ByteDance’s Offshore Chip Access Exposes Enforcement Gap in U.S. Export Controls
Chinese tech giant secures Nvidia's most advanced Blackwell chips in Malaysia while smuggling networks move $160 million in restricted hardware, revealing systematic failures in Biden-era restrictions.
ByteDance has secured access to approximately 36,000 Nvidia Blackwell B200 chips in Malaysia, circumventing U.S. export controls designed to block Chinese firms from acquiring advanced AI hardware. The offshore deployment, confirmed by The Wall Street Journal, comes as federal prosecutors revealed a smuggling ring that illegally exported Nvidia H100 and H200 chips to China between October 2024 and May 2025—underscoring critical gaps in the U.S. strategy to maintain AI superiority through semiconductor restrictions.
Third-Country Cloud Hosting: Legal Workaround or Policy Failure?
ByteDance is working with Southeast Asian firm Aolani Cloud on approximately 500 Nvidia Blackwell computing systems in Malaysia, representing hardware costs exceeding $2.5 billion. Aolani Cloud has been leasing servers with Nvidia H100 chips to ByteDance in Malaysia since February 2025, establishing a template that exploits jurisdictional boundaries in export enforcement.
The arrangement operates within the technical letter of U.S. regulations. An Nvidia spokesman confirmed that “by design, the export rules allow clouds to be built and operated outside controlled countries” such as China. Aolani stated it “adheres fully to all applicable export control regulations” and that its customers do not own the chips or have claims over them—a distinction that permits Chinese firms to access restricted computing power without violating direct export prohibitions.
ByteDance plans to use the computing power for AI research and development outside China and to meet rising AI demand from its global customers, generating about a quarter of its revenue outside China with sizable AI teams in Singapore, San Jose, and Seattle. The company operates more than a dozen AI apps, including chatbot Dola, video creator Dreamina, and video-generation model Seedance, with five of the world’s 50 most popular AI consumer apps originating from ByteDance.
The Biden administration’s Export Controls, introduced in October 2022 and tightened in subsequent years, aimed to restrict China’s access to advanced AI chips and semiconductor manufacturing equipment. The strategy relied on three pillars: blocking chip exports, denying access to manufacturing tools, and preventing domestic production of critical components. Yet enforcement has consistently lagged implementation, with stockpiling, third-country routing, and cloud rental emerging as persistent evasion vectors.
Smuggling Networks Reveal Systematic Enforcement Failure
Parallel to legal workarounds, illicit networks have moved massive volumes of restricted hardware. Alan Hao Hsu initially agreed to buy over 7,000 H100 and H200 GPUs from Lenovo through his company Hao Global, funded by two unnamed Chinese companies through intermediaries in Thailand, Singapore, Malaysia, China and Taiwan.
The smuggling operation employed sophisticated deception. Shipping labels on Nvidia H100 and H200 chip cartons were changed to bear the name of a fictional brand, “Sandkyan,” to evade export controls. When Hsu attempted to export chips through Atlanta, Georgia, agents from the Commerce Department’s Bureau of Industry and Security intercepted and detained them, forcing smugglers to pivot to alternative routes through New York-area Chinese warehouse networks.
The Center for a New American Security estimates 10,000 to hundreds of thousands of AI chips were smuggled to China in the past year—a range so wide it underscores the opacity surrounding enforcement effectiveness. Banned Nvidia GPUs like the A100 and H100 have appeared in Chinese university research and startup documentation, likely acquired through grey-market channels or indirect resellers in third countries.
Policy Whiplash and China’s Strategic Response
The Trump administration’s December 2025 decision to allow Nvidia to sell H200 AI chips to China in exchange for a 25% revenue cut created immediate market confusion. China approved its first major batch, clearing ByteDance, Alibaba and Tencent to purchase more than 400,000 units in total—before Beijing reversed course.
After Trump formally approved limited H200 exports on January 13, Chinese customs authorities informed Nvidia that the chips would not be allowed to enter the country. Nvidia has reportedly halted production of H200 chips intended for China, reallocating TSMC manufacturing capacity to next-generation Vera Rubin hardware, recognizing it could not remain in regulatory limbo between competing U.S. and Chinese policies.
Beijing’s calculus prioritizes long-term technological sovereignty over short-term capability gains. Washington’s approach is to offer China advanced technology and maintain its chipmakers’ market share, while Beijing’s response is to endure short-term pain while accelerating development of core technologies. Huawei’s Ascend 910B, Cambricon’s Siyuan 590 and Biren Technology’s BR100 are now being used for AI training and inference.
- Third-country cloud hosting provides legal access to restricted chips without triggering export controls—a structural loophole that offshore data centers in Malaysia, Indonesia, and Singapore exploit at scale.
- ByteDance’s $22 billion AI procurement budget for 2026 signals continued investment in offshore infrastructure as primary compute strategy, reducing pressure for direct chip imports.
- Smuggling enforcement remains reactive, with visible seizures representing only a fraction of attempted diversions.
- China’s rejection of approved H200 imports demonstrates strategic patience—accepting temporary AI capability gaps to accelerate domestic alternatives and avoid dependence on reversible U.S. policy.
Acceleration Impact on Chinese AI Development Timeline
Access to Blackwell-class hardware materially advances ByteDance’s competitive position. The computing capacity from permitted H200 exports is sufficient to train models matching or exceeding current American frontier models, per analysis from the Center for a New American Security. This represents roughly twice what Chinese fabs are expected to produce domestically in 2026, twice the capacity of the world’s largest data center, and nearly OpenAI’s entire deployed compute worldwide at the end of 2025.
Offshore deployments eliminate the compute constraint that previously bottlenecked Chinese AI labs. Portions of Alibaba’s Qwen family of models and ByteDance’s Doubao systems’ large-scale training runs are now executed on offshore clusters, where the combination of high-bandwidth interconnects and dense GPU racks is comparable to infrastructure used by leading U.S. AI labs, according to Financial Times reporting.
Huawei will not be able to make a chip more powerful than the H200 for two years, with the Ascend 960 planned for Q4 2027, creating a window where offshore Nvidia access provides Chinese firms a capabilities bridge. With three million H200 exports, the amount of AI computing power acquired by China in 2026 would likely be greater than what China could otherwise make until 2028 or 2029, per Council on Foreign Relations analysis.
What to Watch
Cloud rental regulation: The House Foreign Affairs Committee passed the Remote Access Security Act (369-22) to close the cloud GPU rental loophole, but implementation timelines remain unclear. Whether legislation can effectively restrict computing access when physical chips remain outside restricted jurisdictions will test the enforcement limits of extraterritorial controls.
Southeast Asia infrastructure buildout: Malaysia and Indonesia are emerging as neutral ground for AI compute, with ByteDance reportedly in talks to use more than 7,000 B200 chips at a data center in Indonesia. The scale of regional data center investment by Chinese firms will signal whether offshore hosting becomes the dominant access model or a transitional workaround.
Semiconductor supply chain repricing: Opening Chinese demand by granting export licenses reduces global availability and increases prices outside of China as constrained TSMC capacity gets reallocated. Nvidia customers outside China face delayed Blackwell orders if H200 production resumes for Chinese offshore deployments—creating competitive harm that could trigger legal challenges to the licensing regime.
Chinese domestic chip timeline acceleration: Beijing’s willingness to reject approved H200 imports suggests confidence in near-term domestic alternatives. Huawei’s Ascend roadmap includes the Ascend 950PR planned for 2026, followed by the Ascend 960 in 2027 and the Ascend 970 in 2028—each iteration narrowing the performance gap with Nvidia’s previous-generation architectures. Whether this timeline holds will determine if export controls successfully delayed or merely incentivized China’s chip independence.