During his visit to China, Huang Renxun remarked, "It's only a matter of time before Huawei's AI chips replace NVIDIA's." Initially seen as a diplomatic nod to an established competitor, this statement gained significant weight at the 2025 World Artificial Intelligence Conference, where Huawei unveiled its Ascend 384 Super Pod. This announcement signaled the potential arrival of an era dominated by domestic computing power, as China's advancements in computing power chips and large-scale cluster technology have matured to the point of global readiness.

01

Domestic Computing Power Overcomes Challenges

Historically, the development of domestic computing power has been hampered by insufficient single-chip computing power and limitations in large-scale cluster technology. However, significant breakthroughs by companies such as Huawei and Mooresys have addressed these challenges.

The concept of a Super Pod, or "super computing cluster," was pioneered by NVIDIA. Essentially, it involves linking one or several high-performance AI servers as "nodes" through high-speed interconnection networks to form a larger, more powerful "super computing node." This architecture is tailored for the most compute-intensive and complex AI large model tasks. The more nodes connected, the larger the scale and the more it tests the manufacturer's technical prowess. Currently, the Super Pod represents the pinnacle of technical resource allocation for computing power chip manufacturers.

Huawei's Ascend 384 Super Pod stands out as an innovative leader in this field. Comprising 12 compute cabinets and 4 bus cabinets, it achieves the industry's largest-scale 384-card high-speed bus interconnection. Through the MatrixLink high-speed network, 384 Ascend NPUs and 192 Kunpeng CPUs are fully interconnected, forming a single-node "super AI server." Furthermore, the communication bandwidth between any two AI processors within the super pod is 15 times higher than traditional architectures, resolving the issue of low cross-machine bandwidth and enhancing data transmission efficiency within the cluster.

In terms of computing power, the CloudMatrix 384 single-cluster computing power reaches 300 PFlops. Last year, NVIDIA launched a similar cluster system equipped with 72 Blackwell GPUs, including B100, B200, and GB200 models. Based on publicly available test data, Huawei's Ascend 384 Super Pod boasts a total computing power of 300 PFLOPs, while NVIDIA's system offers only 180 PFLOPs. This indicates that Huawei's performance surpasses NVIDIA's by 1.7 times.

Additionally, the 384-card super pod, composed of high-speed network switches, offers robust vertical scalability. The parameter switch supports a cluster size of up to 160,000 cards, with equally impressive horizontal scalability. It can cascade 432 super pods into an ultra-large cluster of up to 160,000 cards, laying a solid foundation for the evolution of larger models in the future.

Regarding inference capabilities, the Ascend 384 Super Pod also excels. Its single-card inference throughput reaches 2300 Tokens/s, enabling rapid analysis and judgment of input data to provide swift inference results. It efficiently supports the inference of large models with mixed experts (MoE), achieving "one card, one expert," and a single super pod can support 384 experts for parallel inference, significantly boosting efficiency.

Huawei is not alone in this endeavor; other manufacturers are also making strides.

Mooresys has introduced the XiYun C600, a new generation of general-purpose GPU based on proprietary core GPU IP, achieving a closed-loop domestic supply chain from chip design to manufacturing, packaging, and testing. Its large-capacity video memory, combined with multi-precision mixed computing power, provides secure and controllable computing power support for the digital economy. Mooresys' product line spans chips, boards, servers, and multi-form factor super pods up to thousand-card clusters, offering a full-link solution from underlying hardware to data center deployment. It has collaborated with over a hundred partners to build a comprehensive industrial ecosystem encompassing chip design, manufacturing, packaging, testing, system integration, and data centers. The company has launched over 40 standard rack servers (suitable for varying AI training, inference, and general computing needs), nearly 10 workstations/all-in-ones (optimized for scientific research and design fields, supporting rapid deployment), and innovated various super pod forms such as optical interconnection super pods (16-64xGPU) and Yaolong 3D Mesh super pods (32/64xGPU).

Hengwei Technology, in partnership with Suanneng, has introduced the industry's first domestic TPU orthogonal architecture super pod—the AS9000 series orthogonal architecture 128 AI super pod. Through orthogonal coupling technology of compute nodes and switch nodes, it innovatively eliminates internal cable connections, surpasses the physical limitations of traditional architectures, achieves ultra-high-density computing power integration and ultra-large-capacity Scale Up expansion, significantly reducing AI chip communication latency and enhancing inference efficiency. This product deeply integrates domestic CPUs and Suanneng TPU chips at the hardware level, employs a hybrid air-liquid cooling solution (liquid cooling for efficient temperature reduction of core computing power modules), and ensures sustained high-performance output.

It is evident that domestic AI computing power has transcended large-scale cluster technology, not only providing robust computing power support but also initially possessing capabilities for large model training.

02

Domestic Computing Power Embarks on Global Expansion

With technical shortcomings addressed, domestic computing power is "fully armed" for global expansion, gaining momentum through the dual impetus of policies and enterprises.

China has unequivocally stated that "artificial intelligence can be an international public good that benefits mankind," emphasizing the bridging of the global intelligence gap through technology sharing, with particular focus on capacity-building needs in global southern countries (developing countries). This proposition transcends traditional geopolitical competition and elevates technological cooperation to the level of common human development. Unlike the technology blockade model, China fosters cooperation through an "open-source ecosystem + standard collaboration" paradigm: promoting the open-source of large models like DeepSeek to lower the technical threshold; leading the establishment of the "China-BRICS AI Development and Cooperation Center" to facilitate cross-border technology transfer; and planning to build over 10 AI joint laboratories in ASEAN and the Middle East to form localized technology empowerment nodes.

Beyond policies, enterprises have spearheaded their global expansion plans. Since the beginning of this year, FT-2000 has accelerated its "general computing + intelligent computing" dual-wheel drive strategy, planning to promote the export of domestic chips in two phases during the "15th Five-Year Plan" period: from 2026 to 2028, it will adopt a "companion go to sea" approach, leveraging customers' overseas infrastructure construction projects to deeply embed domestic chips into overall solutions; after 2029, it will rely on the global resources of the China Electronics Corporation and leverage the compatibility advantages of the FT-2000 CPU ecosystem to achieve "active go to sea," fully entering the international market. Dou Qiang, chief scientist of FT-2000, stated that the company will harness high-quality cooperation under the Belt and Road and the SCO Digital Cooperation Mechanism to serve the global market with China's computing power foundation.

Furthermore, in May this year, Malaysian Deputy Minister of Digital Communications Chang Nianqun announced that Malaysia has officially launched a national AI infrastructure strategy, becoming the first country in Southeast Asia to establish a sovereign, fully autonomous, and full-stack AI ecosystem. This project plans to rely on Chinese AI chips and the open-source DeepSeek large model, marking the first large-scale deployment of Chinese AI chips overseas. However, this plan garnered the attention of the United States. White House AI advisor Sacks commented on social media, "As I have long warned, China's complete AI system has emerged." Subsequently, Chang Nianqun's office staff responded to media inquiries by stating that her relevant remarks would be withdrawn without explanation. Currently, it is uncertain whether the project will proceed as planned, and relevant Chinese AI chip companies have also affirmed that they have not sold chips in Malaysia, nor has the local government procured related equipment.

03

Advantages of Domestic Computing Power's Global Expansion

The advantages of China's computing power going global lie not solely in technological breakthroughs but also in the unique competitiveness of its models and industrial chains.

Firstly, the capability for full-stack solutions is increasingly maturing. With the popularity of the DeepSeek large model earlier this year, domestic computing power chips have achieved large-scale adaptability. In the open-source community, Chinese forces have become key participants in artificial intelligence and open-source large models. The technological output promoted by China is not merely the export of equipment but the construction of a "replicable technological ecosystem." By providing a comprehensive set of solutions, including algorithm optimization and talent cultivation, it aids local regions in establishing complete capabilities from data collection to model training. This "teach a man to fish" model is paving broader avenues for Chinese AI technology to go global, with a potent demonstration effect. However, NVIDIA is also adopting a similar model to compete for the market, as evidenced by Vietnam's agreement with NVIDIA to cooperate on the construction of AI R&D centers and data centers, intensifying competition.

Secondly, the advantages of the entire industrial chain support scenario implementation. China boasts an industrial foundation for the entire AI industry chain, with inherent advantages in vertical application fields and end-side + Agent implementation. By promoting complete solutions, it can more precisely meet market demands: on the 2C side, Alibaba's Quark AI glasses and the AI toy prototype showcased at the CES exhibition earlier this year have entered the sales stage, potentially becoming new commercial breakthroughs; on the 2B side, "out-of-the-box + highly customized" all-in-ones remain the industry standard, suitable for multiple fields such as finance and manufacturing; in the realm of embodied intelligent robots, domestic products have begun shipping in large quantities. To this end, StepLeap has collaborated with 10 chip manufacturers and computing power platforms to establish the "Model-Chip Ecosystem Innovation Alliance" to further strengthen industrial chain coordination.

04

The Significance of China's AI Going Global

The export of domestic computing power is not merely a market behavior of enterprises but holds profound significance for national technological development, global competitiveness, and international influence, manifested in the following five aspects:

Firstly, it validates domestic technology and reduces dependence on the West. For a long time, China has relied on overseas vendors like NVIDIA in the field of high-end computing power, with core links such as chip architecture and cluster technology being restricted by others. Breakthroughs like Huawei's Ascend 384 Super Pod surpassing NVIDIA's similar products and Mooresys achieving a closed-loop domestic supply chain demonstrate that domestic technology already possesses international competitiveness.

Secondly, it contends for standard-setting power, such as promoting China's AI ethics framework to become an international rule. Global AI governance is in a rule-forming stage, with some Western countries pushing for technology blockades under the pretext of "security" and attempting to dominate a single standard. Through initiatives like proposing the establishment of the World Artificial Intelligence Cooperation Organization, releasing the "Global Governance Action Plan for Artificial Intelligence," and building the BRICS AI Cooperation Center, China promotes the governance concept of "technology sharing and inclusive development." For instance, China's "AI Ethics Framework" emphasizes "the priority of the right to development," which aligns more closely with the needs of developing countries. If it can serve as a reference for international rules, it will break the monopoly of Western-dominated governance and strive for more discourse power for global southern countries.

Thirdly, it opens up incremental markets and alleviates internal competition pressure. The domestic computing power market is highly competitive, pressuring the profit margins of leading enterprises. However, emerging markets such as Southeast Asia, the Middle East, and Latin America have a strong demand for AI infrastructure construction. According to estimates, the average annual growth rate of AI computing power demand in global southern countries will reach 35% in the next five years, far exceeding the 18% of developed countries. FT-2000's "companion go to sea" and "active go to sea" plans precisely target this incremental market. Through overseas expansion, enterprises can secure broader growth opportunities.

Fourthly, it promotes the Chinese model and amplifies national influence. Unlike the Western approach of "technology blockade plus monopoly premiums," China's collaborative model of "open-source ecosystem coupled with localized empowerment" is more prone to gaining acceptance among developing nations. For instance, the open-source nature of DeepSeek enables small and medium-sized countries to deploy large models without incurring substantial investments, while the ASEAN AI Joint Laboratory fosters the development of thousands of local AI experts. This "unconditional technology sharing" is transforming the international community's perception of Chinese technology.

Fifthly, it establishes offshore data resource pools to bolster domestic model training. The training of large AI models hinges on extensive and varied data, and domestic data falls short in terms of geographical, cultural, and linguistic coverage. Through overseas AI joint laboratories, data centers, and other nodes, China can construct offshore data resource pools that encompass multiple languages (including Southeast Asian languages and Arabic) and diverse scenarios (like Middle Eastern finance and African agriculture). This "data feedback loop" will expedite the global adaptability of domestic large models, fostering a virtuous cycle of "global expansion - data accumulation - model enhancement - further global expansion." The "Great Navigation Era" of domestic computing power has commenced. From technological breakthroughs to model innovations, and from policy advocacy to corporate endeavors, China is engaging in global AI competition and cooperation with an ever-more open mindset. This signifies not merely the export of computing power but also the export of technological confidence, developmental philosophy, and international responsibility. In the coming years, as AI devices powered by "Chinese chips" proliferate across the globe, the advent of a more inclusive and balanced global intelligent era may be accelerated.