East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- Samsung Electronics is projected to capture 40% of the high-bandwidth memory market in the fourth quarter, positioning the South Korean giant to challenge SK Hynix’s dominance just as it commits to adopting ASML’s advanced High-NA EUV lithography for DRAM production by 2028.
- Microsoft’s plan to triple its global data center capacity by 2032 represents an aggressive infrastructure play to bypass current hardware bottlenecks, even as OpenAI is forced to pause sign-ups for its $200 monthly subscription tier due to immediate compute constraints.
- Tokyo-based Sakana AI’s release of its “Fugu Ultra” model claims specialized performance victories over Western rivals, yet its reliance on evolutionary model-merging rather than massive proprietary training runs demonstrates Japan’s pragmatic, capital-efficient approach to the AI race.
This Issue’s Analysis
The Signal
Microsoft’s Plan to Triple Data Centres by 2032 Secures Long-Term Demand for HBM and Foundry Giants
Microsoft has announced plans to aggressively expand its data center infrastructure globally, aiming to triple its current capacity by 2032. This move is a direct response to persi
Semiconductors & Hardware
Arm Expands CSS and Total Design to Target AI Infrastructure and Robotics
Arm introduced three new solutions at its ‘Arm Everywhere China’ event in Shanghai: Arm CSS for Mobile 2, Arm Neoverse CSS N4 for AI infrastructure, and an expansion of its ‘Arm To
Semiconductors & Hardware
Samsung’s High-NA EUV DRAM Roadmap: Why It Is Accelerating Mass Production to 2028
Samsung Electronics and ASML announced an expanded strategic collaboration to introduce high-numerical aperture (NA) EUV lithography for DRAM mass production by 2028, a move they c
Semiconductors & Hardware
Samsung Targets 40% HBM Share and Foundry Profitability via Integrated AI Nodes
Analysts project Samsung Electronics’ High Bandwidth Memory (HBM) market share to approach 40% in Q4 2026, up from 33% in Q2, driven by increased HBM4 sales. Simultaneously, yield
🧩 Pattern This Issue
- Korea/Taiwan: Samsung eyes 40% Q4 HBM share while targeting 2028 high-NA EUV DRAM mass production
- China: CXMT expands DRAM wafer capacity to rival Korean memory giants
- China: Huawei and Cambricon raise AI chip prices due to domestic HBM shortages
China’s aggressive legacy DRAM expansion is forcing Korean memory giants to accelerate their transition to advanced HBM and next-generation lithography, while leaving Chinese domestic AI chipmakers highly vulnerable to localized HBM supply bottlenecks.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 AI & Machine Learning
Sakana AI Releases Latest Fugu Version, Claims to Exceed GPT-6 Astra and Fable 5.1 in Some Areas; Notes Models Not Used as Partners
Sakana AI, a Japanese startup, announced the latest versions of its Sakana Fugu AI system, ‘Fugu Ultra v2‘ and ‘Fugu Max.’ Fugu Ultra v2 reportedly surpasses OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 in certain coding benchmarks, specifically DeepSWE, while Fugu Max focuses on cost-performance and includes NVIDIA’s Nemotron series as new integration partners.
Why it matters: Sakana AI’s claim to outperform established frontier models from OpenAI and Anthropic in specific benchmarks, without using those models as integration partners for its Ultra v2 offering, represents a significant development from a Japanese startup. This positions Sakana AI as a more direct challenger in the competitive global AI model landscape, rather than merely a synthesizer of others’ models.
For Western readers: Western enterprises considering AI model choices should evaluate Sakana AI’s Fugu Ultra v2 for specialized coding tasks, especially given its claimed performance against leading US models and its cost-optimized Max version.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 AI & Machine Learning2 STORIES
OpenAI Infrastructure Strained as New Advanced Real-Time Models Roll Out
OpenAI has temporarily halted sign-ups for its high-end $200 ‘Pro’ tier due to overwhelming computational demand for its new ‘GPT-6 Astra’ model, while simultaneously releasing the API for ‘GPT-Live-1,’ a full-duplex voice model designed for real-time conversational agents. Together, these moves highlight a critical paradox where OpenAI is rapidly pushing the boundaries of real-time multimodal AI while struggling to scale the hardware infrastructure required to support it.
Why it matters: In East Asia, where hardware giants like Samsung and SK Hynix dominate the global HBM supply chain, OpenAI’s capacity bottlenecks serve as a direct demand signal that will accelerate local sovereign cloud investments and solidify the region’s leverage in AI infrastructure negotiations.
For Western readers: Discard the assumption that major AI providers can scale infinitely; you must immediately design your application architecture with multi-model redundancy to prevent service disruptions when top-tier API providers hit compute ceilings.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 Semiconductors & Hardware
China’s CXMT Accelerates Capacity Expansion, Targets Samsung and SK hynix’s 600K-700K Wafer Monthly Output
📊 Featured Chart
Current and Projected; estimates
ChangXin Memory (CXMT), China’s largest DRAM manufacturer, is initiating significant equipment investment and expansion plans for multiple new fabs. The company aims to rapidly close the production capacity gap with global leaders Samsung Electronics and SK hynix, projecting an increase from a current 250,000 wafers per month to over 600,000 wafers per month by 2028.
Why it matters: CXMT’s capacity targets are not just aspirational; they represent a concrete, state-backed push to achieve self-sufficiency in a critical semiconductor segment. This isn’t merely about market share; it’s about shifting the fundamental supply chain dynamics for a foundational component of modern computing, reducing China’s reliance on foreign suppliers.
For Western readers: Western companies relying on global DRAM supply chains should anticipate increased price volatility and potential oversupply in lower-tier DRAM products as Chinese domestic capacity ramps up, alongside heightened geopolitical pressure on equipment suppliers to China.
🇹🇼 AI & Machine Learning2 STORIES
DeepSeek Launches V4.1-Flash Model, Slashing API Costs While Raising On-Premise Demands
Chinese AI pioneer DeepSeek has released DeepSeek-V4.1-Flash, a 552-billion parameter Mixture-of-Experts model featuring a novel Causal-Encoder-Decoder architecture that slashes cached API input costs to an unprecedented $0.003 per million tokens. While the model delivers superior performance and native multimodal capabilities optimized for agentic AI, its massive scale introduces steep hardware requirements for enterprises seeking private, on-premise deployments.
Why it matters: In East Asia, where data sovereignty concerns run high, this release forces a strategic reckoning for enterprises: they must choose between ultra-cheap, public cloud-based API adoption or investing in massive local hardware clusters to run these increasingly bloated ‘lightweight’ models privately.
For Western readers: Discard the assumption that Chinese AI innovation is stalled by hardware constraints; Western developers must immediately reassess their AI agent unit economics, as DeepSeek’s aggressive pricing model renders Western API cost benchmarks obsolete.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 AI & Machine Learning
OpenAI Tackles Millennium Prize Problems as Benchmarks, Hodge Conjecture Next
QbitAI reports on the ongoing controversy surrounding OpenAI’s Navier-Stokes solution, detailing mathematician Tristan Buckmaster’s refusal of OpenAI’s collaboration terms and the ensuing public dispute. The article also reveals that OpenAI is now claiming “substantial progress” on another Millennium Prize Problem, widely rumored to be the Hodge Conjecture, signaling their intent to publish results soon.
Why it matters: OpenAI’s aggressive pursuit of Millennium Prize Problems, using thousands of agents and immense compute, accelerates scientific discovery in a way human mathematicians cannot match for speed. This challenges traditional notions of authorship and credit, particularly for foundational work that sets the stage for AI’s rapid advancements.
For Western readers: Western researchers and institutions should anticipate more disputes over AI’s role in scientific discovery and develop clear guidelines for collaboration, data use, and credit attribution with large AI labs, as the pace of AI-driven breakthroughs will continue to compress traditional research timelines.
Semiconductors & Hardware
Huawei Ascend 950DT and Cambricon Raise AI Accelerator Quotes on HBM Squeeze
Chinese AI chip manufacturers Huawei (Ascend 950DT) and Cambricon have increased their AI accelerator pricing by 10-20% due to severe shortages and rising costs of High-Bandwidth Memory (HBM). This price hike affects key components for domestic AI model training and inferencing, primarily impacting Chinese cloud service providers and AI startups.
Why it matters: The price increases for Huawei and Cambricon’s AI accelerators signal that China’s domestic AI chip ecosystem remains highly dependent on global supply chains for critical components like HBM. This undermines the narrative of rapid self-sufficiency, as even leading Chinese chip designers are vulnerable to external supply shocks, pushing up costs for their domestic customers.
For Western readers: Western businesses in the AI and cloud infrastructure space should expect persistent HBM supply tightness to continue driving up component costs globally, not just in China. This also indicates that US export controls on advanced manufacturing equipment are effectively limiting China’s ability to localize key components like HBM.
🇨🇳 AI & Machine Learning
Not Simple: Amap’s ‘Foodie Happiness Ranking’ Gets Full AI Upgrade
Amap (Gaode), Alibaba’s mapping and navigation unit, has fully integrated AI into its ‘Sweeping the Streets Ranking’ (扫街榜) for food recommendations. The 2026 version of the ranking now uses AI algorithms, including world models, to understand user navigation and store visit behaviors more precisely, aiming to identify popular and high-quality restaurants.
Why it matters: Amap is framing this as a consumer feature, but the underlying technology, ABot-Earth 0.7, represents a significant push into large-scale 3D world model development and deployment within China. This is not just about restaurant recommendations; it’s about building foundational AI infrastructure for understanding and simulating the physical world at scale, with potential applications far beyond consumer maps.
For Western readers: Western companies competing in geospatial intelligence or developing large-scale world models should note the efficiency claims and global coverage of Amap’s ABot-Earth 0.7. While the ‘foodie ranking’ is a consumer-facing application, the core technology indicates China’s progress in creating scalable, high-fidelity digital twins of the urban environment, which could reduce data acquisition costs for future smart city or autonomous driving applications globally.
🔺 The Prism
Where US and East Asian technology interests intersect
Semiconductors & Hardware
TSMC August Revenues Up 53% YoY, Q2 Profit Jumps 77%
📊 Featured Chart
Source: TSMC, Electronics Weekly
TSMC reported record monthly revenue for August 2026, reaching $16.35 billion, a 53.3% year-over-year increase, and a 10.1% increase from July. The company’s revenue for January through August totaled $107 billion, up 39.3% YoY, with Q2 profits jumping 77% and Q3 revenue projected between $44.6 billion and $45.8 billion.
Why it matters: TSMC’s strong revenue growth indicates that the demand for advanced chips, especially those used in AI, continues to outstrip supply, driving both volume and pricing. This confirms Taiwan’s central role in the global technology supply chain, strengthening its economic leverage and geopolitical importance amidst US-China technology competition.
For Western readers: Western companies relying on leading-edge semiconductors should expect sustained capacity tightness and potentially higher pricing for the foreseeable future, emphasizing the need for diversified supply chain strategies beyond a single dominant foundry.
Semiconductors & Hardware
East Asian Semiconductor Dynamics: Kioxia, SK hynix, TSMC, ASML, and DRAM Prices
SK hynix’s chairman announced plans to build a memory fab in Japan through a joint venture, but Kioxia CEO Hiroo Ota dismissed any tie-up on NAND production, citing antitrust concerns. Meanwhile, TSMC and ASML are collaborating to transition to larger 12-inch photomasks for High-NA EUV lithography, aiming for increased fab productivity and lower chipmaking costs. DRAM contract prices surged by nearly 60% quarter-on-quarter in Q2, pushing quarterly revenue to $154.73 billion, with further growth expected in Q3.
Why it matters: SK hynix’s attempt to establish a Japanese memory fab highlights the strategic push by Korean players to expand production amidst a global semiconductor rebalancing, but Kioxia’s firm rejection indicates the deep-seated rivalry and the challenges of cross-border alliances in a highly competitive market. The ASML-TSMC move towards 12-inch EUV masks aims to maintain Taiwan’s lead in advanced foundry efficiency, further entrenching their dominance in leading-edge chip production. The DRAM price hike directly impacts the profitability of all electronics manufacturers and the cost structure of data centers, with consumer DRAM seeing the largest price growth due to supply constraints.
For Western readers: Western companies relying on DRAM should factor in continued price increases through Q3, especially for server and consumer segments, and anticipate a further widening of the technology gap in advanced lithography efficiency as TSMC and ASML optimize High-NA EUV for mass production. Do not expect rapid consolidation or joint ventures between major Japanese and Korean memory players despite national industrial policy pressures, as competitive and antitrust concerns remain high.
Semiconductors & Hardware
China’s CXMT tops SK Hynix, Micron to take memory profit margin crown
ChangXin Memory Technologies (CXMT), a Chinese chipmaker, has achieved the highest profit margin among global memory competitors, surpassing SK Hynix and Micron in Q2 2026. This move demonstrates a shift in the memory sector, which has been historically dominated by South Korean, U.S., and Japanese suppliers, driven by CXMT’s leveraging of tight supply in commodity-grade DRAM, particularly DDR5.
Why it matters: CXMT’s lead in profit margins for commodity DRAM, especially DDR5, shows China moving beyond mere capacity expansion to achieving competitive profitability. This is not just about producing chips, but about building a viable, self-sustaining memory industry that can compete on market terms rather than relying solely on state subsidies. Their success with DDR5 is particularly telling, as it’s a critical component for many current computing platforms, including AI infrastructure, suggesting they are moving up the value chain even in commodity memory.
For Western readers: Western businesses in the computing and data center sectors should assume that CXMT will continue to gain market share in standard DRAM, influencing pricing and potentially reducing reliance on traditional suppliers in the medium term.