發生了什麼事
The US government is reviewing policies to restrict foreign remote access to cloud computing, aiming to close loopholes that allow Chinese AI firms to train models on Nvidia chips located in Southeast Asia, Japan, and the Middle East. This follows the passage of the Remote Access Security Act by the US House in January, which would extend export controls to virtual server access.
Chinese AI developers have historically bypassed US chip export bans by leasing server space in foreign data centers, often using proxy entities and cryptocurrency payments to maintain secrecy. According to a Shenzhen-based broker cited by the South China Morning Post, this practice has allowed domestic firms to train frontier models on Nvidia processors located outside US jurisdiction.
The US House passed the Remote Access Security Act in January, which, if signed into law, would expand export controls to include remote cloud computing access. While the Biden administration's 'AI Diffusion Rule' was rescinded, the Trump administration is reportedly reviewing remote access arrangements and working on policies to curb them. Legal experts note that current compliance remains a grey area, but the BIS could theoretically impose penalties under 'General Prohibition 10' for transactions involving illegally smuggled chips.
The proposed restrictions threaten to disrupt the data center boom in Southeast Asia, a region where Chinese tech giants like Alibaba, Tencent, and ByteDance have invested heavily. Analysts suggest that while the immediate shock could create a transient supply-demand imbalance, the region's deep demand pool means released capacity would likely be absorbed quickly by other customers.
For Chinese AI firms, losing overseas cloud access exposes the limitations of domestic hardware. While models like DeepSeek-V4 and Moonshot’s Kimi K3 have been adapted for on Chinese chips, training still relies heavily on Nvidia hardware. The H200 chip, approved for import, is available in small quantities that fall far short of the tens of thousands of GPUs required by top-tier tech giants.
為什麼這很重要
This potential regulatory shift targets the critical infrastructure gap in China's AI sector. By cutting off remote access to high-performance Nvidia GPUs, the US aims to force a faster transition to domestic hardware, which currently lags in software ecosystem maturity and training efficiency. The move could significantly impact Nvidia's international revenue and reshape data center demand in Southeast Asia.
The move represents a significant escalation in US tech policy, shifting the focus from physical hardware shipments to virtual infrastructure. This could curtail Nvidia's sales growth momentum, although analysts suggest this shortfall may be offset by strong demand from other international tech giants.
The restriction accelerates China's push for technological self-sufficiency. Huawei has announced that its next-generation Ascend 960DT chip, designed for model training, will be ready in Q1 2027, nine months ahead of schedule. However, migrating from Nvidia's CUDA platform to Huawei's ecosystem presents steep engineering hurdles, potentially adding 50% to training time and cost.
The policy change has broader implications for the global AI ecosystem, challenging the role of Southeast Asia as a neutral hub for AI infrastructure. It also highlights the growing tension between US national security interests and the globalized nature of AI development.
互動機制：它實際上是如何運作的
以互動方式探索這項發展背後的基礎技術。
Which component of an AI application is the machine-learning model itself?
接下來看什麼
Monitor the legislative status of the Remote Access Security Act and any executive actions by the Bureau of Industry and Security (BIS) regarding remote access penalties. Also watch for accelerated adoption of Huawei's Ascend chips by major Chinese tech firms and potential price adjustments in overseas data center markets.
The final status of the Remote Access Security Act and whether it is signed into law or implemented via executive order.
Any new penalties or enforcement actions by the BIS against data centers or model developers using remotely accessed Nvidia chips.
The pace of adoption of Huawei's Ascend 950 and 960DT chips by major Chinese AI labs, and any reported performance benchmarks compared to Nvidia's H100 and B200 series.
Changes in cloud rental rates in Southeast Asia and other regions as Chinese demand shifts or disappears.