Taiwan's Minister of Economic Affairs Kung Ming-hsin confirmed to Bloomberg on October 16, 2026, that the Taiwanese government plans to include AI servers and advanced semiconductors in the export control list under the Foreign Trade Act. Once the new rules take effect, unauthorized shipments of restricted AI hardware to mainland China will directly constitute a criminal offense. This is the first time in Taiwan's history that illegal exports of AI computing hardware have been placed within the scope of criminal prosecution.
In the interview, Kung stated explicitly: “Going forward, the measures we take on certain controlled items will correspond with those of the United States, because we made many commitments in the reciprocal trade agreement.” He was referring to the security provisions covered by the Taiwan-U.S. Reciprocal Trade Agreement signed this year.
From Blacklist to Blanket Ban: A Fundamental Policy Shift
Previously, Taiwan's export control system adopted a “company-by-company blacklist” model—mainly targeting specific sanctioned entities such as Huawei and SMIC. As long as the buyer was not on the list, the same model of AI hardware could be legally shipped.
The new policy will break this logic: the scope of restrictions may no longer be limited to specific companies but expanded to all relevant customers in China. This changes from “prohibiting sales to certain people” to “prohibiting sales to everyone.” For AI hardware suppliers, the cost of determining compliance will rise significantly, and the original gray areas will disappear.
At present, Taiwan has no dedicated restrictive provisions on AI chips and servers. Even when smuggling is discovered, prosecutors can only charge peripheral offenses such as document forgery and breach of trust, resulting in very low prosecution efficiency. Once the new rules are implemented, they will provide prosecutors with direct legal tools.
A Smuggling Case Exposed the Real Loophole in the Law
In August 2026, the Keelung District Prosecutors Office in Taiwan indicted nine people involved, including a Super Micro executive and an Nvidia employee (Taiwanese prosecutors disclosed only the surname “Chang”). The smuggling involved about 50 Supermicro servers equipped with Nvidia GB300 chips, of which 74 units flowed into mainland China through a third country (Japan to Hong Kong) using false documents. Investigators found that the smugglers placed empty-shell fake servers during customs inspections, while the real machines had already been transferred in advance, and they tampered with labels and serial numbers.
In the same period, the U.S. Department of Justice also indicted three individuals involved in similar operations, including Supermicro co-founder Wally Liaw, with the amount involved reaching $2.5 billion.
Nvidia stated regarding the incident that the company itself was not accused of any corporate wrongdoing and emphasized that smuggled products would not receive any service or support.
The Strategic Logic of the Taiwan-U.S. Agreement
Taiwan's action is embedded in a larger geopolitical framework. The security annex of the Taiwan-U.S. Reciprocal Trade Agreement requires Taiwan to align some export control measures with those of the United States.
From the U.S. perspective, this arrangement fills a key gap in the previous control system: the United States has restricted high-end chip exports by Nvidia to China since 2022, but smuggling routes through Taiwan have always been a blind spot in enforcement. Bringing Taiwan into a criminal prosecution framework is equivalent to closing this side channel.
TSMC is investing $265 billion in U.S. plant construction and is in cooperation negotiations with Musk's Terafab project. Under such a highly strategic binding relationship, Taiwan's alignment with Washington on chip export controls is both a treaty obligation and a political signal.
The Real Gap in China's AI Training Compute
To assess the impact of Taiwan's upgraded controls on China's large-model research and development, it is necessary to distinguish between inference and training.
On the inference side, China has already found viable domestic alternatives. DeepSeek is building a gigawatt-scale data center in Ulanqab, Inner Mongolia, and plans to deploy at least 160,000 Huawei Ascend 950DT chips, expected to come online in batches between the end of 2027 and early 2028.
The training side is different. DeepSeek once attempted to train its next-generation model with Huawei Ascend chips but encountered stability problems and inter-chip interconnect failures, ultimately forcing it to limit Huawei chips to inference workloads. According to research by Epoch AI, the gap in training compute between Chinese AI companies and leading U.S. institutions is currently estimated at about three to six months, and this gap continues to widen.
According to BigGo Finance, Nvidia chip shipments to China have halved from their peak, and shipments of domestic non-GPU accelerators have surpassed GPU accelerators for the first time. These figures show that Chinese large-model developers have begun to switch on the inference side, but on the training side, there is still no mature domestic alternative.
Can Algorithmic Efficiency Bridge the Compute Gap?
Over the past 18 months, an optimistic narrative has circulated in China's AI community: as long as the algorithm is smart enough, it can produce models of equivalent quality with less compute. DeepSeek V3 completed training with 2,048 H800 chips at a cost of $5.6 million and achieved near-frontier performance.
Algorithmic innovation can compress training costs and improve inference efficiency, but it cannot break physical laws—when next-generation models require larger-scale pretraining, the absolute ceiling of compute still matters. The H800 is a downgraded product Nvidia specifically developed for the Chinese market; compared with the H100, it already has limitations in interconnect bandwidth. And China has never legally obtained next-generation products such as the GB300 and B200.
Taiwan's new policy closes off the possibility of “using smuggling to make up for the failure of legal channels.” When both legal import and illegal smuggling are blocked at the same time, the ceiling on China's AI training compute will be harder to break through with one-off architectural innovation.
Strategic Judgment: Compute Controls Enter a New Phase
When U.S. chip export controls were introduced in 2022, many people believed they were merely a piece of paper and that the gray market would quickly fill the gap. In the years that followed, reports continued to emerge of Nvidia processors flowing into China through smuggling networks in Southeast Asia.
The significance of Taiwan's policy this time lies in upgrading the control tool from an “administrative ban” to “criminal liability” and extending the enforcement boundary from the destination country to the upstream of the supply chain. Smugglers not only face prosecution by the U.S. Department of Justice but will now also face criminal prosecution by Taiwanese prosecutors. The superposition of the two judicial systems fundamentally changes the cost structure of gray channels.
Kung Ming-hsin did not disclose the specific list of controlled items, exemption clauses, or implementation timetable. The subsequent rulings in the smuggling case indicted in August 2026 will become a touchstone for the credibility of the new framework.
AI compute is becoming a strategically controlled material alongside oil and nuclear technology. For China's AI industry, relying on domestic chips to complete frontier model training is no longer one strategic option; it is increasingly approaching the only way out.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接