Amidst the white-hot global competition in AI computing power, Huawei is about to launch its new generation AI chip, the Ascend 910D. This move is seen as a significant milestone for China’s semiconductor industry in breaking through the US technology blockade and building an independent innovation ecosystem. This article will comprehensively analyze this highly anticipated domestically produced AI chip.
Independent Breakthrough Under US Blockade #
The research and development process of Huawei’s Ascend 910D is inextricably linked to the broader context of Sino-US technological rivalry in recent years. Since the US added Huawei to its Entity List in 2019, the company has faced unprecedented challenges in obtaining advanced semiconductor technology and manufacturing capabilities. Particularly in the field of artificial intelligence computing, the US Department of Commerce further tightened export controls on AI chips to China in October 2023, not only prohibiting Nvidia from selling its flagship A100 and H100 GPUs to China, but also imposing a ban on the H20 chip, which was specifically “downgraded” for the Chinese market.
This series of restrictive measures has directly impacted the development of China’s AI industry. According to industry analysis, Nvidia GPUs currently hold more than 80% of the global data center AI acceleration chip market. Chinese technology companies have long relied on these imported chips for training large-scale AI models. The export controls have led to a severe “computing power hunger” among domestic AI companies, forcing companies like ByteDance and Baidu to adjust their large model development plans, with some projects even being delayed.
It is precisely under this urgent “chokepoint” situation that Huawei has accelerated the research and development process of its independent AI chips. The Ascend 910D, as the latest product in the Ascend series, is Huawei’s third-generation high-performance AI processor, following the 910B and 910C. Notably, in the five years since facing the US technology blockade, Huawei has built a complete technology stack from chip design and advanced packaging to software ecosystem. According to internal sources, the Ascend 910D took only 18 months from project initiation to the sample stage, setting a new record for Huawei’s chip development speed.
Changes in the international industrial environment have had a dual impact in this process. On the one hand, US export restrictions have forced Chinese technology companies to accelerate independent innovation; on the other hand, Huawei has also had to cope with the predicament of not being able to use advanced process foundries like TSMC. To solve this problem, Huawei has deepened its cooperation with domestic semiconductor companies such as SMIC, seeking breakthroughs in chip architecture design and packaging technology. This “forced innovation” model, while increasing R&D difficulty, has also enabled Huawei to forge a technological path different from Western giants.
Ascend 910D – Performance Breakthrough Targeting Nvidia’s H100 #
As the latest flagship product in Huawei’s Ascend series, the Ascend 910D has set its sights on the ambitious goal of surpassing Nvidia’s H100 in technical specifications. According to multiple sources, the chip is expected to provide the first batch of engineering samples to selected customers in late May 2025, and after performance verification, it is planned to enter mass production in the fourth quarter of the same year.
According to an exclusive report by The Wall Street Journal, the Ascend 910D will completely reconstruct its architecture design:
Da Vinci Architecture 3.0: Adopting 3D Cube technology, a single chip integrates 64 AI Cores, increasing computing density by 200%.In-Memory Computing Breakthrough: Self-developed HBM3e high-bandwidth memory uses 3D stacking technology, achieving a bandwidth of 4TB/s, surpassing the H100’s 3.35TB/s.Photonic Interconnect Technology: Using silicon photonics modules to achieve ultra-high-speed interconnection between chips, reducing latency to the nanosecond level.
This architectural innovation allows the Ascend 910D to achieve a theoretical peak computing power of 1.2 PFLOP/s, a 78% increase compared to the H100’s 672 TFLOP/s. More crucially, its energy efficiency ratio is optimized to 2.1 TFLOP/W, a 23% improvement over the H100’s 1.7 TFLOP/W.
From a manufacturing process perspective, the Ascend 910D is very likely to be manufactured by SMIC using an improved 7nm process. Although this process node lags behind the 3nm process provided by TSMC for Nvidia’s latest GPUs, Huawei has successfully narrowed the performance gap caused by the process difference through architectural innovation and system-level optimization.
Due to the inability to obtain the most advanced semiconductor manufacturing process, the performance of the Ascend 910D per unit power may still be inferior to Nvidia’s H100, which uses TSMC’s 4nm process. To compensate for this shortcoming, Huawei’s engineering team has innovated at both the chip and system levels: optimizing voltage-frequency curves and task scheduling algorithms at the chip level; and developing advanced liquid cooling solutions at the system level to enable high-density computing clusters to operate stably.
Huawei has also built a complete MindSpore AI framework and CANN (Compute Architecture for Neural Networks) heterogeneous computing architecture for the Ascend 910D. These software tools can fully leverage hardware performance and support the migration of models from mainstream deep learning frameworks such as TensorFlow and PyTorch.
In addition, unlike the “one-size-fits-all” international general-purpose GPUs, a major selling point of Huawei’s Ascend 910D is its specialized optimization for localized scenarios such as Chinese NLP (Natural Language Processing), intelligent driving, and smart cities. For example, in Chinese large model training, the 910D’s architecture has undergone hardware-level acceleration for tasks such as Chinese character processing and polysemy understanding, which forms a unique differentiated advantage in the global AI chip market.
Reshaping the Global AI Computing Power Landscape #
The launch of the Ascend 910D is not only related to Huawei’s own business development but may also have a profound impact on the AI industry landscape in China and even globally. This chip represents the most powerful competitive attempt by a non-US AI computing platform to date, and its success will largely determine the possibility of future diversification in the computing power market.
Supply chain autonomy is the most direct industrial impact. Against the backdrop of the US continuously tightening semiconductor export controls to China, the R&D progress of the Ascend 910D demonstrates the potential of Chinese technology companies to achieve breakthroughs in key core technologies. According to supply chain sources, Huawei has replaced most of the components of the 910D with domestic alternatives, including chips manufactured by SMIC, advanced packaging services provided by JCET, and supporting power management ICs produced by domestic suppliers. If this vertical integration model is successful, it will significantly enhance the ability of China’s high-tech industry to cope with international supply chain risks.
Diversification of technical standards is another potential impact. Currently, the global AI computing ecosystem is almost entirely built around Nvidia’s CUDA architecture. While this single technical path lowers the development threshold, it also leads to serious “vendor lock-in” issues. The rise of Huawei’s Ascend series with the MindSpore framework provides another option for the market, especially appealing to governments and large enterprises concerned about the risks of technological dependence. Industry analysts point out that even though the Ascend platform currently accounts for only a small portion of the global AI computing power, its existence has already prompted Nvidia to accelerate innovation and improve customer service.
China’s AI industry will be the most direct beneficiary of the 910D. As US restrictions hinder the import of high-end AI chips, many large model R&D projects in China face computing power bottlenecks. Huawei is expected to deliver more than 800,000 Ascend series chips to the domestic market in 2025, which will effectively alleviate the computing power shortage. It is particularly noteworthy that the 910D is not an isolated product. Huawei has launched the CloudMatrix 384 computing system, which integrates 384 910C chips and outperforms systems composed of 72 Nvidia Blackwell GPUs in certain specific workloads. The localization of such large-scale computing clusters provides infrastructure guarantees for Chinese companies to train large models with hundreds of billions of parameters.
The global market landscape may therefore undergo subtle changes. Although Nvidia will continue to dominate the global AI chip market due to its first-mover advantage and technological leadership, Huawei is expected to become a major supplier in China, the world’s second-largest economy. According to The Wall Street Journal’s analysis, if the Ascend 910D can reach the performance level of the H100, the US government’s current chip export controls to China will be greatly weakened. In the longer term, Huawei’s success may inspire more non-US companies to invest in AI chip R&D, further promoting market competition and technological diversification.
Uncertainties Amidst Opportunities #
Although the launch of the Ascend 910D marks a significant step forward for China in the field of AI chips, Huawei still faces multiple challenges in achieving its ambitious goals. These challenges come from both the technical level and involve the market ecosystem and the international competitive environment, which will directly affect the commercialization process and long-term development prospects of the 910D.
Process technology limitations are the most fundamental technical bottleneck. Due to US export controls, Huawei cannot obtain ASML’s most advanced extreme ultraviolet (EUV) lithography machines, resulting in its chip manufacturing partner SMIC only being able to use deep ultraviolet (DUV) equipment to produce the 7nm process 910D. This creates a generational gap compared to TSMC’s 3nm process used by Nvidia’s Blackwell GPUs, directly affecting chip transistor density and energy efficiency. Industry analysts point out that although Huawei has partially compensated for this disadvantage through architectural innovation and advanced packaging technology, the 910D may still struggle to compete with international leading products in edge computing and mobile scenarios that require high energy efficiency.
Building a software ecosystem is another major challenge. Nvidia’s CUDA platform has developed a vast ecosystem of millions of developers over more than a decade, which is its strongest competitive barrier. In contrast, Huawei’s MindSpore framework, although developing rapidly, still has significant gaps in toolchain completeness, community activity, and third-party support. Especially for companies that have already developed complex AI applications based on CUDA, migrating to the Ascend platform requires significant time and resources for code refactoring and performance optimization. Huawei is accelerating ecosystem development by expanding developer training, providing migration toolkits, and establishing industry alliances, but this process is destined to require long-term investment.
International market expansion faces political resistance. Against the backdrop of escalating Sino-US technological competition, the acceptance of Huawei’s products in major markets such as Europe and the United States is significantly affected by geopolitical factors. Even if the Ascend 910D is technically competitive, many international customers may still prioritize solutions from Nvidia or AMD due to supply chain risk considerations. Huawei currently focuses its international market efforts on regions such as the Middle East, Southeast Asia, and Latin America, where political sensitivity to the source of technology is relatively low, but the overall scale and purchasing power are limited.
Supply chain stability also has hidden worries. Although Huawei has made significant progress in chip design, China’s semiconductor industry still relies on imports for key links such as equipment, materials, and IP cores. The US Department of Commerce is constantly updating its export control list, which may further restrict Huawei’s access to necessary technologies and components. For example, key materials and equipment used for advanced packaging have recently been included in the control scope, which will pose challenges to the large-scale mass production of the 910D.
The rapidly evolving competitive environment increases market uncertainty. Just as Huawei is striving to catch up with Nvidia’s H100, the latter has already released more advanced Blackwell architecture B200 and B300 GPUs and plans to launch the next-generation Rubin platform in 2026. It is reported that the Rubin GPU’s FP8 training performance will reach approximately 8300 TFLOPS, about twice that of the B200. This rapidly iterating technology race means that Huawei needs to continuously increase its R&D investment to avoid being left behind again.
Despite facing many challenges, the Ascend 910D still has broad development prospects. The scale advantage of the Chinese market provides a solid backing for Huawei – even if it only meets domestic demand, it is enough to support the sustainable development of its AI chip business. Policy support continues to increase, and the Chinese government has listed AI chips as a key development area, providing comprehensive support in areas such as R&D investment, talent cultivation, and procurement policies. The effect of technological accumulation is gradually emerging. Through the development of three consecutive generations of Ascend products, Huawei has established a complete technological system and talent team, and this accumulation will accelerate future innovation.
From a more macro perspective, the Ascend 910D represents an important attempt under the trend of global technological multipolarity. Outside the US-dominated AI computing power ecosystem, Huawei is striving to build an alternative technological system, which provides more choices and balance for the global technology industry. Regardless of its ultimate commercial success, this effort itself has already changed the competitive landscape of the global AI chip industry and may foster a more diverse and healthy path for technological development in the future.