US Tightens Export Controls on Huawei AI Chips
In a significant escalation of its technology sanctions against China, the US Commerce Department’s Bureau of Industry and Security (BIS) has publicly clarified that any company integrating Huawei’s Ascend AI processors—specifically the 910B, 910C, and 910D variants—into products or services anywhere in the world could face criminal penalties for violating US export controls. This guidance, issued May 14, 2025, reinforces existing rules that restrict the use of US-origin technology in advanced computing systems supplied to Huawei.
Background on the Export Controls
Under US regulations, any semiconductor “designed with certain US software or technology” or “fabricated using US-origin equipment” requires a license before export to Huawei. Although BIS has not enacted a new rule, officials stress that the guidance is a public confirmation of a longstanding interpretation: mere use of advanced Huawei chips—despite no physical export of US technology at the point of use—constitutes a breach of export control measures.
Scope of the New Guidance
- Controlling Ascend 910 Series: The BIS specifically named the Ascend 910B, 910C, and recently announced 910D chips.
- Global Reach: Any entity—OEM, cloud provider, research lab—deploying these chips risks enforcement action.
- Licensing Requirement: Companies must apply for “hard-to-get” licenses to legally incorporate Huawei’s AI accelerators.
Technical Deep Dive: Ascend 910 Architecture
Huawei’s Ascend 910 family—built on TSMC’s 7nm-class process and leveraging the in-house DaVinci architecture—targets high-end AI training workloads:
- Compute Cores: 7,872 DaVinci AI cores at 1.8 GHz delivering up to 256 TFLOPS of FP16 performance.
- Memory Subsystem: 910B: 32 GB HBM2 at 1.2 TB/s; 910C: 64 GB HBM2e at 1.6 TB/s; 910D: 96 GB HBM3 stacks pushing 2 TB/s bandwidth (SMIC 7nm node).
- Interconnect: High-speed CANN fabric for 8-chip NVLink-style clusters, achieving up to 200 GB/s chip-to-chip bandwidth.
- Thermal Design: Advanced vapor-chamber heatsinks enabling sustained power envelopes of 310 W per chip.
Comparative Performance: Ascend vs NVIDIA Hopper
Benchmark data—both internal and from third-party testbeds—paints a nuanced picture:
- Single-chip FP16: NVIDIA H100 (Hopper) peaks at 1,000 TFLOPS vs. Ascend 910C’s 256 TFLOPS.
- Cluster Throughput: 64-chip Huawei clusters can exceed 64 petaFLOPS on mixed-precision tasks, rivaling NVIDIA’s DGX H100 Superpod on certain large-batch training jobs.
- Memory-bound workloads: Huawei’s higher aggregate bandwidth shows strength in transformer-style models with sequences >8K tokens.
Global Supply Chain and Geopolitical Implications
The guidance arrives amid accelerating efforts by the US, Japan, the Netherlands, and the EU to tighten controls on advanced semiconductor equipment bound for China. Huawei—stripped of full TSMC support—has reportedly accelerated in-house fabs at SMIC’s Beijing campus, integrating domestic lithography tools and sourcing alternative EDA software from Cadence and Synopsys under license. Senior BIS officials warn that foreign integrators risk entanglement in US enforcement actions if they indirectly enable Huawei’s AI hardware deployment.
Expert Perspectives
- Kevin Wolf, Akin Gump: “This is not a new control, but a clear articulation that deploying Huawei’s most advanced chips anywhere likely breaches US rules.”
- John Neuffer, Semiconductor Industry Association: “Global alignment on export rules is vital to maintain the technological edge and ensure supply-chain security.”
- Christophe Savare, ICInsights: “Companies must implement robust compliance programs to track chip provenance and avoid inadvertent violations.”
Future Outlook and Regulatory Landscape
Following the rescission of the Biden-era AI Diffusion Rule on May 14—cited as “too bureaucratic” by the Commerce Department—BIS has signaled an upcoming, more targeted framework to govern AI chip exports. Meanwhile, the US Senate is advancing a semiconductor R&D and domestic manufacturing package exceeding $52 billion, designed to underwrite next-generation nodes and curtail dependency on strategic rivals. On the diplomatic front, Washington is rallying allies to adopt parallel restrictions, aiming to forestall a bifurcated global AI ecosystem.
Latest Developments
On May 20, reports surfaced that a consortium of Gulf state cloud providers paused plans to integrate Huawei clusters, pending formal guidance from their legal counsel. Concurrently, industry sources indicate that China’s Ministry of Industry and Information Technology is fast-tracking an indigenous “AI chip trust list” to incentivize non-US solutions. Observers note that the coming months will test both the technical resilience of Huawei’s in-house supply chain and the effectiveness of multilateral export controls.