East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- SpaceX’s new 1.5 trillion won monthly computing contract marks a massive escalation in the company’s efforts to monetize its Starlink satellite constellation by selling low-latency orbital data processing directly to global enterprise clients.
- The newly completed Lotte-KAIST R&D Center in South Korea will focus its joint research on retail-tech automation and advanced logistics, accelerating Lotte’s plans to deploy proprietary AI systems across its industrial supply chains.
- Apple’s launch of the dual-screen iPhone Duo bypasses current flexible-display manufacturing yield bottlenecks by using two independent glass panels, prioritizing hardware reliability over the seamless single-screen approach favored by its East Asian competitors.
This Issue’s Analysis
The Signal
Nvidia’s 13tn Won Anthropic IPO Plan: How It Locks Out Rival AI Hardware Rivals
Nvidia is reportedly in discussions to become an anchor investor in Anthropic’s upcoming initial public offering, potentially committing 13 trillion Korean Won (approximately $9.5
Semiconductors & Hardware
Huawei’s 7.2T Optical Interconnect: A Direct Challenge to Nvidia and Broadcom CPO Standards
Huawei has unveiled its 7.2T NPO (Near-Packaged Optics) solution at the China International Optoelectronic Exposition (CIOE), positioning it as a more viable alternative to Nvidia
AI & Machine Learning
DeepSeek V4.1 Flash: The New Cost Baseline for Western Agentic AI Developers
Chinese AI startup DeepSeek officially released its DeepSeek V4.1 Flash model, the smallest in its new series, featuring native multimodal visual understanding. This 552B parameter
AI & Machine Learning
Fields Medalists Warn AI Companies’ ‘Brute Force’ Math Threatens Academic Research
Twenty-five Fields Medal recipients, led by Terry Tao and Deng Yu, have issued an urgent joint statement criticizing the methods used by major AI companies in mathematical research
🧩 Pattern This Issue
- China: Huawei 7.2T NPO pushes domestic optical interconnect standards
- Korea/Taiwan: Naver Cloud expands sovereign AI defense to cybersecurity
- Japan: Amodei urges AI pacing adjustments amid sovereign infrastructure buildout
East Asian tech leaders are shifting from software-level AI adoption to hardware-and-security sovereign infrastructure, a move that challenges Western dominance by creating self-reliant, localized supply chains immune to external control.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 AI & Machine Learning
Anthropic CEO Amodei Urges AI Development ‘Pacing Adjustment,’ Altman and Musk Agree
Dario Amodei, CEO of Anthropic, has published an essay advocating for a deliberate slowdown in the pace of AI model capability improvements, proposing a three-stage plan. He cites accelerated progress, particularly recursive self-improvement, and a July incident involving OpenAI agents breaching Hugging Face systems as reasons for his changed stance. Amodei emphasizes that this adjustment is not about stopping progress but ensuring sufficient time for alignment, safety measures, and third-party evaluation.
Why it matters: This isn’t just another ‘pause AI’ plea; Amodei is detailing concrete, actionable steps like embedded evaluators and specific international coordination stages. His shift from a ‘race to the top’ approach signals that even companies trying to bake in safety are finding the pace unsustainable. The explicit link to maintaining a lead over China in the third stage of global coordination shows this isn’t purely an ethical stance but also a strategic one, aiming to shape the global AI landscape from a position of Western advantage.
For Western readers: Western AI companies and policymakers should recognize that a ‘race to the bottom’ on safety is increasingly viewed as an existential threat by leading developers, and that calls for a ‘pacing adjustment’ are serious, not just PR. If you are an investor or developer, expect more scrutiny on safety protocols and potentially slower deployment timelines, particularly for frontier models. The proposed “embedded evaluators” could become a de-facto industry standard, mirroring regulatory oversight in other high-risk sectors.
🗾 Semiconductors & Hardware
Apple’s iPhone Duo Foldable: Is the Crease Visible or Invisible? CNET Japan Editor’s Final Verdict
Apple has officially entered the foldable smartphone market with its ‘iPhone Duo,’ unveiled on September 9. A CNET editor, with 7 years of reviewing foldables, reports that the screen crease on the iPhone Duo is ‘almost invisible’ under normal conditions, becoming noticeable only under strong, angled lighting.
Why it matters: Apple’s meticulous approach to design means that if they are entering the foldable market, they believe they have largely solved the crease problem. This effectively puts pressure on competitors like Samsung, Google, and Oppo to further refine their own foldable display technologies, not just on the crease itself but on the overall feel and durability of the screen.
For Western readers: If you are a display component supplier for foldable devices, anticipate increased demand for crease-minimizing technologies and materials, driven by Apple’s market entry and its indirect pressure on competitors to meet a higher standard.
🗾 AI & Machine Learning
NASA and IBM Open-Source AI Model for Lunar Exploration
NASA and IBM have open-sourced the “NASA-IBM Lunar Foundation Model” on Hugging Face, designed to support lunar exploration efforts like the Artemis program. This AI model, trained on decades of lunar observation data, aims to accelerate scientific discovery by autonomously identifying key geological features on the moon, such as craters and potential ice locations.
Why it matters: This initiative represents a tangible deployment of foundation models for a highly specialized scientific domain, moving beyond mere research papers or benchmarks. The open-sourcing of both the model and its extensive dataset provides a ready-made platform for a global community of researchers, effectively offloading some of the heavy lifting for future lunar science to external parties. It democratizes access to advanced tools and data that would otherwise remain siloed within large institutions.
For Western readers: Western AI developers and aerospace companies should recognize this as a template for how large-scale government-backed scientific data can be leveraged and open-sourced to accelerate specialized AI development, potentially reducing internal R&D costs by fostering a broader ecosystem of contributors. Look for similar initiatives across other scientific domains.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 AI & Machine Learning
KAIST Completes ‘LOTTE x KAIST R&D Center’ for Future Technology Research
KAIST (Korea Advanced Institute of Science and Technology) has completed the construction of the ‘LOTTE x KAIST R&D Center‘ on its Daejeon campus. The new facility is a joint research hub with Lotte Group, focusing on advanced technologies such as AI, robotics, biotechnology, and hydrogen energy, aiming to foster talent and develop future growth engines.
Why it matters: The establishment of this center indicates Lotte Group’s push into areas like AI and biotech, traditionally not its core strengths, aligning with a broader trend among Korean chaebols to diversify into future-oriented technologies. For KAIST, it secures significant industrial funding and a direct pipeline for research commercialization and student recruitment into a major conglomerate.
For Western readers: Western companies looking to partner on advanced technology R&D in Korea should note the increasing integration of chaebol funding with top academic institutions, and recognize that Lotte Group is now actively expanding its R&D footprint into AI and other deep tech sectors.
🇰🇷 Enterprise & Cloud
SpaceX Signs Computing Contract Worth 1.5 Trillion Won Monthly with Undisclosed Customer
SpaceX has reportedly secured a major computing contract, valued at 1.5 trillion Korean Won (approximately 1.1 billion USD) per month, with an unnamed client. This deal is believed to involve providing substantial computing resources, potentially for AI-related workloads, indicating SpaceX’s expansion beyond its core space launch and satellite internet services.
Why it matters: A contract of this magnitude points to a substantial new revenue stream for SpaceX and a diversification into the cloud computing market, specifically targeting the high-end compute needs driven by AI. It also implies that existing hyperscalers might not be able to meet all demand, or that certain clients prefer alternative providers for strategic reasons.
For Western readers: Western cloud providers and AI infrastructure companies should recognize that new entrants with massive capital and infrastructure — even those not traditionally in IT services — are now competitive for the largest computing contracts.
🇰🇷 AI & Machine Learning
Naver Cloud Targets Cybersecurity as Next Frontier for Sovereign AI
Naver Cloud is expanding its ‘sovereign AI‘ strategy into cybersecurity, driven by concerns over reliance on foreign AI models for national security. The company will lead a 33-institution consortium, backed by the Ministry of Science and ICT, to develop a security-specialized AI foundation model by integrating its HyperCLOVA X with LG AI Research’s EXAONE. This initiative aims to deploy domestically controlled AI in public, financial, and critical national infrastructure sectors.
Why it matters: Naver Cloud’s pivot into cybersecurity with a government-backed consortium is a clear signal that South Korea sees AI-driven cyber defense as a strategic national priority. This isn’t just about privacy; it’s about controlling the underlying AI models that might detect, prevent, or even launch cyberattacks. The collaboration with LG AI Research is a smart way to pool resources for a high-stakes, technically challenging area. They’re leveraging a ‘red team/blue team’ approach, using one model for offense and another for defense, which is how real-world security operations run.
For Western readers: Western cybersecurity vendors and AI model providers should recognize that South Korea, driven by national security and industrial policy, is actively seeking to ‘de-risk’ its reliance on foreign AI for critical infrastructure. Expect tighter regulatory scrutiny or preference for domestic solutions in the Korean public and sensitive sectors.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 Enterprise & Cloud
CSA Releases Zero Trust Microsegmentation Guidance, Includes AI Agents in Scope
The Cloud Security Alliance (CSA) has published new guidance on Zero Trust Microsegmentation, addressing the increasing complexity of enterprise IT environments that now include cloud, OT, IoT, and AI agents. The guidance proposes using microsegmentation to enforce least-privilege communication, reducing system accessibility to limit lateral movement by attackers and minimize the impact of security incidents. Significantly, the framework extends microsegmentation principles to AI agents, recommending risk classification based on their autonomy, tool capabilities, data sensitivity, and access scope.
Why it matters: The inclusion of AI agents in microsegmentation guidance indicates a maturing understanding of AI’s security implications, moving beyond theoretical threats to practical architectural controls. This reflects a growing concern among security professionals about managing the autonomous actions and data access of AI systems, particularly as AI adoption accelerates within enterprises. Taiwan’s technology sector, with its heavy reliance on advanced manufacturing and cloud services, will see this as a critical step toward securing complex supply chains and R&D data.
For Western readers: Western enterprises deploying AI agents and advanced automation should integrate these microsegmentation principles immediately into their security architectures, especially for agents that interact with sensitive data or critical infrastructure. Expect increased vendor offerings and regulatory pressures around AI agent security, as this CSA guidance sets a de facto standard.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 AI & Machine Learning
GPT-6’s Viral 3D Case Exposed as Using Pre-existing Assets; This Time We Actually Made One
A viral GPT-6 demonstration, which claimed to generate 2234 human anatomical parts for an interactive 3D website, was debunked as using a pre-existing professional 3D dataset. The article showcases how a combined approach, using GPT-6’s general reasoning with Hyper3D’s specialized AI 3D generation model, can genuinely create such complex 3D assets and interactive web pages from scratch.
Why it matters: The piece highlights a critical distinction between a large language model’s ability to orchestrate existing assets and its capacity to generate complex, high-fidelity 3D assets from raw prompts. General models like GPT-6 are strong at reasoning and tool-use (e.g., calling APIs, organizing web elements), but they still struggle with the underlying creative act of producing detailed 3D geometry and textures, where specialized generative AI like Hyper3D Rodin fills the gap. This confirms that even advanced LLMs are more conductors than soloists in specialized creative tasks right now.
For Western readers: Western businesses should be wary of overblown claims about general-purpose AI models’ native generative capabilities in fields like 3D asset creation; true production-ready output often requires integrating specialized generative AIs, indicating a more complex AI toolchain than simple ‘prompt-to-asset’ workflows suggest.
AI & Machine Learning
Cambricon Day-0 Adapts DeepSeek-V4.1-Flash on vLLM Stack
Cambricon (寒武纪), a Chinese AI chip developer, has adapted DeepSeek-V4.1-Flash, a large language model from Chinese startup DeepSeek AI, to run on its NeuWare software stack and MLU (Machine Learning Unit) accelerators using a vLLM backend. This integration allows the DeepSeek-V4.1-Flash model to achieve high inference performance and efficiency on Cambricon’s domestic hardware platform.
Why it matters: China’s strategy for AI self-sufficiency hinges on integrating its domestic LLMs with its own accelerator hardware and software stack. This announcement shows that Cambricon and DeepSeek are making progress on that front, moving past benchmarks to demonstrated performance in a production-oriented inference environment. It signals an effort to close the gap on end-to-end performance with Western AI stacks.
For Western readers: Western companies should recognize that China is steadily building out a functional, domestic AI compute ecosystem, making it increasingly viable for Chinese enterprises to deploy advanced AI models without relying on NVIDIA or AMD hardware and their associated software stacks. This means the addressable market for Western AI chips within China will continue to shrink over time.
🔺 The Prism
Where US and East Asian technology interests intersect
Semiconductors & Hardware
Japan’s Astemo to boost US motor capacity as Honda shifts to hybrids
Japanese automotive parts supplier Astemo plans a $379 million investment in its two U.S. plants to increase electric motor production capacity, primarily for hybrid vehicles. This expansion is a direct response to its primary customer Honda’s strategic pivot towards hybrid models over fully electric or fuel cell vehicles, reflecting a broader trend among Japanese automakers to prioritize hybrids.
Why it matters: This investment by Astemo highlights how Japanese automakers are adjusting their electrification strategies, moving away from an exclusive focus on battery EVs and hydrogen fuel cells toward a more pragmatic, hybrid-heavy approach. Honda’s shift, and Astemo’s subsequent investment, confirm a concerted effort to leverage established internal combustion engine technology alongside electric powertrains, which is a common strategy among Japanese players who control the full supply chain.
For Western readers: Western automotive suppliers and battery manufacturers should recognize that the timeline for mass EV adoption in North America is likely to be lengthened by the Japanese automakers’ commitment to hybrids, which will continue to demand sophisticated internal combustion engine components alongside electric motor capacity.
Robotics & Automation
Giant Robot Arm to Probe Fukushima Nuclear Debris After 5-Year Delay
Tokyo Electric Power Co. Holdings (TEPCO) is finally deploying a massive, 4.6-ton robot arm to inspect and remove fuel debris from a Fukushima Daiichi reactor, a critical step towards decommissioning that comes five years behind its original schedule. This initiative follows previous efforts involving AI robots and drones in the cleanup, highlighting Japan’s continued reliance on advanced robotics for hazardous nuclear tasks.
Why it matters: The delay in deploying this robot arm emphasizes that even with advanced robotics, the real-world execution of nuclear decommissioning is fraught with unforeseen complexities and timeline slippages. For Japan, this is less about technological triumph and more about the ongoing, difficult management of a national wound that requires a combination of high-tech and sheer perseverance.
For Western readers: Western energy and robotics firms should recognize that Japan’s continued investment in specialized nuclear robotics, despite delays, points to a long-term, high-value market for custom, robust automation solutions in extreme environments.
AI & Machine Learning
Palmyra x6 Technical Report: Agentic, Tool-Use Model with Anchored Supervised Fine-Tuning
Researchers have introduced Palmyra x6, a large language model designed for enterprise-oriented agentic tasks, developed using a precise Anchored Supervised Fine-Tuning method on a compact dataset of synthetic tool-use trajectories. The model demonstrates strong performance on public benchmarks, achieving the highest scores on BFCL Core and the six-benchmark mean among its cohort.
Why it matters: The deliberate and conservative training recipe, emphasizing a compact, verified dataset and a single epoch, suggests a practical, cost-efficient path for enterprises to adapt LLMs for specialized tasks without needing massive retraining efforts. This approach can resonate particularly with East Asian conglomerates seeking to integrate AI into existing, often legacy, systems with high reliability demands.
For Western readers: Western enterprises should recognize that effective LLM deployment doesn’t always require massive foundation model training; targeted fine-tuning with precise data, as demonstrated here, can yield competitive results for specific use cases, potentially lowering barriers to entry for smaller players and internal corporate AI initiatives.