US Closes Chinese Chip Loophole: Overseas Subsidiaries Now Blocked

The US Department of Commerce confirmed on June 1, 2026 that its export controls on advanced AI chips apply to any company with a Chinese parent or headquarters, regardless of where that company's subsidiaries are physically located. The guidance closes a loophole that US officials believe enabled hundreds of thousands of Nvidia and AMD AI chip servers to reach Chinese-controlled entities through subsidiaries established in Malaysia, Singapore, and the United Arab Emirates — countries not on the restricted list.

This is the most aggressive extraterritorial AI chip enforcement action the US has taken to date. It moves export control enforcement from a geography-based model (where the chip ships) to a beneficial ownership model (who ultimately controls the entity receiving the chip) — the same standard the US Treasury's OFAC has used for financial sanctions for decades.

What the BIS Guidance Actually Changed

Before the June 1 guidance, the primary mechanism for blocking AI chip shipments to China was a combination of entity list designations and country-level licensing requirements for exports to the People's Republic of China. That framework applied to shipments going to China, or to entities already listed on the BIS Entity List.

The loophole: an entity not on the entity list, not physically located in China, could order advanced chips without triggering the same controls — even if that entity was a wholly owned subsidiary of a Chinese parent company. A ByteDance subsidiary in Singapore, a Baidu AI research center in Malaysia, or an Alibaba Cloud node in Amsterdam were not treated as "Chinese" for export control purposes as long as they were incorporated outside China.

The June 1 guidance ends that. BIS now states that the licensing requirement for advanced AI chips attaches to the ultimate beneficial owner or parent company, not just the immediate end-user location. If a company has headquarters or a parent in China, its subsidiaries anywhere in the world are subject to the same AI chip export restrictions that apply to China-located buyers.

How the Loophole Worked: Malaysia, Singapore, and the UAE as Pass-Through Points

The mechanism was straightforward. Chinese AI and cloud firms incorporated operating subsidiaries in jurisdictions with strong commercial infrastructure and no chip export controls: Malaysia (specifically Johor, adjacent to Singapore), Singapore, the UAE (Dubai and Abu Dhabi), and the Netherlands.

These subsidiaries ordered Nvidia Blackwell GB200 NVL72 systems, Nvidia Rubin-class hardware (released in late 2025), and AMD MI350x accelerators through normal commercial channels. The orders were fulfilled by Nvidia's distribution partners who had no obligation to check the ownership structure above the subsidiary level — the subsidiary itself was a legitimate legal entity in a non-restricted country.

Once delivered, the hardware supported the AI compute needs of the parent Chinese company's operations globally, including — in some cases — sending data and model outputs back to China-based research teams. The chips never physically crossed into China, but the compute power effectively did.

The Scale: Hundreds of Thousands of Servers Through One Blind Spot

A source cited by the South China Morning Post estimated the volume of hardware that flowed through this loophole at "probably in the range of hundreds of thousands of servers." Tom's Hardware's reporting referenced a BIS internal estimate suggesting the number could be even higher.

At the Nvidia Blackwell GB200 NVL72 configuration — 72 GPU-equivalents per server rack — hundreds of thousands of servers translate into tens of millions of GPU-equivalent compute units flowing outside the intended control framework. This is not a small compliance gap. It is a material breach in the export control regime that may have partially offset years of primary chip restrictions.

BIS's year-long lag in closing this loophole raises questions about whether the guidance is reactive (they discovered the loophole through intelligence) or proactive (they knew and tolerated it until it became too large to ignore). Either way, the scale confirms that entity list additions and country-level restrictions alone are insufficient when companies can restructure their procurement through legal entities in non-restricted jurisdictions.

Which Companies and Industries Are Now Directly Affected

The guidance affects every Chinese-parented company that has been buying or planned to buy advanced AI chips through non-China subsidiaries. The categories include:

Chinese AI cloud providers — Alibaba Cloud (Singapore, Malaysia, and Europe operations), Baidu AI Cloud (multiple Asian markets), Tencent Cloud (international data centers), and ByteDance (global data science infrastructure).

Chinese AI research labs with international offices — DeepSeek, Zhipu AI, Moonshot AI (Kimi), and others that operate research outposts in Singapore or the US.

Chinese financial technology firms operating in Singapore's financial district that use GPU compute for quantitative modeling, fraud detection, and AI-driven trading systems.

Joint ventures and subsidiaries where a Chinese company holds majority or controlling interest in an entity incorporated in a third country — even if minority Western investors are also present.

The guidance leaves one significant ambiguity: what percentage of Chinese ownership triggers the restriction? OFAC's 50% rule (the US person prohibition applies to any entity 50% or more owned by a sanctioned party) is the likely model, but BIS has not yet published a specific ownership threshold in the June 1 guidance.

How Cloud Providers in Asia Are Now Exposed to Compliance Risk

The immediate compliance problem is not just for Chinese companies — it is for the cloud providers who have been serving them.

AWS, Google Cloud, and Microsoft Azure all operate large data centers in Singapore and Malaysia. Enterprise cloud customers in these regions include subsidiaries of Chinese parent companies that are now in scope for the new guidance. If a Chinese-parented company's Singapore subsidiary is running workloads on AWS Singapore compute infrastructure using Blackwell or Rubin GPUs, that is potentially an export control violation — by the Chinese company, and potentially by AWS for providing the service with knowledge of the ownership structure.

Cloud providers now face a compliance review requirement for their enterprise customer base in Asia Pacific. Enhanced due diligence on account ownership structures — currently not standard practice — becomes necessary. Service termination for non-compliant tenants becomes a real option that legal teams must model.

This creates a short-term revenue disruption for hyperscalers in APAC and a longer-term restructuring of how AI compute is provisioned for Chinese-affiliated businesses outside China.

Our Analysis: Extraterritorial Chip Controls and What Comes Next

The June 1 guidance is the first clear statement that the US intends to apply AI chip controls extraterritorially — following the beneficial owner, not just the shipping address. This mirrors how the US has applied financial sanctions for the past 30 years.

The practical implication is that geography is no longer a reliable shield. A Chinese AI company cannot solve its chip access problem by incorporating a Singapore subsidiary, hiring local staff, and ordering chips in Singapore's name. The beneficial owner test looks through the corporate structure to the ultimate controller.

What likely comes next: BIS adding Alibaba Cloud Singapore, ByteDance Ireland, and similar entities to an expanded version of the entity list — not because of specific violations, but as a prophylactic measure to eliminate ambiguity. The June 1 guidance creates the legal authority; entity list additions provide enforcement clarity.

For developers and DevOps teams working in APAC, the practical question is simpler: if your company has Chinese ownership at any level, verify your cloud provider and hardware procurement chain against the new guidance before making your next GPU infrastructure commitment. The compliance window for remediating existing deployments is likely measured in months, not years.

Cloud infrastructure built for AI training and inference in Singapore and Malaysia is about to go through a significant ownership and compliance audit cycle. That affects compute availability, potentially pricing, and definitely the speed of provisioning for any project in that region with complex ownership structures.

Key Takeaways

• Beneficial owner rule confirmed — BIS guidance: export controls on Nvidia Blackwell, Rubin, and AMD MI350x apply to any company with a Chinese parent, regardless of subsidiary location
• Hundreds of thousands of servers — SCMP source estimate of hardware that flowed through the loophole over approximately 12 months
• Pass-through jurisdictions — Malaysia, Singapore, UAE, Netherlands were the primary routes
• Cloud providers at risk — AWS, Google Cloud, Azure in APAC must audit existing enterprise tenants for Chinese parent ownership structures
• Ownership threshold unclear — BIS has not published a specific percentage; the OFAC 50% rule is the likely model
• For developers: if your company has Chinese investment or ownership and you're using cloud compute in APAC, conduct an export control compliance review before your next GPU infrastructure commit
• What to watch: BIS entity list additions targeting Chinese-parented companies in Singapore and Malaysia; hyperscaler customer audit announcements in Q3 2026

Sources

• Al Jazeera: US says ban on AI chip shipments applies to Chinese firms outside China (June 1, 2026)
• CNBC: US takes step to halt Nvidia AI chip shipments to Chinese firms outside China (May 31, 2026)
• TrendForce: US moves to block AI chip exports to overseas Chinese units (June 1, 2026)
• Tom's Hardware: US closes loophole that allowed Chinese-owned subsidiaries outside China to buy AI chips
• South China Morning Post: US labels Chinese chips an economic threat; sources estimate hundreds of thousands of servers shipped through loophole