East Asian Technology Intelligence
Japan & China tech news — translated, contextualized, and delivered to your inbox.
Free. Unsubscribe anytime.
3 Takeaways This Issue
- Nvidia’s record $96.22 billion fiscal Q2 revenue has locked in demand for South Korean memory giants SK Hynix and Samsung, making their high-bandwidth memory production yields the primary bottleneck for global AI hardware deployment.
- OpenAI’s claims that its custom “Jalapeño” inference chip outperforms Nvidia’s GB300 will accelerate efforts by cloud giants to bypass merchant silicon, shifting the competitive landscape from model-building to proprietary chip-design execution.
- TSMC’s capacity constraints in CoWoS packaging are opening a critical entry point for Intel to capture high-bandwidth memory packaging market share from the Taiwanese foundry.
This Issue’s Analysis
The Signal
Nvidia’s Vera Rubin Bottleneck: Why HBM Suppliers Hold the Key to Its 70% Growth Forecast
Nvidia reported record-breaking fiscal Q2 (May-July) revenue of $96.22 billion, exceeding market forecasts. The company expects another record quarter for Q3, projecting $108 billi
Semiconductors & Hardware
Nvidia’s Supply Bottlenecks: How Packaging and Memory Constraints Threaten Its 70% Growth Target
NVIDIA reported Q2 revenue of $96.2 billion, double year-over-year, driven by strong data center performance at $89 billion. CEO Jensen Huang highlighted that demand far outstrips
Semiconductors & Hardware
Intel’s Advanced Packaging Gains Traction as TSMC CoWoS Bottlenecks Persist
Analysts suggest TSMC’s CoWoS advanced packaging, while a key technology, is becoming a production bottleneck for integrating AI accelerators with HBM due to cost and yield risks.
Semiconductors & Hardware
Nvidia’s Revenue Surge Solidifies HBM Demand for South Korean Chipmakers
Nvidia’s stellar Q2 earnings, featuring a 106% year-on-year revenue surge and eased margin pressure concerns, have reinforced a highly optimistic long-term outlook for the global A
🧩 Pattern This Issue
- Korea/Taiwan: Nvidia blowout Q2 earnings solidify HBM packaging demand for SK Hynix and Samsung
- Japan: Fujitsu taps TSMC 2nm and 3D packaging for 144-core Monaka processor
- Korea/Taiwan: TSMC CoWoS capacity bottlenecks open advanced packaging market share for competitors
As Nvidia’s explosive growth collides with persistent TSMC packaging bottlenecks, the semiconductor supply chain is shifting from raw wafer fabrication to advanced 3D and HBM packaging integration, challenging TSMC’s sole dominance and forcing system designers like Fujitsu to lock in next-generation packaging capacity early.
Also This Issue
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 AI & Machine Learning
Chinese AI Company Z.ai Reveals ‘Ox Alpha’ Stealth Model as GLM-5.3-Flash, Running on Domestic Chips
Chinese AI developer Z.ai (Z.ai) announced that its ‘Ox Alpha’ stealth model, which gained significant traction on OpenRouter, was in fact its own GLM-5.3-Flash model. The company revealed that inference for this model during its verification period was handled by a cluster of tens of thousands of Chinese-made AI chips, achieving cost parity with NVIDIA GPUs. Z.ai has now released the model’s weights under an MIT license for commercial use.
Why it matters: The reveal of GLM-5.3-Flash operating at scale on Chinese AI chips challenges the common Western assumption that advanced AI model inference fundamentally relies on NVIDIA’s high-end GPUs. This isn’t merely a software release; it’s a statement on the operational viability and cost-effectiveness of China’s domestic semiconductor and AI ecosystem for foundational model inference.
For Western readers: Western businesses and policymakers should adjust their assumptions about the dependency of Chinese frontier AI on imported chip technology; expect China to accelerate investment in domestic AI chip production and cluster deployment, potentially impacting global supply chains for specialized AI hardware.
🗾 Semiconductors & Hardware
OpenAI’s Custom Chip ‘Jalapeño’ Significantly Outperforms NVIDIA GB300
📊 Featured Chart
Max improvement/reduction from InferenceX benchmark
OpenAI has revealed performance details for its custom AI inference chip, ‘Jalapeño,’ showing it can achieve up to 1.9 times better power efficiency and 4.1 times better performance for interactive workloads compared to NVIDIA GB200/GB300 systems. The chip, which integrates the entire system design from silicon to software, was developed in nine months, leveraging OpenAI’s own AI models for optimization and design acceleration. Jalapeño is slated for deployment in OpenAI’s infrastructure by the end of 2026, with second and third generations already in development.
Why it matters: OpenAI’s ‘Jalapeño’ is a defensive play against NVIDIA’s pricing power and a proactive step to optimize its operational efficiency. By leveraging its own AI models for chip design, OpenAI demonstrates a vertically integrated approach, which aligns with how major East Asian chaebol and keiretsu approach core technology development: self-sufficiency and control over the full stack are prioritized.
For Western readers: Western businesses heavily reliant on NVIDIA’s GPU ecosystem should recognize that major AI developers are actively seeking alternatives, which could lead to shifts in long-term procurement strategies and potentially impact NVIDIA’s market dominance in specific inference applications. Assume this is the start of a trend, not an isolated move.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
🇰🇷 Policy & Regulation
Naver Cloud and LG CNS Partner to Develop National Cyber Security Model with Full-Stack Technology
Naver Cloud and LG CNS have signed a Memorandum of Understanding (MOU) to jointly develop a national-level cybersecurity model utilizing their respective full-stack AI and cloud technologies. The partnership aims to integrate Naver Cloud’s hyperscale AI and cloud infrastructure with LG CNS’s system integration expertise and domain-specific security solutions. The focus will be on AI-driven threat analysis and response for critical national infrastructure.
Why it matters: This partnership is less about technological breakthroughs and more about a concerted effort to build a robust domestic cybersecurity framework within South Korea, leveraging existing chaebol resources and AI capabilities. It exemplifies the current Korean industrial policy push to bring critical infrastructure and data management under national champions, rather than relying on foreign providers, even if those foreign providers offer technically superior solutions.
For Western readers: Western cybersecurity providers should recognize that South Korea’s national security market is increasingly being ring-fenced for domestic players, driven by government policy and chaebol collaboration; direct market access for core infrastructure security will likely diminish. Focus instead on partnerships that augment, rather than replace, indigenous South Korean capabilities.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
🇹🇼 Semiconductors & Hardware
Fujitsu’s 144-Core Monaka Processor Details Revealed: TSMC 2nm and 3D Packaging Integration
Fujitsu unveiled further details of its Monaka server processor at Hot Chips 2026, featuring a 144-core Arm v9.3-A architecture and a 3D multi-die design that integrates cores, cache, and I/O using TSMC’s 2nm and 5nm processes. The computational cores utilize TSMC’s N2P 2nm process, while SRAM and I/O are placed on 5nm dies, connected via hybrid bonding and silicon interposers, with mass production slated for 2027.
Why it matters: Fujitsu’s Monaka processor represents a defensive strategy by a Japanese hardware player to remain competitive in high-performance computing, relying on TSMC’s most advanced foundry technologies. The split 2nm/5nm die approach for cores and cache/I/O suggests a pragmatic move to manage costs and yields while pushing the performance envelope, rather than a pure leading-edge play across the entire chip.
For Western readers: Western businesses sourcing HPC and AI hardware should note that Fujitsu’s reliance on TSMC’s 2nm for critical compute portions further cements Taiwan’s foundry leadership in next-generation processors, reinforcing the existing supply chain dynamics rather than diversifying them. Plan for continued tight capacity at the most advanced nodes.
🇹🇼 AI & Machine Learning
Anthropic Secures Next-Gen NVIDIA Vera Rubin Chips with $45 Billion Deal for Future Compute Power
Anthropic has signed a six-year, $45 billion agreement with Nscale, a cloud infrastructure provider, to lease AI computing capacity from Nscale’s data center in West Virginia. This deal secures future access to NVIDIA’s unreleased Vera Rubin series chips, with compute capacity expected to be delivered starting late 2027.
Why it matters: The rush to secure future GPU capacity well ahead of market availability shows that competitive advantage in generative AI is increasingly defined by access to advanced hardware and large-scale data center infrastructure, not just algorithm development. This forward-looking deal reflects Anthropic’s need to demonstrate a clear path to sustained compute power to justify its nearly $1 trillion valuation ahead of a rumored IPO, especially as rivals like OpenAI and Google also aggressively lock in resources.
For Western readers: Western tech investors and enterprises should recognize that the ‘AI race’ is now fundamentally about securing supply chain access years in advance, particularly for high-end GPUs. Do not assume that current market availability of AI compute capacity reflects future competitive dynamics; the true battle is for unannounced hardware and specialized data center space through long-term contracts.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
🇨🇳 AI & Machine Learning
Mysterious ‘Ox Alpha’ Model Confirmed as Zhipu’s GLM-5.3 Flash, First Native Multimodal GLM Model Powered by Domestic Chips
📊 Featured Chart
Source: QbitAI, 2026-08-27
Chinese AI company Zhipu AI has revealed its new GLM-5.3 Flash model as the mysterious ‘Ox Alpha’ model that recently topped OpenRouter and OpenCode rankings. GLM-5.3 Flash is the first native multimodal model in Zhipu’s GLM-5 series, and notably, it was developed and runs entirely on domestic Chinese hardware. Despite its smaller 320B parameter size (with 18B active parameters), it reportedly outperforms the larger 753B GLM-5.2 and achieves scores comparable to Claude Opus 4.8.
Why it matters: Zhipu AI’s GLM-5.3 Flash model is another indication that Chinese firms are rapidly closing the gap in large multimodal models, not just in benchmarks but in practical application and cost efficiency. The explicit mention of the model running on domestic cards reinforces the ongoing industrial policy push for self-sufficiency in AI compute, positioning China to control more of its AI supply chain.
For Western readers: Western businesses and AI developers should recognize that top-tier AI capabilities, including competitive multimodal models, are increasingly accessible from Chinese providers, often at a lower cost, and built entirely on a domestic hardware stack. This reduces the leverage of US chip export controls in the long term and creates a viable alternative supply chain for AI development.
🇨🇳 Robotics & Automation
Unitree and LimX Dynamics Robots Share a Single Brain in a 10-Minute Uncut Demo
Chinese AI company QbitAI showcased a 10-minute, unedited demo video of embodied AI robots from Unitree and LimX Dynamics (智元) working collaboratively in a complex indoor environment. The robots, despite having different hardware architectures, shared a single AI model, performing tasks like window cleaning, object retrieval, and cooperative organization, demonstrating advanced real-time perception, dynamic planning, and task resumption capabilities.
Why it matters: This demo, if genuine and replicable, suggests a significant leap in general-purpose embodied AI. The ability for a single model to control disparate robot hardware from different manufacturers—and for those robots to autonomously collaborate and adapt to complex, interruptible tasks—moves beyond scripted demonstrations toward true robotic autonomy. It signals that Chinese AI developers are not merely replicating Western research but are potentially innovating on core architectural challenges in embodied AI.
For Western readers: Western hardware manufacturers and AI model developers in robotics should re-evaluate their timelines for achieving cross-platform general intelligence and autonomous multi-robot collaboration, as this demo indicates Chinese firms may be closer than previously assumed. Assume the domestic market for general-purpose robotic platforms in China will accelerate, creating a challenging competitive environment for non-Chinese players there.
🇨🇳 Cross-Regional Analysis
Morning News: Apple iPhone launch Sep 10, Tesla China refutes data center withdrawal, MIIT bans unverified car products
📊 Featured Chart
Source: The Information, unconfirmed by DeepSeek
Apple is set to launch new iPhones and potentially its first foldable iPhone on September 10th. Meanwhile, DeepSeek reported 475 million RMB in revenue and 715 million RMB net loss for the first seven months of the year, while its AI infrastructure investment reached 11 billion RMB. Tesla China denied rumors of withdrawing its Shanghai data center, confirming it remains operational for in-country data storage.
Why it matters: DeepSeek’s 11 billion RMB investment in AI infrastructure over seven months, primarily for leasing AI chip-equipped servers, highlights the immense capital requirements to compete in the foundational model space and the supply-side constraints for advanced AI compute in China. Their 82.9% gross margin on API sales suggests that efficient inference infrastructure is critical for profitability, even as overall losses mount due to massive CAPEX. The Tesla data center situation directly addresses a sensitive point for foreign automakers in China: data localization requirements for autonomous driving and the need for clear communication to avoid regulatory missteps or public speculation.
For Western readers: Western investors in Chinese AI companies should understand that high revenue growth for model providers can still be accompanied by substantial losses due to infrastructure costs, indicating a long path to profitability for many in the sector. Western automakers should note that clarity on in-country data storage and compliance with Chinese data security standards for advanced driving features is non-negotiable for market access and avoiding operational disruptions in China.
🔺 The Prism
Where US and East Asian technology interests intersect
Robotics & Automation
Humanoid Robot Investment Is Booming Globally, East Asian Implications
Global investment in humanoid robot startups reached $4.39 billion in the first half of 2026, surpassing the total investment of the past decade. While the article covers global trends, this surge will directly impact East Asian robotics firms and national AI strategies, intensifying competition and accelerating development in a region already heavily invested in automation.
Why it matters: The massive influx of capital into humanoid robotics, even if originating globally, will inevitably accelerate development and market formation in East Asia. This means greater competition for local firms like Boston Dynamics’ former owner SoftBank, or Hyundai Robotics, and could lead to faster deployment of humanoids in manufacturing and logistics across the region.
For Western readers: Western robotics firms and investors should expect heightened competition and faster technology cycles from East Asian players, as this global funding wave enables rapid scaling and product iteration among well-capitalized regional startups and incumbents.
Semiconductors & Hardware
Kioxia to Invest $6.3 Billion in New Memory Production Facility in Japan
Kioxia Holdings plans to invest over 1 trillion yen ($6.27 billion) to construct a new memory chip production facility at its plant in northern Japan. This move positions Japan’s leading NAND flash maker to increase output, following similar capacity expansion strategies by Korean rivals Samsung Electronics and SK hynix.
Why it matters: This investment by Kioxia represents a significant commitment to maintaining Japan’s position in the global NAND flash market. While Samsung and SK hynix are also expanding, Kioxia’s move, supported by Japanese industrial policy, aims to secure a domestic supply base and push advanced memory technology within Japan, rather than relying solely on foreign production.
For Western readers: Western businesses sourcing NAND flash memory should anticipate Kioxia’s increased output to provide a more stable, geographically diversified supply option over the next few years, potentially reducing over-reliance on South Korean manufacturers. If you are sourcing advanced flash memory, watch for announcements of specific product generations and capacity timelines from this new facility.
Policy & Regulation
Huawei Licenses Key Wi-Fi Tech to HP Ahead of Trump-Xi Meeting
Huawei has licensed core Wi-Fi connectivity technologies to HP under a multi-year deal, allowing the American PC maker to use them in computers and peripherals sold globally. This agreement represents a significant milestone for Huawei in monetizing its intellectual property and continuing to operate internationally despite ongoing US sanctions.
Why it matters: Huawei’s licensing of key Wi-Fi technology to HP indicates a broader strategy to leverage its extensive patent portfolio as a revenue stream, adapting to US export controls that restrict its hardware business. This move by Huawei to monetize its IP assets suggests a shift from manufacturing dominance to a more service- and licensing-oriented model, which could be a blueprint for other Chinese tech firms facing similar restrictions.
For Western readers: Western businesses should recognize that even sanctioned Chinese tech companies like Huawei will seek to generate revenue and maintain influence through intellectual property licensing, meaning a ‘decoupling’ of technology is far more complex than simply blocking hardware sales.