firm’s role in the core infrastructure enabling next-generation AI at scale.
Semiconductors & Hardware2 STORIES
Power, Not Pixels: East Asia Redefines the AI Infrastructure Race
As Asia-Pacific’s AI expansion faces critical energy constraints, the region is shifting focus from raw compute power to grid efficiency, highlighted by soaring data center demands in Malaysia and Singapore. In response, partnerships like Germany’s Infineon and South Korea’s LS ELECTRIC are pioneering next-generation DC power solutions to re-architect data center energy delivery from the ground up.
Why it matters: The focus on power, cooling, and overall system efficiency for AI infrastructure is a more practical, boots-on-the-ground concern than the typical Western emphasis on benchmarked AI model performance or raw chip specifications. This shift highlights how operational realities—like a stable power supply and efficient thermal management—are becoming differentiators in the race to scale AI deployments, especially in regions with diverse energy grids and environmental considerations like Southeast Asia.
For Western readers: Western firms planning AI data center investments or supply chain build-outs in Asia Pacific should prioritize detailed energy infrastructure assessments and power management technology integration over solely optimizing for compute power, as regional growth will be gated by energy access and efficiency.
Semiconductors & Hardware
China Startup Yuanjiwei Claims World’s First 2D Semiconductor Pilot Line
📊 Featured Chart
Source: Yuanjiwei via SCMP
Shanghai-based Yuanjiwei has established what it claims is the world’s first pilot production line for 2D semiconductors, releasing a PDK for its 500nm 8-inch line and launching foundry tape-out services. The company aims for a 90nm equivalent process by late 2026 and a 5nm equivalent by 2029, entirely without EUV lithography. Announcing a ‘world’s first’ pilot line for 2D semiconductors signals China’s intent to build domestic capabilities that are fundamentally different from mainstream silicon, allowing them to circumvent current geopolitical restrictions on leading-edge manufacturing. The claim of reaching 5nm equivalent without EUV is audacious, and the critical question is execution. It shifts the focus from purely silicon-based process competition to materials science and novel architectures, an area where Chinese research has been strong.
For Western readers: If Yuanjiwei can genuinely achieve even a 90nm equivalent process using 2D materials by 2026, Western semiconductor roadmaps that assume continued reliance on EUV for advanced nodes will need a critical re-evaluation of long-term supply chain dependencies. Watch for concrete yield and performance data, not just process claims.
🧩 Pattern This Issue
- Japan: Nvidia cements sovereign AI partnerships with industrial and regional leaders
- Taiwan/US: TSMC commits $100 billion to scale cutting-edge Arizona fabs
- China: Xiaomi deploys proprietary humanoid robots onto its Beijing EV assembly lines
The AI race is shifting from software benchmarks to physical scale, positioning hardware manufacturing powerhouses as the ultimate gatekeepers of global AI infrastructure.
AsiaAI.FYI ·
Written by Dick Weisinger ·
Subscribe