I. Introduction and Overview
Amid escalating U.S. technology sanctions, Huawei has emerged as a formidable competitor in the artificial intelligence (AI) semiconductor market with its Ascend AI chip family. This move is more than a commercial achievement; it is a technological response to geopolitical pressure and a key driver of China's semiconductor self-sufficiency goals. This report provides a comprehensive analysis of the development, production, technical capabilities, and future direction of Huawei's Ascend chips. By comparing them to the market leader, NVIDIA, it offers a profound look at their impact on China's AI industry.
A. Key Analysis Summary
Huawei's AI chip strategy focuses on overcoming the limitations of individual chip performance through system-level integration technology. With access to advanced EUV (Extreme Ultraviolet) lithography equipment blocked, Huawei has pioneered an innovative workaround by leveraging its foundry partner SMIC's 7nm process and advanced packaging technologies like 'chiplets' to close the performance gap. 1
In terms of performance, Huawei's Ascend 910B is said to offer 80% of the performance of the NVIDIA A100 at only 30% of the price. 2 Furthermore, the latest Ascend 910C is reported to achieve 60% of the NVIDIA H100's performance in inference tasks, establishing it as a powerful alternative, especially within the Chinese domestic market. 3 Huawei's "quantity-over-quality" strategy compensates for the inefficiency of individual chips through large-scale clustering. For example, a CloudMatrix 384 system composed of 384 Ascend 910C chips can achieve a total AI system performance that is nearly double that of a system with 72 NVIDIA GB200 NVL72 chips. 4
Underpinning this technological advancement is immense financial and policy support from the Chinese government. Through the 'National Integrated Circuit Industry Investment Fund' (the "Big Fund"), the government has injected trillions of won into the semiconductor industry, concentrating support on a few key companies like Huawei and SMIC. 5 These government subsidies offset the high production costs resulting from initial low yields and provide the financial foundation that enables the low selling price of Huawei's chips. 7
B. Strategic Implications
The U.S. sanctions have directly hit Huawei, but they have also had the paradoxical effect of compelling Chinese companies to reduce their reliance on NVIDIA and turn to Huawei's AI solutions. China is accelerating the construction of an independent AI technology stack centered around Huawei's hardware and its proprietary CANN software ecosystem, separate from NVIDIA's CUDA. 8 While it will be difficult to replace NVIDIA's dominant position in the short term, this movement could fundamentally change the global technology supply chain by creating two distinct technological poles.
II. The Ascend AI Chip Portfolio: In-depth Technical and Architectural Analysis
A. Product Overview and Core Development Philosophy
Huawei's Ascend AI chips are a prime example of the company's internal AI chip development capabilities. Since launching its high-performance Ascend series in 2018, Huawei has continuously evolved its products, from the original Ascend 910 to the 910B, and the latest versions, the 910C and 910D. Notably, the Ascend 910C and 910D are not entirely new architectures but rather evolutionary upgrades based on existing chips. This demonstrates Huawei's strategy of maximizing the use of existing resources in a technologically constrained environment.
The core philosophy behind Huawei's AI chip development is to compensate for the limitations of individual chips with system-level integration technology. While market leaders like NVIDIA maximize single-chip performance with cutting-edge fabrication processes, Huawei has adopted an approach of enhancing the overall system's computing power by advancing its architecture for stacking and interconnecting chips. This approach has been acknowledged even by NVIDIA's founder, Jensen Huang, who noted that "AI is a parallel problem, so if a single computer's performance is insufficient, you can just add more computers."
B. Performance Benchmarks: Ascend vs. NVIDIA
1. Chip-to-Chip Performance Comparison
The performance of Ascend AI chips is directly compared to NVIDIA's market-standard products.
- Ascend 910B vs. NVIDIA A100: The Ascend 910B offers similar specifications to the NVIDIA A100 and has become a key alternative for Chinese IT companies. In particular, Chinese AI startup DeepSeek stated that the Ascend 910B chip offers 80% of the performance of the NVIDIA A100 but at only 30% of the price. 2 This means AI models can be implemented at a cost approximately 54% lower than an NVIDIA system with equivalent performance. 2
- Ascend 910C vs. NVIDIA H100 and H20: The latest Ascend 910C is reported to achieve a level of performance similar to NVIDIA's most advanced chip, the H100, and has shown the potential to deliver 60% of its performance in inference tasks. 3 It is also said to not fall behind the NVIDIA H20 chip, which was specifically created for the Chinese market to circumvent U.S. export regulations, making it a lifeline for China's AI industry.
2. Training vs. Inference Performance
Huawei's AI chips show higher performance in the inference stage of AI rather than the training stage. This is because, unlike the immense computational power required for training models, inference focuses on running pre-trained models, which is relatively less technically demanding. Huawei is exploiting this exact point, competing with NVIDIA primarily in the inference market.
3. System-Level Performance Comparison
Huawei's true strength lies not in single-chip performance but in its overall system design. By parallelly connecting 384 Ascend 910C chips, Huawei's CloudMatrix 384 supernode system achieves a computing capacity of 300 petaflops and can offer nearly double the system performance of the NVIDIA GB200 NVL72. 4 Even with its lower power efficiency (2.3 times less efficient per watt and consuming 3.9 times more power than NVIDIA), Huawei is adopting a "quantity over quality" strategy to enable large-scale AI model training.
C. 'Teardown' and Architectural Analysis
1. Manufacturing Process: SMIC's 7nm DUV Process
Huawei's Ascend 910B and 910C chips are produced using SMIC's 7nm (N+2) process. This is the result of SMIC achieving a 7nm-class process with multi-patterning technology using existing deep ultraviolet (DUV) lithography equipment, as the import of extreme ultraviolet (EUV) lithography equipment has been blocked due to U.S. sanctions. While there is a technological gap compared to global tech giants like NVIDIA, AMD, and Intel, who use 4nm or smaller processes, China is making significant progress with this process alone.
2. Advanced Packaging and Chiplet Technology
With its access to advanced fabrication processes restricted, Huawei has found a breakthrough in post-processing technology. Huawei has filed a patent for 'quad-chiplet' technology, which connects four semiconductors together. 1 This technology, which integrates multiple individual chips to enhance performance, is evaluated to be on par with TSMC's CoWoS (Chip-on-Wafer-on-Substrate) technology. 1 Huawei packaged two 910B chips into a single module, effectively doubling computing performance and memory.
This technological choice clearly shows how geopolitical pressure can spur technological innovation. Unable to access advanced process technology due to U.S. sanctions, Huawei shifted its strategy from increasing the performance of a single die to a modular architecture that effectively connects multiple dies. This has created a new competitive landscape where Huawei can achieve performance close to NVIDIA's H100 by utilizing SMIC's mature processes (7nm, etc.) combined with packaging technology alone.
3. Component Supply Chain
The core components of the Ascend chips are also sourced from within China. For example, the Ascend 910C is known to be equipped with Chinese CXMT's 3rd-generation HBM2E memory, which aligns with Huawei's self-sufficiency efforts in smartphone components.
III. Production and Supply Chain Paradox
A. Manufacturing Reality: The Role of SMIC
1. Exclusive Foundry and Production Capacity
SMIC is currently the exclusive foundry for Huawei's AI chips. 6 To circumvent U.S. regulations and increase AI semiconductor self-sufficiency, the two companies have planned to triple their AI chip production next year. SMIC plans to double its 7nm process production capacity by 2026 to meet the demand from its largest customer, Huawei.
2. Yield Challenges and Improvements
SMIC's 7nm process faced initial yield issues. In 2023, reports indicated a low yield in the 20% range for Huawei's smartphone chips, with most products being discarded as defective. 7 However, recent reports suggest that SMIC's 7nm process yield has significantly improved to as high as 70%. 7 This yield improvement is a crucial development as it can more than double the production output without expanding the production line. 7
Table 1: SMIC 7nm Process Yield Change and Strategic Implications
B. The Economic Paradox of Self-Sufficiency
Huawei's AI chips have a strong price competitiveness. The Ascend 910B provides 80% of the performance of the NVIDIA A100 but is only 30% of the price. 2 This aggressive pricing cannot be explained by free-market economics alone. The gap between high production costs and low selling prices is filled by massive subsidies from the Chinese government. 5
Through its "Big Fund," the Chinese government has invested approximately 340 billion yuan (about 60 trillion KRW) in two phases to support the semiconductor industry. 5 This serves as a strategic tool, allowing the government to absorb the losses incurred from low yields and inefficient production processes, thereby enabling Huawei's chips to be price-competitive in the market. 5
This production model is less about maximizing commercial profit and more about a strategic project for national security and technological independence. The government subsidies allow Huawei to implement an aggressive pricing strategy without worrying about profitability, which is a key driver for rapidly seizing market share from NVIDIA's chips in the Chinese domestic market.
Table 2: China's "Big Fund" Semiconductor Investment Status
IV. Geopolitical and Government Support Trends
A. The Backfire Effect of Sanctions
The U.S. government's sanctions have unintentionally accelerated Huawei's self-reliance, creating a paradoxical outcome. After the U.S. banned the export of NVIDIA's H100 chips to China in 2022 and later restricted even the lower-spec H20 chips designed for the Chinese market, Chinese AI companies were forced to seek alternatives. This situation created an opportunity for Huawei's Ascend 910C to emerge as the primary option for the Chinese industry.
This phenomenon reveals a side of geopolitical competition that cannot be explained by market logic alone. The U.S. aimed to slow down China's progress through sanctions, but China succeeded in filling the resulting market vacuum with its own domestic companies. As a result, the U.S. sanctions have strengthened China's self-sufficiency efforts and solidified its resolve to build an independent technology ecosystem.
B. State-Led Industrial Policy
The Chinese government is driving semiconductor self-sufficiency through a powerful state-led industrial policy. 5
- Big Fund and Focused Support: The 'National Integrated Circuit Industry Investment Fund' is responsible for providing massive subsidies to the semiconductor industry and adopts a strategy of concentrating funds on a small number of successful companies. SMIC and other foundries were a major target of this fund, which has helped them grow into the world's third-largest company.
- Government Mandates for Domestic Use: Chinese provincial governments are implementing policies that mandate the use of domestically produced chips. Shanghai requires 70% of data center chips to be Chinese-made by 2027, while Beijing aims for complete self-sufficiency in the same year. 9 Such policies guarantee stable demand for domestic companies like Huawei and play a decisive role in their rapid market share expansion.
V. The Rise of China's AI Ecosystem
A. Building an Independent Ecosystem: Software is Key
NVIDIA's CUDA platform, built over two decades, has become the standard for AI development and is a key factor in solidifying NVIDIA's dominant position, as much as its hardware performance. 8 Huawei has accurately recognized this and developed its own software toolkit, 'CANN (Compute Architecture for Neural Networks),' to rival NVIDIA's CUDA. 8
Huawei's CANN provides functions similar to NVIDIA's CUDA, but a key difference is its openness. Huawei is pursuing a strategy of making CANN open-source to increase accessibility for more developers and accelerate ecosystem expansion. 8 Furthermore, Huawei is actively recruiting a large number of engineers from NVIDIA to focus on developing software that can challenge the CUDA ecosystem.
B. Industry Adoption and Market Impact
Huawei's AI chips are already being adopted by major AI companies within China.
- The DeepSeek Case: Chinese AI startup DeepSeek successfully utilized Huawei's Ascend 910C for the inference tasks of its large language model (LLM), 'DeepSeek R1.' 10 DeepSeek stated that by porting its model to the Ascend 910B chip, it experienced only a 5% performance loss for the same task while achieving a 70% cost reduction. 10 This shows that Chinese AI companies are adopting a hybrid model, using NVIDIA's high-performance chips (like the H800) for training and Huawei's low-cost chips for inference.
- Widespread Market Adoption: In addition to DeepSeek, major tech companies like Alibaba and new startups like Z.ai, Cambricon, and Byron Technology are also adopting Huawei AI chips, expanding the China-centric AI ecosystem.
VI. Future Outlook and Challenges
A. Future Roadmap and Development Speed
Huawei has laid out a robust AI chip roadmap, including the Ascend 920, with mass production targeted for as early as the second half of this year. 11 This rapid development speed suggests that Huawei is putting all its efforts into closing the technology gap with NVIDIA.
However, the biggest long-term challenge remains overcoming the process technology gap. Huawei's chiplet and advanced packaging technologies are innovative workarounds that navigate current limitations, but they cannot fundamentally replace NVIDIA's latest EUV-based process technology. The process technology gap will continue to exist until China successfully develops its own EUV equipment.
B. Key Challenges
- Power Efficiency Issues: While Huawei's cluster-based systems have gained performance competitiveness, their power efficiency is significantly lower than NVIDIA's solutions. This could lead to increased data center operating costs and pose a constraint on long-term competitiveness.
- Supply Chain Vulnerability: Huawei's reliance on a single foundry, SMIC, makes its supply chain highly vulnerable. Additional geopolitical pressure or production disruptions at SMIC could have a devastating impact on the entire ecosystem.
- Software Ecosystem Maturity: Although CANN is growing rapidly, it will take a considerable amount of time to reach the level of the vast developer community, rich libraries, and stable support that NVIDIA's CUDA, accumulated over two decades, provides. 8
C. Strategic Recommendations for Mass AI Chip Consumers (Research Institutions, Schools, Companies, Governments, etc.)
Ultimately, the Huawei AI chip ecosystem is a strategic alternative born out of geopolitical pressure and technical constraints. When acquiring large quantities of AI chips, institutions such as research centers, universities, companies, and government bodies should formulate their procurement strategies by considering the following perspectives:
- Cost-Effectiveness: Huawei's Ascend chips offer excellent price competitiveness compared to NVIDIA's. 2 For inference tasks, in particular, 10 Huawei's chips can be a practical alternative that significantly reduces costs with minimal performance loss. This is especially important for institutions with limited budgets that need to build large-scale AI infrastructure.
- Supply Chain Stability: The current Huawei AI chip supply chain relies on a single foundry, SMIC, which can be vulnerable to external shocks. 6 Conversely, while NVIDIA utilizes a global supply chain, it is also directly affected by U.S. export regulations. Therefore, when planning AI chip adoption, it is crucial to diversify suppliers and thoroughly assess how geopolitical volatility could impact supply stability.
- Technology and Ecosystem Compatibility: NVIDIA's CUDA has become the industry standard for AI model development and operation. 8 If an organization already uses a CUDA-based system or requires the extensive support of a vast developer community and libraries, NVIDIA's chips might be a better choice. However, Huawei is building its own CANN ecosystem 8 and is accelerating its growth by making it open-source. 8 In the long term, a careful evaluation of which ecosystem better suits the institution's research and business model is essential.
By considering these factors holistically, organizations can find the optimal balance between short-term cost savings and the long-term risk of technological dependency.