EAST ASIAN TECHNOLOGY INTELLIGENCE
Issue #80 · 2026-08-29
U.S. Cloud Export Controls: Closing the Remote Access Loophole for Chinese AI Firms
3 Takeaways This Issue
- The US Commerce Department's move to block Chinese firms from using cloud-based remote access to bypass export controls will force Chinese AI developers to rely more heavily on domestic hardware, accelerating the growth of local alternatives like Huawei's Ascend chips.
- Kioxia and Western Digital's ¥5 trillion expansion of their joint-venture flash memory plants in Japan secures a massive domestic supply of NAND flash, insulating global AI storage supply chains from geopolitical disruptions in the Taiwan Strait.
- Changxin Memory Technology’s leap to a 1.8 trillion yuan net profit in the first half of 2026 demonstrates that China's state-backed memory sector is rapidly scaling production to capture the surging domestic demand for AI hardware.
Core Move
US Moves to Block Remote Access to AI Chips in China, Closing Export Control Loophole
The US Commerce Department wants to block remote access to advanced AI chips in China. This move shifts the battleground of US-China tech competition. The rivalry is moving from hardware sales to compute-as-a-service. This is not just about closing a loophole. It shows that the physical location of a chip matters less than who controls its power. Washington wants to make sure China cannot build a top-tier AI ecosystem using US-designed compute. This rule applies even if that compute sits in Singapore instead of Shenzhen.
This change moves beyond the first phase of export controls. Those initial rules focused on physical hardware. Now, the goal is to cut off access to performance. For Beijing, this makes the need to build domestic AI compute even more urgent. This rule turns a supply chain risk into a real bottleneck for Chinese cloud firms and AI developers. It forces them to rely on domestic options, even if those options are worse.
Korean media, like the AI Times, sees this as a natural but bold extension of US policy. Regional players understand that these controls are tightening. Many Western writers focus on the financial loss for US chipmakers. In contrast, East Asian observers focus on national AI competitiveness. For Seoul, this policy means a new focus on its own domestic AI supply chains. South Korea must build its own compute infrastructure as the global market splits.
Many Western experts thought that restricting chip shipments would stop China's AI progress. This new rule challenges that idea. It directly targets how Chinese firms find other paths to compute. However, enforcing this rule will be hard. It is difficult to track and police cloud usage across complex global networks. This task is much harder than stopping physical shipments at a border.
To judge this policy, we must watch the investments of Chinese chip firms like Huawei and Alibaba. We should also track the capacity plans of major Chinese cloud providers. We need to monitor their new compute capacity and the specific chips they use. If these US rules work, China will invest more in its own chip factories. We will see a push for new compute architectures within the next 12 to 18 months.
🗾 Japan Radar
As reported in Japan — what the Japanese-language press is covering
🗾 Semiconductors & Hardware2 STORIES
Kioxia and SanDisk Detail ¥5 Trillion Flash Memory Expansion in Japan
Joint venture partners Kioxia and Western Digital's SanDisk have announced a massive ¥5 trillion ($31 billion) investment through 2032 to boost advanced NAND flash memory production at their Yokkaichi and Kitakami plants in Japan. This long-term expansion, contingent on government subsidies, is designed to meet the soaring global demand for AI data storage and secure Japan's position in the international semiconductor race against rivals in the US, China, and South Korea.
Why it matters: This investment solidifies the US-Japan semiconductor alliance as a dominant counterweight to South Korea's HBM dominance and China's state-backed memory expansion, reshaping the East Asian supply chain hierarchy for AI-era data centers.
For Western readers: Western tech leaders must abandon the assumption that state-backed semiconductor reshoring is primarily a US or European phenomenon; Japan is rapidly leveraging foreign joint ventures to establish itself as the most reliable, subsidized hub for next-generation AI memory production.
🗾 Semiconductors & Hardware
China's CXMT Reports $1.8 Trillion Profit in H1 2026 as AI Drives Memory Demand
Changxin Memory Technology (CXMT), a leading Chinese semiconductor memory producer, announced a net profit of 77.6 billion yuan ($1.8 trillion) for the first half of 2026, a significant turnaround from a 2.3 billion yuan loss in the same period last year. Sales revenue soared to 150.3 billion yuan, nearly ten times higher than the previous year, driven by surging AI demand and a semiconductor supply shortage that pushed up selling prices. The company also significantly increased its R&D staff and budget, exceeding its own financial forecasts.
Why it matters: This remarkable profit surge solidifies CXMT's position as a formidable competitor in the global memory market, directly challenging established players like Samsung, SK Hynix, and Micron. Its aggressive expansion and R&D investment, backed by significant capital from its recent IPO, indicate a determined effort to capture greater market share and reduce China's reliance on foreign memory suppliers.
For Western readers: Western businesses in the semiconductor industry should now assume that China's domestic memory production capacity is accelerating significantly, potentially altering global supply chain dynamics and increasing competition in the DRAM market.
🇰🇷 Korea Signal
As reported in Korea — memory, chips and platform moves from Korean sources
South Korea's memory bottlenecks and platform dependencies leave its hardware giants highly vulnerable to US-China decoupling and Nvidia's market moves.
🇰🇷 Policy & Regulation
US Moves to Block Remote Access to AI Chips in China, Closing Export Control Loophole
The US Department of Commerce is planning new regulations to prevent Chinese companies from circumventing AI chip export controls by using cloud services or remote access to advanced chips hosted outside China. This move aims to close a loophole that has allowed Chinese firms to access high-performance AI compute power, even if they cannot directly purchase the chips. The Commerce Department expects to issue these new rules this autumn.
Why it matters: This move signals the US is systematically tightening every avenue for China to access advanced AI compute, regardless of where the physical hardware is located. It shifts the focus from merely restricting chip sales to controlling compute access, directly impacting how Chinese AI companies can operate globally and scale their operations.
For Western readers: Western cloud providers and data center operators that host AI chips and serve Chinese clients should review their compliance frameworks, as they will likely face new restrictions on providing compute access to entities in China. Assume the scope will broaden, not narrow.
🇰🇷 AI & Machine Learning
Why Nvidia Might Acquire Hugging Face: From Open Model Download to Distribution
AI Times Korea discusses speculation that Nvidia could acquire Hugging Face, suggesting such a move would shift Hugging Face's role from a platform for downloading open-source AI models to a robust distribution channel. This acquisition would deepen Nvidia's integration into the AI software ecosystem, leveraging Hugging Face's developer community and model repository.
Why it matters: Nvidia's dominance in AI hardware is established, but control over software and model distribution is the next frontier. Acquiring Hugging Face would give Nvidia direct influence over the open-source AI community and the flow of models, providing a strategic defense against competitors developing their own full-stack AI solutions or alternative hardware architectures. It also means Nvidia would control a key layer of the AI stack currently seen as neutral.
For Western readers: Western AI developers and companies relying on Hugging Face for open-source models should recognize that Nvidia's potential ownership could shift the platform's neutrality and introduce new commercial incentives or integration points with Nvidia's ecosystem, potentially altering how models are accessed and deployed.
🇰🇷 Semiconductors & Hardware
Surge in Memory Prices Impacts NVIDIA's Performance Outlook
NVIDIA is reportedly lowering its revenue projections for its fiscal third quarter due to a sharp increase in HBM (High Bandwidth Memory) prices. This situation is compelling NVIDIA to adjust its product pricing and potentially delay some shipments to manage costs.
Why it matters: NVIDIA's decision to revise its outlook due to HBM costs directly contradicts the prevailing narrative that AI chip demand is impervious to pricing pressures. This indicates that even a market leader like NVIDIA faces constraints from its supply chain, affecting both its margins and delivery schedules.
For Western readers: If you are an investor in NVIDIA or an enterprise client planning large AI infrastructure deployments, assume a tighter supply and higher pricing environment for HBM-intensive chips will persist, potentially delaying rollouts or increasing CapEx. Focus on the actual availability of HBM capacity, not just headline demand.
🇹🇼 Taiwan Silicon
As reported in Taiwan — foundry, hardware and enterprise IT from the Taiwanese press
Taiwan's power grid cannot support Nvidia's local expansion, threatening the island's bottleneck control over the global AI hardware supply chain.
🇹🇼 AI & Machine Learning
Google Launches Gemini 3.5 Transcribe Model for Speech-to-Text, Speaker Diarization, and Oral Fluency Processing
Google has introduced Gemini 3.5 Transcribe, a speech-to-text model available in public preview via the Gemini API and Gemini Enterprise Agent Platform. It features advanced capabilities like automatically removing filler words, handling speaker corrections, and organizing punctuation, numbers, and paragraphs into readable text.
Why it matters: Google's move with Gemini 3.5 Transcribe signals a continued investment in improving AI's utility for enterprise applications, particularly in Asia where accurate transcription of complex spoken languages and dialects can unlock significant business value. The focus on cleaning up oral speech and distinguishing multiple speakers addresses key pain points for companies dealing with large volumes of audio data, from customer service calls to international business meetings.
For Western readers: Western businesses operating in East Asia, especially those with multilingual teams or customer bases, should evaluate Gemini 3.5 Transcribe's capabilities for Mandarin Chinese (Simplified) and Cantonese (Traditional) to enhance communication and data analysis, assuming Google expands its language support for specific regional variants like Taiwanese Mandarin.
🇹🇼 Semiconductors & Hardware
Nvidia Earnings Send Stock Soaring, But Revolving Financing Risks Now A Focus As Guarantees Exceed Quarterly Revenue
📊 Featured Chart
Source: Nvidia Q2 FY2027 Earnings
Nvidia reported strong Q2 FY2027 earnings, exceeding market expectations with $96.2 billion in revenue and a 106% year-over-year increase, driven by a 117% surge in data center revenue. The company also issued an optimistic Q3 forecast of $108 billion, signaling continued strong demand for AI chips. However, concerns regarding Nvidia's revolving financing deals, where total guarantees to AI partners (including a $105 billion guarantee for OpenAI's data center lease with SoftBank) now exceed its quarterly revenue, are drawing scrutiny.
Why it matters: Nvidia's financial engineering to support its AI ecosystem, particularly the substantial guarantees for OpenAI's data center expansion, suggests a deeper commitment to fostering its market dominance beyond just selling chips. This move could entrench Nvidia's position by enabling its partners to scale, but it also ties Nvidia's balance sheet more directly to the success of specific AI ventures and their infrastructure buildouts. The memory shortage issue is a genuine constraint on Nvidia's growth that local coverage understands acutely.
For Western readers: Western investors should understand that Nvidia's growth projections are increasingly tied to its ability to manage both the physical supply chain, especially memory, and the financial stability of its key AI cloud partners. If you are an investor, look for further details on how these guarantees are structured and what recourse Nvidia has if a partner faces financial difficulties. If you are in the AI infrastructure buildout space, understand that Nvidia's guarantees are a strong incentive for partners to prioritize its ecosystem, but also signal the immense capital requirements to scale AI.
🇹🇼 Policy & Regulation
NVIDIA Wants Power, Not Substations? Northern Taiwan Grid Nears Overload as Public Misunderstands Energy Mix
📊 Featured Chart
Source: TechNews, based on current electricity mix
NVIDIA's new 'Star Cluster' R&D and manufacturing hub in Beitou-Shilin Technology Park faces local opposition to a critical substation project, highlighting Taiwan's intensifying electricity grid challenges. A recent survey reveals that while the public supports AI and semiconductor industry expansion, there's a significant lack of understanding regarding power generation sources and the need for new infrastructure.
Why it matters: Taiwan's energy supply is critical to the global semiconductor supply chain, and local resistance to infrastructure directly threatens the expansion plans of companies like NVIDIA and TSMC. The public's misinformed opposition to substations could create a severe bottleneck, preventing new AI and semiconductor facilities from securing the necessary power to operate, directly impacting global tech production capacity.
For Western readers: Western companies relying on Taiwanese manufacturing should anticipate potential delays and increased costs due to energy infrastructure bottlenecks; this issue could tighten capacity for CoWoS packaging and other power-intensive processes. Watch for how the Taiwan government resolves local opposition to grid projects, as it will signal the viability of future high-tech investments there.
🇨🇳 China Watch
As reported in China — from Chinese-language technology media
China’s tech giants are pivoting from frontier LLM racing to open-sourcing highly specialized, domain-specific AI tools for immediate commercial utility.
AI & Machine Learning
Tencent Open-Sources Hy4 LLM with 770B Parameters and 1M-Token Context
📊 Featured Chart
Source: Tencent internal blind evaluation
Tencent Hunyuan has open-sourced a preview of its Hy4 large language model, featuring 770 billion total parameters and a 1 million token context window. The model, designed for productivity tasks, is accessible through Tencent's various applications and via API, with a limited free trial period.
Why it matters: Tencent's decision to open-source a model of this scale, even as a preview, indicates a shift towards a more collaborative, platform-driven strategy to capture developer mindshare in China. The company's internal evaluation claims of efficiency gains in its own systems are notable, suggesting this isn't just a public-facing play but a sign of genuine integration and optimization within Tencent's tech stack.
For Western readers: Western businesses operating in China, or those with Chinese partners, should expect to see Tencent's Hy4 ecosystem gain traction in enterprise applications, potentially becoming a de facto standard for certain productivity and development tasks within the Chinese market. It will be important to understand how its API and enterprise offerings integrate into local supply chains.
🇨🇳 AI & Machine Learning
Coding No Longer Exclusive to Programmers! Alibaba's Qoder Shows Impressive Advances
Alibaba's Qoder, an AI agent for coding, has released a new desktop version featuring a 'desktop pet' interface that allows users to generate code and manage projects through natural language conversations. The updated Qoder can understand vague verbal requirements, plan development tasks, write code, run tests, and self-correct, operating beyond traditional Integrated Development Environments (IDEs). It has already served 6 million users and 100,000 enterprise clients globally in the past year.
Why it matters: Alibaba is pushing the boundaries of AI-driven code generation beyond the developer community, moving it to the desktop where non-programmers can interact with it conversationally. This reflects a strategic pivot from merely assisting developers to enabling entirely new classes of users to build digital tools, positioning Qoder not just as a coding agent but as a 'digital execution' layer for various tasks.
For Western readers: Western businesses should recognize that Chinese AI players like Alibaba are not just competing on model benchmarks but on user accessibility and integration across diverse workflows; assume this will lead to a broader, faster diffusion of AI-driven tool creation within Chinese enterprises compared to Western counterparts still focused on developer-centric tools.
AI & Machine Learning
Ant Group Launches Finance-Tuned Ling-3.0-flash-Fin AI Model, Plans Open-Source Release
Ant Group has introduced Ling-3.0-flash-Fin, a large language model specifically optimized for financial tasks, leveraging the existing Ling-3.0-flash architecture. The model, with 124 billion total parameters and 5.1 billion active parameters, is designed for financial document analysis, investment analysis, and banking applications. Ant Group plans to open-source the model weights next week, offering a one-month free API period for professionals.
Why it matters: Ant Group's move to open-source a finance-tuned model signals a strategy to accelerate adoption and foster an ecosystem around its AI capabilities within China's financial industry. This could enable smaller financial institutions to integrate advanced AI without heavy upfront development costs, potentially reinforcing Ant Group's platform dominance in adjacent financial services through its model's ubiquity.
For Western readers: Western financial technology firms and data providers should recognize this as a direct challenge to their market share in China, as a potent, open-source domestic alternative will likely be favored by local regulators and customers.
🔺 The Prism
Where US and East Asian technology interests intersect
East Asian tech giants are moving aggressively beyond software to lock down the raw hardware, power, and agentic infrastructure powering global AI.
Semiconductors & Hardware
Kioxia Plans $31.4 Billion NAND Investment, SK hynix Considers Deeper Ties
Kioxia, with Western Digital's SanDisk, plans to invest $31.4 billion over six years into new NAND flash capacity, with $11.3 billion allocated to a new fab in Kitakami, Japan, seeking significant government subsidies. Concurrently, SK hynix is evaluating closer collaboration with Kioxia, hinting at potential strategic moves beyond its current indirect stake through a Bain Capital investment vehicle.
Why it matters: Kioxia's investment, backed by Japanese government subsidies, aims to secure Japan's position in advanced NAND manufacturing, leveraging the familiar 'national champion' consortium model to shore up domestic tech supply chains. SK hynix's explicit interest in 'co-developing' the NAND market with Kioxia signals a potential strategic alignment that could shift the competitive balance in the global memory sector, moving beyond its existing indirect financial holding.
For Western readers: Western companies relying on NAND supply should track Kioxia's Kitakami fab buildout and the progress of Japanese government subsidies, as this capacity could stabilize or increase future supply. They should also watch for any concrete announcements from SK hynix regarding its relationship with Kioxia, as a deeper alliance could impact pricing power and technology roadmaps in the NAND market.
Cross-Regional Analysis
SoftBank's US Gas Power Plant Faces Mixed Local Reaction Amid AI Data Center Buildout
SoftBank Group's plans for the largest natural gas-fired power plant in the U.S., located in Waverly, Ohio, have received a mixed reception from local residents at a public meeting. This plant is intended to power a major AI data center, highlighting SoftBank's strategy to secure energy infrastructure for its growing AI investments.
Why it matters: This initiative underscores how crucial energy supply has become for AI infrastructure, driving Japanese firms like SoftBank to directly invest in power generation in key overseas markets. It shows that the buildout of AI capacity is less about pure compute and more about the entire integrated supply chain, from power to cooling.
For Western readers: Western businesses involved in AI infrastructure development, or those planning large-scale data center operations, should factor direct energy production or long-term energy contracts into their strategic planning, as the scramble for reliable power becomes a bottleneck.
AI & Machine Learning
Automata from Agent Traces: Failure and Next-Step Prediction in LLM Agents
Researchers, including Seonglae Cho, published a paper introducing a method to extract compact finite-state machines (FSMs) from large language model (LLM) agent traces. These FSMs provide a structural substrate for auditing and runtime monitoring, enabling more accurate next-step prediction and early failure detection in complex multi-step tasks. The FSMs demonstrate high fitness in replaying data and are computationally efficient to build.
Why it matters: The core problem this paper tackles — making LLM agent behavior predictable and auditable — is a major barrier to enterprise adoption, not just a research curiosity. East Asian industrial firms, particularly in manufacturing and logistics, are keen to automate complex workflows with AI agents, but they demand reliability and clear accountability that current black-box LLMs often lack. This method offers a path to meeting those industrial requirements.
For Western readers: Western enterprises deploying LLM agents, especially in regulated industries, should assess how this FSM-based approach could be integrated into their existing MLOps and safety frameworks to enhance observability and reduce operational risk, as the need for explainable AI becomes more pressing globally.
🧩 Pattern This Issue
- Policy: US closes remote cloud access loophole for Chinese AI developers
- China: CXMT books record profits on domestic AI-driven memory demand
- Taiwan/Korea: Nvidia faces Taiwan power grid constraints and rising memory costs
As Washington squeezes China's access to compute via cloud loopholes, Beijing is successfully pivoting to domestic hardware self-sufficiency, while Nvidia's supply chain faces physical limits from Taiwan's energy grid and rising Korean memory costs.
AsiaAI.FYI · Written by Dick Weisinger · Subscribe