East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- South Korea’s record $40 billion in semiconductor exports for August confirms that global demand for AI infrastructure remains decoupled from the public equity market’s immediate anxieties over monetization.
- South Korea’s Upstage secured direct GPU allocation commitments during its recent visit to Nvidia’s headquarters, bypassing standard distributor bottlenecks to scale its localized large language models across East Asia.
- China’s CXMT doubled its DRAM sales in the second quarter of 2026 to capture a larger share of the $154.7 billion global memory market, demonstrating that domestic production expansion is successfully insulating Chinese hardware supply chains from Western export controls.
This Issue’s Analysis
The Signal
$2T IPO and Two Profitable Quarters: Why Anthropic Is Shifting the AI Valuation Playbook
Anthropic, a leading AI startup, has reportedly achieved profitability for two consecutive quarters, a significant milestone ahead of its anticipated initial public offering (IPO).
Semiconductors & Hardware
DRAM Market Surges 59%: How AI Infrastructure Demand Is Driving Up Average Selling Prices
The DRAM market expanded by 59.5% quarter-over-quarter to $154.7 billion in Q2 2026, driven by strong demand for HBM3E, LPDDR5X, and high-capacity RDIMMs for AI applications. Samsu
Semiconductors & Hardware
Murata’s iPaS Substrates: Solving the Power Delivery Bottleneck in AI Accelerators
Murata Manufacturing, the world’s largest producer of Multi-Layer Ceramic Capacitors (MLCCs), plans to mass-produce its Integrated Package Solution (iPaS) next year. iPaS integrate
AI & Machine Learning
US AI Safety Warning Wipes Billions From SoftBank and Kioxia
AI-related stocks, including SoftBank and Kioxia, fell sharply on September 14th after CEOs of leading US AI model developers warned about the need to slow development for safety r
🧩 Pattern This Issue
- Japan: Murata’s iPaS targets vertical power delivery for next-generation AI chips
- Korea/Taiwan: SEMCO and Qualcomm codevelop organic bridge packaging for advanced chipsets
- Korea/Taiwan: Samsung and CXMT drive DRAM market growth to $154.7 billion
While Western markets obsess over software deceleration, East Asian hardware giants are aggressively re-engineering the physical packaging, power delivery, and memory architectures required for the next hardware generation, ensuring they remain the indispensable foundation of the global AI supply chain regardless of model-level hype cycles.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 Policy & Regulation
Anthropic Details AI Misuse Cases, Including Biological Weapon Development
Anthropic published a threat intelligence report on September 10 detailing specific cases of its Claude AI being misused. The report highlighted instances where advanced AI capabilities, particularly in scientific knowledge, were exploited for dual-use research such as enhancing pathogen virulence or developing biological weapons, marking the first time a private AI company has disclosed such concrete examples. Misusers attempted to bypass safety filters by disguising their intentions as legitimate research or submitting ambiguous prompts, automatically resending to less restrictive models if initially rejected.
Why it matters: Japanese industrial readers, often focused on practical applications and safety, will see this as confirmation that leading-edge AI carries significant, immediate risks. The emphasis on dual-use technology and the sophisticated methods used to bypass safety filters underscores how difficult it will be to implement effective AI governance, especially when dealing with expert-level malicious actors. This isn’t just about general AI ethics; it’s about specific, verifiable threats.
For Western readers: Western policymakers and AI developers should adjust their threat models, recognizing that sophisticated actors are already actively testing and bypassing AI safety measures for highly dangerous applications, particularly in biology. Assume that current safety filters are insufficient against determined, expert-level misuse and focus regulatory efforts on pre-emptive supply chain controls for critical AI capabilities.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 AI & Machine Learning
Upstage Visits NVIDIA Headquarters to Discuss GPU Support and Five Other Agenda Items
Korean AI startup Upstage recently visited NVIDIA’s headquarters to discuss enhanced collaboration, including securing NVIDIA GPU support. The discussions covered five main agenda items aimed at strengthening their partnership in AI development and infrastructure.
Why it matters: Upstage is a recognized force in the Korean LLM space, and their direct engagement with NVIDIA signals that even well-funded regional players cannot take GPU access for granted. Securing this support is a necessary condition for them to compete effectively against larger domestic and international models, particularly in Korea’s competitive enterprise AI market.
For Western readers: If you are a Western AI startup, assume that securing reliable access to NVIDIA’s latest GPUs will require direct, strategic engagement with NVIDIA or its key partners, rather than simply relying on market availability.
🇰🇷 Semiconductors & Hardware
August ICT Exports Near $60 Billion, Semiconductors Lead with $40 Billion
South Korea’s Information and Communications Technology (ICT) exports for August are on track to reach nearly $60 billion, primarily driven by strong semiconductor sales, which accounted for approximately $40 billion of the total. This surge indicates a robust recovery in the semiconductor market and a significant contribution to the nation’s overall export performance.
Why it matters: The export data reinforces the ongoing recovery in the memory semiconductor market, which directly benefits major Korean players like Samsung Electronics and SK hynix. A strong rebound in demand, particularly for high-end memory, translates to improved financial performance and potential increases in capital expenditure, which ripple through the global supply chain.
For Western readers: Western companies relying on Korean memory semiconductors for their products should factor these strong export numbers into their inventory and procurement planning, anticipating potentially tighter supply and firmer pricing for the next 6-12 months.
🇰🇷 Semiconductors & Hardware
Samsung Electro-Mechanics and Qualcomm Partner on 2.1D Organic Bridge Packaging Development
Samsung Electro-Mechanics (SEMCO) and Qualcomm are collaborating to jointly develop ‘Organic Bridge‘ packaging technology, a 2.1D solution. This partnership aims to enhance advanced packaging capabilities, particularly for high-performance computing (HPC) and AI chips, leveraging SEMCO’s substrate expertise and Qualcomm’s chip design leadership.
Why it matters: This partnership signals a shift in advanced packaging, moving beyond the traditional 2.5D interposer to a 2.1D ‘Organic Bridge’ approach that offers cost and performance advantages, potentially disrupting the packaging market. For SEMCO, it diversifies their customer base beyond Samsung Electronics, strengthening their independent market position. For Qualcomm, it is about securing access to leading-edge packaging technology that is critical for their high-performance mobile and AI chips, especially as packaging becomes a key differentiator in chip performance.
For Western readers: Western chip designers and foundries should assess the viability and potential advantages of 2.1D Organic Bridge technology as an alternative to established 2.5D solutions, particularly regarding cost-efficiency and performance for AI accelerators and HPC applications. This move by Qualcomm suggests a push for more modular and cost-effective advanced packaging solutions for future designs.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 AI & Machine Learning
OpenAI’s GPT-6 Astra Released to ChatGPT Work, Codex, and API
OpenAI has officially released its latest large language model, GPT-6 Astra, for enterprise and developer use through ChatGPT Work, Codex, and its API. This model is notable for being the first to meet a ‘Critical’ cybersecurity threshold under OpenAI’s Preparedness Framework, and it is positioned for professional tasks, software development, and cybersecurity applications. OpenAI has also unveiled new enterprise plugins for ChatGPT Desktop, including integrations with Oracle Analytics and Power BI.
Why it matters: OpenAI is signaling its intent to capture a larger share of the enterprise AI market by focusing on practical application, data security, and integration with existing business software. The emphasis on a ‘Critical’ cybersecurity rating and better adherence to corporate directives addresses the real-world friction points that often slow down AI adoption in conservative large organizations.
For Western readers: Western businesses considering large-scale AI deployment should evaluate Astra’s security claims and its integration capabilities with their existing enterprise software, as it directly impacts data governance and operational efficiency within their stack.
🇹🇼 AI & Machine Learning
SpaceX Aims to Launch Nvidia AI Data Centers into Orbit by 2027
SpaceX CEO Elon Musk announced high confidence in launching Nvidia Vera Rubin NVL72 AI supercomputers into low-Earth orbit by 2027, as part of its ambitious ‘Starmind’ project. This initiative aims to establish an orbital AI computing infrastructure, with a customized satellite planned for Q4 2027 and scaled operations by 2028, significantly advancing earlier timelines.
Why it matters: Musk’s announcement, while characteristic in its ambition, marks a concrete step towards shifting high-performance computing infrastructure off-planet. This isn’t just a satellite; it’s an orbital data center using the same top-tier Nvidia GPUs that currently drive on-Earth AI development. The implications for latency-sensitive applications and potential military applications, free from terrestrial physical and regulatory constraints, are substantial. It also creates another demand vector for Nvidia’s most advanced chips, potentially further tightening supply.
For Western readers: Western AI infrastructure providers and defense contractors should recognize this as a move towards decentralized, space-based compute resources, potentially altering future data processing and sovereign AI strategies. Keep an eye on how existing ground station networks and satellite communication infrastructure adapt to support or compete with this orbital compute model.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 AI & Machine Learning
Chinese Physical AI Breaks Through: PhysBrain 1.5 Tops Global Open-Source Ranking, Spatial Intelligence Rivals GPT-6 Astra
📊 Featured Chart
28 public benchmarks
Chinese company DeepCybo (深度机智) has released its physical foundation model, PhysBrain 1.5, which achieved a comprehensive score of 72.5 across 28 benchmarks, leading all open-source models. The model’s performance in spatial intelligence and embodied cognition has narrowed the gap to within one point of top closed-source models like OpenAI’s GPT-6 Astra (73.3) and Google DeepMind’s Gemini 3.6 Flash (73.0). DeepCybo has open-sourced both 2B and 8B parameter versions, emphasizing a strategy focused on foundational physical intelligence rather than just scale.
Why it matters: The core problem of physical AI is not raw compute or parameter counts; it’s the ability to translate high-level understanding into precise, real-world physical actions. DeepCybo’s focus on foundational physical intelligence and its demonstrated performance in millimeter-level tasks, rather than just raw scale, shows a pragmatic approach. This indicates that China is not just playing catch-up in AI, but developing specialized solutions that address critical bottlenecks in embodied AI.
For Western readers: Western robotics and AI companies should recognize that China is developing highly competitive, specialized physical AI models. If these open-source models gain traction, they could become a viable alternative for developers seeking to build out embodied AI applications, potentially shifting some R&D and deployment away from Western-led ecosystems.
🇨🇳 Startups & Funding
Anthropic Reportedly Accelerating IPO Plans, Aiming to Match SpaceX’s Fundraising Record
📊 Featured Chart
Source: QbitAI. All figures approximate.
Despite its co-founder Dario Amodei publicly advocating for AI deceleration, Anthropic is reportedly pushing ahead with plans for a Nasdaq IPO as early as October, aiming to raise at least $86.3 billion, matching or exceeding SpaceX’s record-setting offering this year. The company’s valuation has surged dramatically, reaching $965 billion after its H-round funding in May 2026, and its annualized recurring revenue (ARR) has grown sevenfold since late last year to over $65 billion.
Why it matters: The reported scale of Anthropic’s IPO, if realized, would set a new benchmark for AI company valuations and funding, potentially eclipsing rivals like OpenAI and further fueling the capital markets’ enthusiasm for AI. This aggressive move suggests a scramble for resources to secure compute and talent, which are becoming the primary constraints for AI development, rather than a genuine concern for decelerating the field.
For Western readers: Western investors and AI companies should recognize that the ‘AI safety’ narrative can sometimes be used as a strategic tool in competitive fundraising and market positioning, rather than a sole indicator of a company’s actual operational intent. This event confirms that the race for AI dominance remains primarily a race for capital and compute, regardless of public statements.
Robotics & Automation
Unitree Opens UnifoLM-WLA-1.0 Humanoid Foundation Model Project Page
Chinese robotics company Unitree Robotics has launched a project page for its UnifoLM-WLA-1.0 humanoid foundation model, designed to integrate large language models with robotic control. The model aims to enhance perception, decision-making, and control for general-purpose humanoid robots, leveraging deep learning for motion planning and task execution.
Why it matters: Unitree’s move to open-source aspects of its humanoid foundation model, even if just a project page initially, serves to attract broader developer engagement and accelerate iteration cycles within China’s AI and robotics ecosystem. This contrasts with more closed, proprietary development seen in some Western counterparts and could foster rapid domestic skill development.
For Western readers: Western robotics firms and researchers should pay attention to how quickly Chinese efforts like Unitree’s transition from project announcements and benchmarks to deployed, scalable hardware and software. A robust Chinese open-source robotics ecosystem could rapidly advance capabilities without direct access to Western proprietary systems.
🔺 The Prism
Where US and East Asian technology interests intersect
AI & Machine Learning
Deictic Ambiguity Challenges in LLM Draft-Verify-Revise Pipelines
📊 Featured Chart
Source: arXiv:2609.12162
A new research paper explores how Large Language Models (LLMs) in draft-verify-revise pipelines struggle with ‘deictic shifts,’ where context-dependent expressions like ‘previous’ are misinterpreted across stages. The study, testing models like GPT-5.2 and Gemini 3 Pro, found significant variations in accuracy, with Gemini 3 Pro performing notably better at lower computational cost in resolving these ambiguities.
Why it matters: The findings highlight a fundamental reliability challenge for LLM orchestrations that are widely adopted in East Asian AI development. Poor performance in handling deictic shifts means deployed AI systems can produce inconsistent or incorrect outputs, impacting the trustworthiness of automated processes and enterprise AI applications from Tokyo to Shenzhen.
For Western readers: Western businesses adopting or developing advanced LLM-powered applications should pressure their East Asian partners and suppliers to transparently address deictic ambiguity in multi-stage AI systems, particularly when sourcing solutions from China, Korea, or Japan, where such pipeline architectures are common.
AI & Machine Learning
Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work
Chinese researchers have introduced Occamy-1.0, a 35B-parameter language model developed by further training the Qwen3.6-35B-A3B checkpoint. This model is designed for cost-efficient co-work applications, excelling in complex workflows involving information gathering, tool use, coding, and file manipulation.
Why it matters: Occamy-1.0’s emphasis on efficiency and ‘co-work’ capabilities points to a strategic push within China to develop AI that is immediately applicable in real-world business settings, particularly where cost and latency are major concerns. This contrasts with the Western narrative’s heavier focus on raw benchmark scores for foundational models.
For Western readers: Western businesses should recognize that Chinese AI development is not just about scaling up parameter counts; it is increasingly about optimizing models for specific enterprise workflows to deliver practical, cost-effective solutions in areas like automated agentic task execution.
AI & Machine Learning
Huawei rolls out AI-based tourism promotion service for Xi’an region
Huawei has launched an AI-powered tourism promotion service for China’s Shaanxi province, home to Xi’an, leveraging its BoGuan large language model (LLM) to generate content and itineraries based on regional history. This initiative aims to attract more visitors by offering tailored tourism experiences. The service uses a proprietary LLM trained specifically on local historical data to enhance its relevance and accuracy.
Why it matters: Huawei is finding commercially viable applications for its large language models within China’s domestic market, moving beyond just raw compute and foundational model development to monetize its AI capabilities. This is less about tourism and more about proving that a sanctioned tech giant can still build a robust AI business by focusing on specialized, government-linked enterprise services.
For Western readers: Western businesses and investors should recognize that Huawei is strategically deploying its AI capabilities into specific industry verticals within China, creating a formidable domestic competitor across a range of enterprise applications where it might otherwise be overlooked in broader AI competition narratives. Do not assume Huawei’s AI is solely focused on hardware or generic chatbots; it’s about practical integration into the national economy.