A new East Asia Forum analysis rejects the claim that US semiconductor export controls have failed, arguing that recent advances in Chinese AI models reflect algorithmic adaptation rather than a breakthrough against the restrictions.

The piece, published by the Australian National University-hosted policy journal in Canberra, takes aim at commentary treating capable Chinese models as proof of policy collapse. The author frames such conclusions as conflating visible model output with the underlying compute threshold the controls were designed to defend.

Why is the timing of this argument significant?

The article lands in the third year of US AI-chip restrictions. The Commerce Department's Bureau of Industry and Security (BIS) issued its first sweeping AI-chip rules on October 7, 2022, capping data-center accelerators above specific compute and interconnect thresholds. BIS updated the rules on October 17, 2023, closing loopholes around customized parts and capturing the Nvidia H800, A800, and L40S. The subsequent H20 — designed as a compliant part — was added to the restricted list in 2024.

Chinese AI labs have nonetheless released models trained within those constraints:

• DeepSeek released the R1 reasoning model on January 20, 2025
• Alibaba's Qwen family has expanded across parameter sizes, including open-weight releases
• Baidu continues to update its Ernie series
• Tencent, ByteDance, Moonshot AI, and Zhipu AI operate production model lines

The article's central point: model performance under constrained compute shows engineering efficiency, not parity with restricted frontier chips such as the Nvidia H100, B100, or the forthcoming Blackwell Ultra generation.

What does the article say about policy goals?

The analysis separates two objectives often conflated in public debate:

• Denying Chinese military and surveillance end-users access to frontier compute
• Preserving US and allied leadership in advanced AI hardware

The piece argues the controls can succeed on the first goal while Chinese firms remain competitive in commercial model deployments, because the compute ceiling for many commercial workloads sits below the threshold the rules restrict.

This framing echoes the "small yard, high fence" formulation senior US officials used in 2023 and 2024 to describe the policy's narrow targeting of military and surveillance use cases.

What does the supply-chain picture look like?

The piece reportedly engages with the hardware side of the equation, where the bottleneck sits in fabrication rather than model training. Nvidia CEO Jensen Huang said publicly in 2024 that the controls had accelerated Chinese domestic chip development — a dynamic the article appears to weigh on both sides.

Domestic Chinese accelerators face structural constraints that the article likely addresses:

• SMIC produces advanced nodes without access to ASML's extreme ultraviolet (EUV) systems
• High-bandwidth memory (HBM) sourcing from SK Hynix and Samsung is restricted
• CoWoS-equivalent advanced packaging capacity remains limited inside China
• EDA tool access for leading-edge design remains constrained under US rules

These supply-chain factors, not model releases, will determine whether Chinese AI compute closes the gap with US-controlled frontier hardware.

Forward-looking: The article's thesis will face its sharpest empirical test if Chinese fabs shift from limited advanced-node production to sustained volume, a transition that would alter the supply-chain calculus underpinning current US export policy.