Ten Chinese Companies, Including ByteDance, Receive US Approval for AI Chip Imports Amid Export Restrictions
In a development highlighting the complexities of international trade regulations in the artificial intelligence sector, the US Department of Commerce has reportedly issued special export licenses to ten Chinese technology firms. These licenses permit the companies to acquire certain American-made AI chips, despite ongoing stringent US export controls aimed at limiting China’s access to advanced semiconductor technology. Among the recipients is ByteDance Technology (Beijing) Co Ltd, the Chinese arm of ByteDance, the parent company of the popular short-video platform TikTok.
This information comes from sources familiar with the matter, as reported by Reuters. The licenses were granted by the Bureau of Industry and Security (BIS), the entity within the Commerce Department responsible for enforcing export controls. These approvals represent a narrow exception to broader US policies designed to curb the transfer of cutting-edge AI hardware to China, a policy intensified since 2022 under both the Biden and preceding Trump administrations.
Background on US AI Chip Export Controls
US export restrictions on AI chips originated from national security concerns, particularly the potential for advanced semiconductors to enhance military capabilities through AI applications. High-performance chips, such as Nvidia’s A100 and H100 GPUs, are central to training large language models and other AI systems due to their massive parallel processing power, high memory bandwidth, and tensor core architectures optimized for machine learning workloads.
In October 2022, BIS introduced comprehensive rules requiring licenses for exporting advanced chips to China, including performance thresholds based on total processing performance (TPP) metrics exceeding 4800. Subsequent updates in 2023 and 2024 tightened these rules further, blacklisting additional Chinese entities and closing loopholes for server assemblies incorporating restricted chips. Companies like Nvidia responded by developing China-specific variants, such as the A800, H800, and most recently the H20, which are engineered to fall just below the TPP thresholds while retaining substantial AI compute capabilities.
The H20 chip, for instance, features reduced interconnect speeds and memory capacities compared to its unrestricted counterparts. It uses Nvidia’s Hopper architecture with 94 billion transistors, 96GB of HBM3 memory, and a TDP of around 700W, delivering approximately 296 teraflops of FP8 performance for AI inference and training. While powerful enough for many commercial AI tasks, it lags behind the H100’s 4 petaflops in similar metrics, making it a compliant but downgraded option for the Chinese market.
The Licensed Chinese Recipients
The ten companies receiving these licenses operate in sectors critical to China’s AI ambitions, including video processing, autonomous driving, and cloud computing. ByteDance Technology (Beijing) Co Ltd stands out due to its global profile via TikTok and its domestic counterpart Douyin, which rely heavily on AI for content recommendation algorithms powered by recommendation systems involving billions of parameters.
Other recipients reportedly include subsidiaries or affiliates of major players like Shanghai Sensetime Technology Development, a leader in computer vision AI, and affiliates of companies focused on edge AI and data centers. These firms were previously added to the US Entity List, a designation that presumes denial of export privileges unless a specific license is obtained. Placement on the Entity List triggers rigorous review processes, where applicants must demonstrate that the end-use will not support military end-uses or proliferation activities.
The licenses are tailored and time-limited, often specifying exact chip models, quantities, and end-users. They do not grant blanket approval for future purchases, requiring renewals subject to ongoing compliance checks.
The Paradox: Clearance with Strings Attached
Despite securing these licenses, the Chinese firms face a peculiar irony: the approved chips may not meet their performance needs, rendering them effectively “not allowed to accept” in practical terms. The H20 and similar variants, while accessible, deliver significantly reduced capabilities compared to state-of-the-art alternatives available elsewhere. For instance, benchmarks show the H20 achieving about 40-50% of the H100’s training throughput on models like Llama 70B, limiting scalability for frontier AI research.
Moreover, China’s domestic semiconductor ecosystem, bolstered by initiatives like “Made in China 2025,” promotes self-reliance. Huawei’s Ascend 910B, fabricated on a 7nm process by SMIC, offers competitive inference performance and is increasingly preferred by local hyperscalers. Regulatory pressures within China encourage adoption of indigenous hardware, with subsidies and procurement preferences tilting the scales away from US imports even when permitted.
This situation underscores the cat-and-mouse dynamic of the US-China tech rivalry. US controls aim to slow China’s AI progress by denying access to the most advanced nodes (e.g., TSMC’s 3nm), estimated at 5-7 years ahead of domestic equivalents. However, licenses for compliant chips maintain some commercial flow, supporting US firms like Nvidia, which derived billions in revenue from China pre-restrictions.
Industry analysts note that these approvals could signal selective easing for non-sensitive commercial uses, balancing security with economic interests. Nvidia’s fiscal reports indicate continued China sales via compliant products, totaling over $12 billion in the region last year despite curbs.
Implications for Global AI Supply Chains
The granting of these licenses illustrates the nuanced application of export controls. BIS evaluates applications case-by-case, factoring in factors like foreign availability, technology leadership, and human rights concerns. For the approved firms, compliance involves detailed reporting, audits, and restrictions on re-export or disassembly.
From a technical standpoint, reliance on downgraded chips forces optimizations in software stacks. Chinese developers leverage frameworks like Huawei’s MindSpore or adaptations of PyTorch to maximize H20 efficiency, employing techniques such as model quantization, sparse inference, and distributed training across clusters.
Looking ahead, evolving controls may further narrow the gap between permitted and prohibited hardware. Proposed rules target even inference workloads, potentially requiring licenses for chips above lower TPP thresholds.
This episode reveals the intricate balance in AI hardware geopolitics: US approvals enable limited access, yet strategic and performance limitations ensure approved shipments do not fully empower recipients, aligning with policy goals.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.