East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- LG CNS’s deployment of liquid cooling at its Samsong data center reflects a broader infrastructure pivot driven by Gartner’s projection that the global semiconductor market will nearly double to $1.555 trillion by 2026.
- The extreme shortage and rising prices of high-capacity SSDs will squeeze margins for enterprise AI buyers just as Anthropic seeks a $2 trillion valuation in its planned September IPO.
- Perplexity’s collaboration with NVIDIA on local agents and Intel’s debut of its 480GB LPDDR5X Crescent Island GPU demonstrate that the physical bottleneck of agentic AI is rapidly shifting from centralized cloud data centers to high-performance edge hardware.
This Issue’s Analysis
The Signal
$837B Memory Market Shift: Japan’s Chip Equipment Giants Pivot to AI Bandwidth
Gartner predicts the global semiconductor market will grow 92% year-over-year to $1.555 trillion in 2026, driven by continued AI infrastructure investment and rising memory prices.
Semiconductors & Hardware
SSD Prices to Quadruple by 2026: Why Enterprise Storage Budgets Face an Unprecedented NAND Pinch
The storage market is facing extreme disruption due to a severe shortage and rapid price increases for SSDs. Market research firm Techno Systems Research (TSR) revised its 2026 for
Semiconductors & Hardware
Arm’s AGI CPU Architecture: Why Agent Workloads Are Driving a $2B Infrastructure Market
Arm has unveiled details of its AGI CPU architecture at the Hot Chips 2026 conference, designed specifically for AI agent workloads like tool calling and workflow orchestration. Ba
Semiconductors & Hardware
Nvidia’s $750B AI Financing Risk Triggers Record Rise in Bond Market Credit Swaps
Nvidia is set to announce its Q2 earnings, with projected revenue doubling from 2025 to $92.18 billion, and an estimated $104.2 billion for Q3. However, market attention has shifte
🧩 Pattern This Issue
- Japan: Gartner forecasts global chip sales hit $1.6T by 2026 as memory takes 54% share
- Korea: LG CNS deploys liquid cooling at Samsong data center to handle AI density
- Taiwan/US: Intel and Arm detail new high-bandwidth, local inference chips at Hot Chips
The semiconductor industry is aggressively shifting from raw compute scaling to memory-centric and localized inference architectures, exposing Western chip designers who fail to secure advanced packaging and thermal management supply chains in East Asia.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 Semiconductors & Hardware
Intel Details New ‘Crescent Island’ GPU with Up to 480GB LPDDR5X for Agentic AI Inference
Intel announced its next-generation GPU, code-named ‘Crescent Island,’ at Hot Chips 2026, designed specifically for agentic AI inference. The GPU, featuring the Xe3P architecture, prioritizes memory capacity over speed by adopting LPDDR5X memory, allowing for up to 480GB on custom designs while maintaining a 350W air-cooled PCIe card form factor. This approach addresses the growing memory demands of complex multi-step AI workloads and large models, aiming for higher overall throughput and efficiency.
Why it matters: Intel is explicitly targeting inference rather than training, a segment that often gets less attention than frontier model development but is ultimately where most AI value is captured. Their decision to prioritize LPDDR5X capacity over HBM bandwidth suggests a pragmatic view of cost and power efficiency for enterprise deployments, rather than solely chasing benchmark leads.
For Western readers: Western enterprises planning large-scale AI inference deployments should evaluate Intel’s Crescent Island for its potential to lower TCO by accommodating larger models or more concurrent sessions per GPU, especially if their workloads are memory-capacity bound rather than bandwidth-bound.
🗾 Startups & Funding
Anthropic Targets Record-Breaking $2 Trillion IPO, Eyes September Listing
US AI developer Anthropic is reportedly planning an initial public offering (IPO) as early as September, aiming for a valuation of up to $2 trillion, which would make it the largest IPO in history. The company privately submitted its S-1 prospectus in June and is currently in final discussions for a listing, with funds intended to enhance its generative AI model, Claude, and secure computing resources. Anthropic has outpaced rival OpenAI in B2B revenue growth and recently achieved its first operating profit.
Why it matters: Anthropic’s potential record-setting IPO signifies a major shift in market valuation for AI companies, rewarding its B2B focus and profitability over OpenAI’s consumer-oriented strategy. This could encourage other AI startups to prioritize enterprise solutions and demonstrate a clear path to profitability, rather than solely focusing on user growth, setting a new standard for AI investment.
For Western readers: Western investors and tech executives should now assume that a clear path to B2B profitability, even with a later market entry, can lead to a significantly higher valuation than consumer-focused AI models, potentially shifting investment strategies and partnership considerations towards enterprise AI solutions.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 Enterprise & Cloud
LG CNS Adopts Liquid Cooling Infrastructure at Samsong Data Center
LG CNS, LG Group’s IT services arm, has introduced liquid cooling technology at its Samsong data center. This move aims to enhance the center’s capacity to handle high-performance computing, particularly for AI workloads, by improving power efficiency and cooling capabilities.
Why it matters: The shift to liquid cooling by LG CNS is a direct response to the escalating power and thermal demands of AI servers. This investment indicates that large Korean enterprises are already building out the physical infrastructure necessary to support their own AI development and deployment, which means more demand for advanced cooling solutions in the region.
For Western readers: Western providers of liquid cooling technology and data center infrastructure should recognize that Korean chaebols are moving quickly to implement these solutions, creating a new market for specialized hardware and services. Assume that this type of upgrade will become standard practice across major data centers in East Asia within the next 24 months, not just for hyperscalers.
🇰🇷 AI & Machine Learning
Perplexity Unveils ‘Portable Computer’ Local Agent in Collaboration with NVIDIA
Perplexity, known for its conversational AI search engine, announced a collaboration with NVIDIA to launch ‘Portable Computer,’ a local AI agent designed for real-time, on-device information processing. This initiative aims to address data sovereignty and latency concerns by running AI models directly on user devices, enhancing privacy and efficiency.
Why it matters: This initiative represents Perplexity’s entry into the burgeoning local AI market, leveraging NVIDIA’s hardware expertise to move beyond cloud-based search. It reflects a strategic bet that a segment of users will prioritize data privacy and real-time processing over purely cloud-dependent AI, signaling a potential new vector for competition in AI services.
For Western readers: Western businesses should watch how the market responds to local AI agents like ‘Portable Computer’; if adoption is strong, it indicates a significant user appetite for on-device processing and could necessitate shifts in AI product development and data strategy to align with privacy and real-time demands.
🇰🇷 AI & Machine Learning
Kakao AI’s Inaugural Board of Directors Takes Shape, Including Former Ministers, National Tax Service Chief, and KT CTO
Kakao AI, the new entity spun off from Kakao’s AI division, is establishing its first board of directors, drawing high-profile talent from government and industry. The provisional board includes a former minister, a former commissioner of the National Tax Service, and the former CTO of KT, signaling a strategic focus on strong governance and industry expertise for the new AI venture.
Why it matters: Kakao is positioning Kakao AI not just as a technology startup but as a nationally relevant enterprise, leveraging the credibility and network of former senior government officials and seasoned corporate executives. This is typical chaebol behavior: when a new business line is deemed strategically important, it gets political and industry heft from the start, a practice aimed at smooth regulatory navigation and securing national champion status.
For Western readers: Western firms looking to partner with or compete against South Korean AI entities should recognize that top-tier local players like Kakao AI will likely have deep ties into government and established industries, not just venture capital. This structure allows them to exert greater influence on local policy and secure large-scale domestic contracts more readily than a pure-tech startup might.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 AI & Machine Learning
Google’s Gemma Open AI Model Exceeds 1 Billion Downloads, Spawning Over 100,000 Derivative Models
Google announced that its open AI model family, Gemma, has surpassed 1 billion cumulative downloads and has led to over 100,000 derivative models developed by the community over the past two years, forming a ‘Gemmaverse’ ecosystem. Unlike Gemini, which is primarily cloud-accessed, Gemma is designed for local, edge, and cloud deployment, fostering diverse applications in fields such as medical research, satellite imagery analysis, and animal communication studies.
Why it matters: Google’s push with Gemma, especially in enabling a vast ecosystem of fine-tuned models and edge deployments, directly competes with the closed-source, API-driven model offerings that are common among Western AI providers. This strategy, backed by significant community adoption, positions Google to capture developer mindshare in regions sensitive to data localization and seeking more control over their AI infrastructure, particularly in industrial and public sector applications.
For Western readers: Western businesses and developers reliant on cloud-centric, proprietary AI models should recognize that Google’s open-source Gemma initiative is gaining significant traction, especially for edge and specialized applications, which could shift the competitive landscape for AI tools and talent.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 Robotics & Automation2 STORIES
A Ten-Minute Breakthrough in Embodied AI and Universal Multi-Robot Coordination
Recent breakthroughs are pushing the limits of embodied AI, highlighted by Skild AI’s S1 model enabling robots to learn complex 10-minute tasks from a single video demonstration, and a continuous Chinese demo showing Unitree and LimX Dynamics robots collaborating seamlessly under a single shared brain. Collectively, these developments demonstrate a massive leap forward in long-sequence task execution, zero-shot generalization, and hardware-agnostic robot intelligence.
Why it matters: In the East Asian tech ecosystem, this shift toward hardware-agnostic, unified AI brains allows nimble Chinese robotics manufacturers to bypass proprietary software bottlenecks and rapidly scale cooperative, multi-agent industrial automation.
For Western readers: Western tech leaders must abandon the assumption that robotics software must be custom-built for specific hardware; instead, prepare for a market dominated by universal AI ‘brains’ capable of operating diverse, commoditized Asian robotic fleets out of the box.
Semiconductors & Hardware
NAND maker YMTC’s RMB33 billion Shanghai IPO application accepted
China’s leading NAND flash memory producer, Yangtze Memory Technologies Holdings (YMTC), has had its application to list on Shanghai’s STAR Market accepted. The company aims to raise RMB33 billion (US$4.9 billion) to fund production-line upgrades and R&D, positioning it as one of the market’s largest offerings.
Why it matters: YMTC’s planned RMB33 billion IPO is a direct measure of China’s commitment to building out its domestic semiconductor capabilities, particularly in memory, despite external pressures. Access to this capital will enable YMTC to invest heavily in its manufacturing processes and product development, directly challenging the dominance of non-Chinese memory manufacturers.
For Western readers: Western memory suppliers and equipment manufacturers should anticipate increased competition and a continued push for domestic substitution within China, potentially impacting their market share and sales of advanced tooling in the long term.
🔺 The Prism
Where US and East Asian technology interests intersect
AI & Machine Learning2 STORIES
Engineering Reliability: East Asian Research Solves the LLM Agent Usability Crisis
Recent research from East Asian teams addresses the core hurdles of deploying AI agents in the real world, focusing on scaling agentic intelligence for complex environments and extracting structured state-machines to audit opaque agent behaviors. Together, these breakthroughs in Apodex 1.1 and trace-based automata transition LLM agents from unpredictable chatbots into highly structured, deployable enterprise tools.
Why it matters: In East Asia’s highly integrated industrial and corporate ecosystems, AI adoption hinges on zero-tolerance reliability; these technical advancements directly enable regional giants to safely integrate autonomous agents into critical manufacturing, logistics, and financial workflows.
For Western readers: Western tech leaders must abandon the assumption that raw frontier model power is enough to win the enterprise AI race; they must immediately pivot toward investing in agentic middleware and structured auditing frameworks to match East Asian operational readiness.
AI & Machine Learning
Gated Activation Steering for Reducing Sycophancy & Hallucination in Medical Question Answering
📊 Featured Chart
4B parameter model vs. 100B+ parameter model
Researchers have developed a Gated Activation Steering (GAS) framework to combat sycophancy and hallucination in Large Language Models (LLMs), particularly within clinical question answering. This method uses Inference Time Intervention (ITI) with behavior-specific gates to learn separate steering directions for these issues, applying them only when necessary to preserve correct responses. The framework significantly improved a 4-billion-parameter model’s robustness under pressure, making it comparable to models over 100 billion parameters.
Why it matters: The paper demonstrates a method to significantly enhance LLM reliability without retraining, which is crucial for East Asian firms deploying domestic models in fields like medicine. This approach allows smaller, more efficient models to achieve robustness comparable to much larger ones, which can reduce inference costs and dependency on high-end hardware, a persistent concern for many East Asian AI developers.
For Western readers: Western AI developers focused on model scaling should evaluate inference-time intervention methods as a cost-effective way to improve model safety and trustworthiness, especially when competing with or deploying alongside East Asian models in regulated industries.