Why China Is Phasing Out Foreign AI Chips Despite Their Higher Compute Power?

Why China Is Phasing Out Foreign AI Chips Despite Their Higher Compute Power?

China’s AI Chip Ban: Sovereignty Over Performance as Domestic Alternatives Replace Nvidia, AMD, and Intel in Data Centers

In a significant escalation of its tech self-sufficiency drive, the Chinese government has issued new guidance mandating the use of domestically produced Artificial Intelligence (AI) chips in all new, state-funded data center projects.

This move—reported to include orders for data centers less than 30% complete to remove foreign chips—marks a bold step towards technological decoupling, even as Chinese-made chips are generally understood to offer lower raw computing power than their foreign counterparts.

Foreign chips previously widespread in China include Nvidia H100, H800, A100, A800, H20, AMD MI-series, and Intel Gaudi accelerators—these are set to be removed or banned from new infrastructure investment.

The main domestic replacements are Huawei Ascend 910B/910C, Cambricon Siyuan 590, and Enflame chips, with Alibaba and Baidu also developing in-house chips for their platforms.

China’s “Nvidia Rival” Cambricon Sees 1300% Revenue Surge Benefiting from US and China’s bans

Why the Ban Despite Lower Domestic Computing Power?

China’s decision to ban superior foreign chips like those from Nvidia is primarily driven by geopolitical and national security concerns, outweighing the short-term performance gap. The key reasons include:

Technological Sovereignty and National Security: The most crucial factor. China views reliance on foreign (especially US) technology for its critical infrastructure as a major national security risk. Escalating US export controls, which repeatedly restrict China’s access to the most advanced AI chips (like Nvidia’s H100 and new Blackwell series), have underscored the vulnerability of the foreign supply chain.
Response to US Export Controls: Beijing’s move is a direct countermeasure to Washington’s restrictions, which are justified by the US on the grounds of preventing the chips’ use by the Chinese military. By mandating domestic chips, China aims to remove this vulnerability entirely and insulate its core AI development from foreign policy decisions.
Boosting the Domestic Industry: The mandate provides a guaranteed, massive market for local chipmakers. This influx of public investment is intended to accelerate the development, iteration, and scaling of domestic AI chip technology, ultimately closing the performance gap.
Mitigating Operational Costs: While domestic chips are currently less power-efficient, the Chinese government is reportedly offering substantial electricity subsidies (up to a 50% cut in energy bills) to data centers that use domestic chips. This financial incentive aims to offset the higher operational costs associated with less efficient hardware, making the domestic option economically viable.

Foreign AI Chips Used and Banned

Nvidia: H100, H800, A100, A800, H20—all of which are AI training and inference accelerators formerly used broadly in data centers and for LLM training.
AMD: MI250, MI300 accelerators, used in AI and supercomputing applications.
Intel: Gaudi series (designed for AI workloads), also subject to removal or exclusion under the new policy.

Domestic AI Chip Replacements

Huawei Ascend Series: 910B/910C chips, the most widely promoted as a direct domestic alternative, with growing support in LLM and data center deployments.
Cambricon Siyuan 590: Targeted for efficient AI inference and mid-level training, adopted by some data center projects.
Enflame: Specialized for data center AI, but deployment scale is smaller than Huawei.
Alibaba/Baidu In-house Chips: Both companies have accelerated use of their own AI silicon for part of their cloud and LLM workloads.

UBIOS: China’s Alternative to UEFI and the New Era of Firmware Standards

Performance Comparison Table

Chip Model	Origin	Peak FP16 (TFLOPS)	Main Use	Key Notes
Nvidia H100	USA	~200	AI training/inference	Industry software, top-tier performance
Nvidia A100	USA	~156	AI training/inference	Widespread use before export controls
Nvidia H20, A800, H800	USA	80–120	China-specific, “cut”	Restricted or recently banned
AMD MI300	USA	~150+	Training/inference	Facing export bans
Huawei Ascend 910C	China	~128	Model training/inference	Top Chinese domestic alternative
Cambricon Siyuan 590	China	~120	AI inference/training	Lower ecosystem/training support
Enflame SNS580	China	~80	Cloud/data center AI	Smaller install base, inference focus
Alibaba Hanguang 800	China	~60	Alibaba-specific	Used internally for cloud AI

Feature	Foreign AI Chip (e.g., Nvidia A100/H100 or compliant variants)	Domestic AI Chip (e.g., Huawei Ascend 910B)
Raw Performance (Theoretical)	High – Industry benchmark for speed and throughput.	Lower – Generally considered a generation or more behind the top-tier foreign chips.
Energy Efficiency	High – State-of-the-art power-to-performance ratio.	Lower – May require 30-50% more electricity for equivalent compute tokens.
Ecosystem & Software	Mature – Backed by the CUDA ecosystem, the global standard for AI development.	Developing – Dependent on proprietary or open-source local frameworks; still maturing.
Supply Chain Security	Vulnerable – Subject to US export controls and geopolitical tensions.	High – Domestically controlled and resilient to foreign sanctions.
Market Share (China)	Rapidly declining – Near 0% in new state-funded projects.	Rapidly increasing – Mandated to be 100% in new state-funded data centers.

Domestic AI chip clusters can match foreign single-card performance through “chip stacking,” but this increases energy use and complexity.
Software Ecosystem: The biggest gap is ecosystem and compatibility; Nvidia dominates with CUDA, while China’s open-source and custom approaches are still maturing.

Does China Have Domestic Software Substitutes Amid US Export Ban Risks?

Conclusion

Despite clear performance gaps at the top end, China’s AI chip ban is driven by sovereignty and resilience imperatives, not current hardware parity.

Domestic chips—especially Huawei Ascend—are rapidly improving, with broad national policy support and incentives for further in-house innovation—even at the cost of slowing AI progress during the transition period.

Why China Is Phasing Out Foreign AI Chips Despite Their Higher Compute Power?