Semiconductor Update: US Tightens Export Controls on AI Chips to China — Key Takeaways
On February 15, 2025, the US Bureau of Industry and Security (BIS) released its most aggressive revision yet to export controls on advanced semiconductors, lowering the effective performance threshold for AI chips bound for China to a total processing performance (TPP) of 3,200 or higher — a 33% reduction from the previous cap of 4,800 TPP. This update follows a pattern of escalating restrictions since October 2022, when the US first blocked the export of NVIDIA A100 and H100 GPUs to China, and signals a structural tightening that is reshaping the global semiconductor supply chain.
The new rules close existing loopholes, expand “deemed export” coverage to Chinese nationals working at foreign firms, and add 22 Chinese entities — including subsidiaries of 华为技术有限公司 (Huawei Technologies Co., Ltd., Huáwéi Jìshù Yǒuxiàn Gōngsī) — to the Entity List. For foreign executives evaluating China market entry, the implications are immediate: hardware procurement timelines have doubled, compliance costs have surged, and the risk profile for joint ventures involving 先进制程工艺 (advanced node processes, xiānjìn zhìchéng gōngyì) has fundamentally changed.
What Changed: The New Export Control Rules Explained
The February 2025 BIS revision is not a single rule but a package of three coordinating amendments: a revised Performance Parameter Threshold, an expansion of the Foreign Direct Product Rule (FDPR), and a tightened “footprint” clause that extends US jurisdiction over any chip designed with American electronic design automation (EDA) tools. The most impactful change is the threshold reduction: any chip with a TPP of 3,200 or higher — or a performance density exceeding 5.92 TOPS per square millimeter — now requires a BIS license for export to China.
To put this in context, the October 2023 rules covered chips with a TPP of 4,800 or higher, while the original 2022 rules only covered chips exceeding 4,800 TPP on a different measurement metric. The new threshold effectively captures NVIDIA’s full H100 line (which has a TPP of approximately 7,900 under the new metric) as well as the upcoming B200 chip. For comparison, the H800, a “China-compliant” chip designed after the 2022 rules, had a TPP of approximately 4,900 — previously legal but now captured under the new filter.
The FDPR expansion is equally consequential. Previously, the rule applied only to chips fabricated at facilities with US-origin equipment. Now, it also applies to chips designed using US-origin EDA software — meaning that any chip designed with Cadence, Synopsys, or Siemens EDA tools (which dominate over 85% of the global market) is subject to US export control, regardless of where the fabrication takes place.
Impact on China’s AI Chip Ecosystem
The immediate effect on China’s 人工智能 (artificial intelligence, réngōng zhìnéng) semiconductor ecosystem has been severe. Chinese AI chip designers like 寒武纪科技 (Cambricon Technologies, Hánwǔjì Kējì) and 壁仞科技 (Biren Technology, Bìrèn Kējì) rely on TSMC’s advanced nodes (7nm and 5nm) for their highest-performance chips. With TSMC now required to obtain BIS licenses for any “advanced node” production destined for Chinese AI chip designers — and those licenses effectively denied since late 2023 — these companies face a 12-18 month production gap.
Huawei’s in-house chip design subsidiary 海思半导体 (HiSilicon, Hǎisī Bàndǎotǐ) has been the most resilient, having stockpiled significant wafer capacity at TSMC before the 2023 restrictions and maintaining limited production via SMIC’s 7nm-class N+2 process. However, SMIC’s 中芯国际集成电路制造有限公司 (Semiconductor Manufacturing International Corporation, SMIC, Zhōngxīn Guójì Jíchéng Diànlù Zhìzào Yǒuxiàn Gōngsī) yields on 7nm remain at approximately 40%-50% compared to TSMC’s 95%+, making high-volume production of competitive AI chips economically challenging.
For foreign companies with operations in China, the rules create a dual-speed compliance burden. Hardware procurement teams must now verify not only the TPP of imported chips but also the supply chain provenance of every subsystem. Since the FDPR now captures chips designed with US EDA tools, Chinese contract distributors face a 4-6 week delay in clearing customs for any high-performance computing (HPC) equipment, compared to 1-2 weeks before the rule change.
Strategic Responses and Compliance Imperatives
Foreign executives must adapt their China AI hardware strategy along three dimensions: sourcing, legal structuring, and technology isolation. On sourcing, the window for purchasing pre-licensing inventory has closed — BIS’s new “prior authorization” requirement for any chip meeting the revised threshold means that spot-purchasing of NVIDIA H100s or equivalent products is no longer viable. Companies should instead evaluate domestically-chinese alternatives such as 华为昇腾910 (Huawei Ascend 910, Huáwéi Shēngténg 910), which now receives strong government procurement preference in state-linked projects.
On legal structuring, the expanded deemed export rule — which requires US and foreign companies to obtain licenses before allowing Chinese nationals (including permanent residents) to access controlled technology — now applies to software, toolchains, and SDKs. Any foreign company with a China R&D center that works on AI model training or optimization using US-origin tools must now implement separate clean-room facilities or face potential penalties under IEEPA. The cost of non-compliance can reach RMB 50 million or $7 million in fines per violation, plus reputational damage.
Technology isolation is the most operationally complex response. Companies should segment their China operations into two distinct technology stacks: one for China-local AI development using approved hardware (Huawei Ascend, Cambricon MLU series) and one for global development using unrestricted hardware. This separation requires separate IT infrastructure, separate data storage, and separate personnel access — a restructuring that typically takes 6-9 months and costs an estimated RMB 15-20 million for a mid-sized R&D center.
| Parameter | October 2022 (Original) | October 2023 (Revision 1) | February 2025 (Revision 2) |
|---|---|---|---|
| Performance threshold | TPP > 4,800 (original metric) | TPP > 4,800 (revised metric, covering H800) | TPP > 3,200 + performance density > 5.92 TOPS/mm² |
| FDPR coverage | Equipment only (US-origin fab tools) | Equipment + “any US technology” | Equipment + EDA tools + “any US software” |
| Entity list additions | 28 entities (Huawei, SMIC) | 13 additional entities | 22 additional entities (incl. HiSilicon subsidiaries) |
| Impact on accessible NVIDIA chips | H800 compliant | H800 restricted | All NVIDIA chips except repurposed L4s |
| License processing time | 4-6 weeks (presumption of approval for non-military) | 8-12 weeks (presumption of denial for China) | 12-16 weeks (presumption of denial — only partner-country exceptions) |
NEXT STEPS
- Audit your AI hardware supply chain. Review all outstanding purchase orders and in-transit shipments against the new TPP threshold and FDPR rules. Use our Supply Chain Risk Audit Guide to identify exposure within 48 hours.
- Evaluate Chinese AI chip alternatives. Schedule a technical evaluation of the Huawei Ascend 910B or Cambricon MLU370 against your specific AI workloads. Read our China AI Chip Comparison 2025 for performance benchmarks and procurement timelines.
- Restructure your China R&D operations. Implement the dual-stack technology isolation framework described above. Our R&D Compliance Restructuring Blueprint provides a 12-week implementation plan with cost estimates and legal checkpoints.
— China Gateway 360 —
Remote China market entry support, built around execution.
