East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- Yang Zhilin’s rejection of Apple to build Moonshot AI illustrates how China’s top tier of US-trained research talent is increasingly opting to build domestic competitors to Anthropic rather than join Silicon Valley giants.
- Foxconn’s defense of its position in Nvidia’s modular server supply chain shows that Taiwan’s manufacturing giants maintain lock-in on high-end AI hardware execution despite intense competition from rivals aiming to chip away at their margins.
- The projection by Gartner that $234 billion in enterprise SaaS spending will vanish by 2030 because of agentic AI indicates a massive structural shift where software-as-a-service licensing models will be hollowed out by autonomous, continuous-execution workflows.
Core Move
China-Born AI Agent ‘Manus’ to Spin Off from Meta; Chinese Government Opposes Acquisition, Prompting Deletion of Some User Data #
The Chinese government blocked Meta’s $2 billion buyout of AI agent startup Manus. This move shows a hard geopolitical limit for global AI deals. Beijing now views consumer-facing AI agents as critical national security assets. The National Development and Reform Commission (NDRC) forced Manus to spin off. The agency also ordered Manus to delete user data created during Meta’s brief custody.
Beijing has stopped the flow of AI tech and data from Chinese-founded firms to US tech giants. This represents a structural split of the global AI software market. Western executives usually view startups as targets for eventual buyouts. For them, this split shows how Chinese regulators enforce data sovereignty. Western media focused on the financial loss of the $2 billion deal. In contrast, Chinese media praised the NDRC for protecting domestic AI talent and behavioral data.
Beijing planners refuse to let a top Chinese AI agent be bought by a US giant like Meta. These agents can perform complex tasks across the web. To Beijing, losing this tech is like exporting the best tools of industrial automation. This intervention is China’s version of the US Committee on Foreign Investment (CFIUS). Yet, Beijing applies it to software layers rather than hardware chips.
Washington blocks Chinese investment in US chip factories. In the same way, Beijing now protects the software where AI interacts with the real economy. Many assumed Chinese AI startups could bypass local rules by registering overseas or selling to Western buyers. That view is now dead. If the engineering and data trace back to China, Beijing claims control.
This forced split will hurt early-stage Chinese AI founders. It cuts off their access to Western money. These startups must now rely on Chinese state funds. Those state funds demand strict alignment with Beijing’s goals. Without cash from Western buyers or global stock listings, these startups may struggle. They need massive funding to buy the computing power to compete globally.
To see how deep this split will go, keep track of three signs over the next six months. First, see if Manus can raise at least $500 million from Chinese giants like Tencent or Baidu. Second, watch the Cyberspace Administration of China (CAC) audits to verify that Manus deleted the data. Third, observe whether Western venture firms like HongShan change how they fund global AI teams.
🗾 Japan Radar #
What Japanese media is reporting that Western outlets miss
The race for AI dominance has shifted from frontier models to hardware integration and agent-driven industrial execution.
🗾 AI & Machine Learning
SpaceXAI has launched the beta version of ‘Grok Bot,’ a continuous, cloud-based AI agent designed to act as an autonomous digital colleague. Unlike traditional AI agents that rely on restricted developer APIs, Grok Bot operates through its own virtual machine on the cloud, interacting with web interfaces and applications just like a human user by logging in and executing workflows. The service is available to subscribers of premium tiers, specifically Grok’s ‘SuperGrok Heavy’ and Cursor’s ‘Ultra’ and ‘Premium Teams’ plans.
Why it matters: Operating agents inside dedicated cloud virtual machines sidesteps the integration bottlenecks of API ecosystems, allowing immediate compatibility with any legacy enterprise software that has a web login. The ability of these bots to collaborate in hierarchical groups and learn custom routines by observing user demonstrations suggests a fast path to automating routine white-collar workflows without costly software overhauls.
For Western readers: Western enterprises should prepare for a rapid shift in SaaS consumption, where the primary user of enterprise application licenses shifts from human employees to autonomous, cloud-hosted virtual agents. 🗾 Enterprise & Cloud
Research firm Gartner projects that up to $234 billion (approximately 37 trillion yen) of enterprise software spending will be exposed to “agentic arbitrage” by 2030, representing about 20% of the global SaaS market. This phenomenon occurs as autonomous AI agents execute tasks across multiple systems, rendering traditional user interfaces obsolete and breaking the link between user seat-count and software vendor revenue. Consequently, the enterprise software market is shifting from seat-based subscription pricing to performance and outcome-based monetization models.
Why it matters: The shift toward agentic AI means corporate IT buyers will stop paying for seat licenses and start paying strictly for completed tasks. Established SaaS giants unable to capture the underlying agent workflows will see their margins cannibalized by nimbler, cross-platform AI agent startups that control the actual execution layer.
For Western readers: If your company is renegotiating multi-year enterprise SaaS contracts, push for performance-based pricing tier options now to avoid overpaying for seat licenses that autonomous AI agents will render redundant within 3 to 4 years. Robotics & Automation
China’s Humanoid Robot Makers Race to Public Markets Chinese humanoid robot developer Unitree and its domestic competitors are accelerating plans for public listings to fund capital-intensive commercialization efforts. Driven by local government subsidies and industrial mandates to automate manufacturing, these firms are transitioning from trade-show prototypes to factory-floor deployments.
Why it matters: Chinese hardware players are executing a classic volume-and-cost play, utilizing state-guided industrial funds to subsidize early-stage losses. The rush to IPO is less about market maturity and more about securing a capital runway before domestic subsidy pools dry up or consolidate around a few state-selected champions.
For Western readers: Western robotics developers must prepare for a flood of low-cost Chinese humanoid components and platforms entering the global market, which will squeeze hardware margins and shift the competitive moat entirely to proprietary AI control software. Semiconductors & Hardware
Foxconn Downplays Competitor Threats in Nvidia Modular AI Server Supply Chain Foxconn’s newly appointed rotating CEO Michael Chiang stated that Nvidia’s transition to modular AI server platforms will help the company increase its market share rather than dilute it. The contract manufacturing giant claims its massive vertical supply chain and unmatched production scale insulate it from emerging competitors targeting the AI infrastructure boom.
Why it matters: Nvidia’s modular design shift acts as a consolidator for players with massive balance sheets rather than an equalizer for smaller assemblers. Foxconn’s ability to manufacture everything from connectors to liquid cooling systems internally means it can absorb margin pressure that would crush rivals trying to enter the AI server space.
For Western readers: Western hyper-scalers should design their infrastructure pipelines around the reality that Foxconn and a few select Taiwanese tier-one partners will maintain a virtual lock on next-generation rack-level deployments, limiting procurement negotiation leverage. Startups & Funding
Moonshot AI’s CEO turned down Apple and built Chinese rival to Anthropic
Yang Zhilin, a prominent AI researcher who earned his doctorate in the United States, rejected job offers from major American technology firms like Apple to return to China. He co-founded Moonshot AI, the startup behind the Kimi chatbot, positioning it as a direct domestic competitor to leading Western laboratories like OpenAI and Anthropic.
Why it matters: US-educated Chinese researchers returning home accelerated the speed at which Chinese foundational models caught up with Western counterparts. This pipeline of high-caliber talent enables Chinese startups to build highly competitive architectures, such as Moonshot’s Kimi, which can run efficiently on constrained domestic hardware.
For Western readers: Do not assume that tightening US visa or trade restrictions will permanently lock in a Western AI talent advantage; instead, expect these pressures to act as a catalyst that drives top-tier Chinese researchers back to the domestic ecosystem where they will build formidable competitors.
🇨🇳 China Watch #
China’s technology moves, framed for Western readers
U.S. chip curbs are forcing Chinese AI giants to abandon frontier LLM races for infrastructure workarounds and practical agent applications.
Startups & Funding
Lin Junyang, the former chief architect of Alibaba’s Qwen LLM team, has departed to launch Pragmatik Labs, securing a $2 billion valuation. The new venture signals a decisive shift away from foundational model pre-training toward specialized AI agent workflows.
Why it matters: Pragmatik’s immediate multi-billion-dollar valuation demonstrates that Chinese venture capital is still highly concentrated around top-tier talent, even as it retreats from funding general foundation models. By skipping the infrastructure phase and focusing entirely on agents, Lin is positioning his new venture to capture enterprise utility before the underlying LLM market completely commoditizes.
For Western readers: Western enterprise software vendors should prepare for a wave of highly agile, agent-native Chinese productivity tools competing aggressively in global markets, bypassing the high compute-cost barriers that hobble traditional LLM startups. AI & Machine Learning
Zhipu’s API User Base Nears 7 Million as It Adds 50,000-Plus Chinese AI Chips Chinese generative AI pioneer Zhipu AI has expanded its model-as-a-service platform to nearly 7 million registered API users, adding 2 million users in just one month. To support this massive surge in inference demand, the company activated over 50,000 domestic AI chips and lifted purchase restrictions on its coding assistant model.
Why it matters: Zhipu’s rapid scaling proves that domestic Chinese silicon is no longer just a laboratory substitute; it is actively handling production-level inference workloads for millions of users. This successful deployment of 50,000 local chips indicates that China’s domestic AI ecosystem is bypassing the Nvidia bottleneck for mainstream software applications.
For Western readers: Western enterprises assuming that US chip sanctions will choke off China’s generative AI application market must revise their timelines, as domestic software players are already running highly scalable API businesses on local hardware. Enterprise & Cloud3 STORIES
Chinese AI Giants Pivot to Physical Infrastructure Amid U.S. Chip Curbs To survive tight U.S. export controls on advanced silicon, Chinese tech leaders are aggressively taking direct control of their physical infrastructure, as seen in DeepSeek hiring civil engineers for in-house data centers, Tencent’s massive capex surge for AI clusters, and Alibaba cutting facility construction times to 100 days. Collectively, these moves show a strategic shift from relying on public cloud scaling and raw chip performance to optimizing physical deployment, modular design, and hardware-software co-design.
Why it matters: In the East Asian business ecosystem, this infrastructure arms race is forcing a structural realignment where software supremacy is now entirely dependent on proprietary, highly customized physical hardware stacks, raising the capital barrier to entry for any aspiring local AI challenger.
For Western readers: Western leaders must abandon the assumption that restricting advanced chip exports will stall China’s AI ecosystem; instead, prepare to compete against Chinese firms that are becoming hyper-efficient at squeezing maximum performance out of legacy and domestic silicon through unprecedented physical infrastructure optimization.
🔺 The Triangle #
Where US, Japan, and China technology interests intersect
As the hardware supply chain tightens, the AI battleground has shifted from frontier model benchmarks to physical manufacturing sovereignty.
Semiconductors & Hardware
Sony-TSMC Joint Venture in Kumamoto Secures Japan’s Legacy Node Supply TSMC and Sony’s joint venture to construct a $4.7 billion semiconductor fab in Kumamoto, Japan, directly addresses Japan’s critical deficit in legacy logic nodes. The fab will produce 22nm and 28nm chips, securing local supply chains for Sony’s image sensors and Japan’s automotive sector. This strategic move leverages TSMC’s manufacturing scale to counter Samsung’s aggressive expansion in the image sensor market.
Why it matters: Sony secures a dedicated, domestic foundry partner to fabricate the logic layers of its advanced image sensors, protecting its dominant market share from Samsung’s vertically integrated threat. For TSMC, the venture deepens its integration into the Japanese industrial ecosystem, building a defensive moat of political and financial support that rivals cannot easily replicate.
For Western readers: Western automotive and industrial OEMs sourcing components from Japan should expect improved supply chain resilience for legacy nodes by 2024, but must prepare for higher domestic chip packaging costs as Japan establishes its own packaging ecosystem. Semiconductors & Hardware
Testing and Metrology Emerge as a New Front in the AI Race At a press conference in Taiwan, SEMI and key local industry leaders, including Chroma ATE and KYEC, called for dedicated policies and shared infrastructure to support the testing and metrology sectors. The industry group forecasts that global semiconductor testing equipment sales will surge 31% in 2026 to reach $15.3 billion, driven by the extreme complexity of advanced AI logic chips and high-bandwidth memory (HBM). Taiwan’s domestic metrology industry output already hit a record NT$193.6 billion last year, with early 2026 data showing a 32.8% year-over-year jump.
Why it matters: Yield is the ultimate arbiter of profitability and volume in advanced nodes. By coordinating a localized testing and metrology ecosystem through SEMI Taiwan, local players are building an operational moat around TSMC’s fabs that prevents overseas competitors from easily duplicating the advanced packaging ecosystem.
For Western readers: Western chip designers must accept that outsourcing fabrication to Taiwan means relying on Taiwan’s domestic ecosystem for post-fab yield verification; any attempt to onshore assembly to the US or Europe will stall without equivalent localized investment in advanced metrology and testing infrastructure. Semiconductors & Hardware
Advanced PCB Shortages Trigger 2028 Delivery Quotes from Chinese Manufacturers 📊 Featured Chart
Averages based on current industry reports. FR-4 values use midpoints.
A severe global shortage of printed circuit boards (PCBs) is forcing Chinese manufacturers to quote delivery lead times stretching into 2028 for orders placed last month. The crisis is driven by insatiable AI server demand, where a single rack requires up to ten thousand boards, and is severely compounded by a supply freeze of critical high-purity PPE resin from Saudi Arabia.
Why it matters: While the market focuses on high-bandwidth memory and advanced GPU yields, the humble PCB has become the actual physical bottleneck for AI scaling. Chinese fabricators are prioritizing high-margin AI clients, leaving legacy industrial, automotive, and consumer electronics customers completely stranded at the back of a multi-year queue.
For Western readers: If your product roadmap relies on advanced low-loss or standard FR-4 laminates and you do not have pre-allocated capacity, design-in alternative mechanical layouts or qualify non-Chinese, non-Indian fabrication partners immediately to avoid production halts in 2027. Policy & Regulation
China Blocks Meta’s $2B Acquisition of Singapore-Based Manus AI, Forcing Divestment Singapore-headquartered AI agent startup Manus AI is separating from Meta and deleting user data generated after December 2025 following a regulatory block by Chinese authorities. The divestment undoes a $2 billion acquisition completed in late 2025, forcing Manus to return to operating as an independent entity.
Why it matters: Beijing is demonstrating that physical relocation to Singapore does not insulate Chinese-founded tech startups from domestic regulatory vetoes, especially when high-value AI IP is being transferred to American tech giants. This intervention protects domestic AI capabilities from foreign absorption and signals tighter regulatory chokeholds on outbound technology transfers.
For Western readers: US tech buyers must assume that any acquisition of a startup with Chinese founders or historical R&D ties to China will face regulatory blocks from Beijing, even if the target is currently headquartered in Singapore or another neutral hub. Semiconductors & Hardware
AMD EPYC 9006 ‘Venice’ Server Processors Leverage TSMC 2nm Process for Enterprise AI AMD has announced its 6th Generation EPYC 9006 ‘Venice’ server processors, featuring up to 256 Zen 6 cores and support for PCIe Gen 6. The architecture relies on TSMC’s advanced 2nm silicon fabrication technology to deliver the density and memory bandwidth required for high-concurrency agentic AI pipelines.
Why it matters: TSMC’s 2nm yield and capacity allocation will directly dictate the rollout schedule of next-generation enterprise AI servers. AMD’s early commitment to this node for its flagship server lineup increases the competitive pressure on Intel and domestic Chinese server chip designers who lack access to TSMC’s most advanced nodes.
For Western readers: Western infrastructure buyers planning agentic AI deployments must tie their hardware procurement timelines directly to TSMC’s 2nm production milestones rather than AMD’s nominal release dates. 🧩 Pattern This Issue
China: Tech giants shift investments from raw compute to physical infrastructure and AI agentsChina: Former Qwen and Meta AI engineers pivot from foundation models to agentic platformsJapan: Sony-TSMC Kumamoto joint venture secures legacy chip supply for industrial manufacturing
As U.S. chip curbs choke raw compute access, East Asian players are shifting their capital from chasing frontier foundation models to securing the physical supply chain and deploying practical agentic AI, meaning Western firms focusing solely on LLM benchmarks are missing where the actual commercial battle is being fought.
AsiaAI.FYI · Written by Dick Weisinger ·