Research Note: Huawei Ascend AI Chips
Corporate
Huawei Technologies Co., Ltd. is a Chinese multinational technology company with its headquarters located at Bantian, Longgang District, Shenzhen, Guangdong, China. Founded in 1987, Huawei has evolved from a telecommunications equipment provider to a diversified technology giant with significant investments in artificial intelligence technologies. The company has faced substantial challenges from US sanctions since 2019, which have restricted its access to advanced semiconductor manufacturing technologies and critical supply chains. Despite these sanctions, Huawei has continued to innovate and develop its Ascend AI chip series, positioning these products as alternatives to Nvidia's dominance in the global AI chip market. The Ascend chip development is being led by Huawei's HiSilicon subsidiary, with manufacturing partnerships with Chinese foundry SMIC (Semiconductor Manufacturing International Corporation) following restrictions on accessing Taiwan's TSMC production capabilities. Huawei's resilience in the face of geopolitical challenges has prompted significant government support and investment from Chinese authorities, positioning the company as China's best hope for technological independence in advanced AI computing.
Market
The global AI chip market exceeded $23 billion in 2023 and is projected to grow at a CAGR of 31.2% to reach $165 billion by 2030, creating a significant opportunity for Huawei's Ascend series. In this rapidly expanding market, Nvidia currently dominates with approximately 80-90% market share, particularly in high-performance AI training accelerators, while Huawei's Ascend chips have gained approximately 6% of the Chinese market as of 2024. The vertical AI market, where Huawei is focusing many of its efforts, exceeded $10.2 billion in 2024 and is projected to showcase around 21.6% CAGR from 2025 to 2034, driven by increasing demand for industry-specific AI solutions. North America currently dominates the global vertical AI market with around 40% share, but China's domestic market is growing rapidly, partly due to US export restrictions creating a protected space for local vendors like Huawei. Huawei aims to significantly increase its market penetration in 2025, with plans to produce approximately 100,000 units of the Ascend 910C and 300,000 units of the Ascend 910B chips. Recent reports indicate Huawei is expected to deliver over 800,000 Ascend 910B and 910C chips to customers in 2025, with industry analysts projecting that Chinese companies, led by Huawei, could have a majority share of AI chips in every major category within China by 2030, reflecting a major shift in the global semiconductor landscape.
Huawei's CloudMatrix 384 AI cluster, which incorporates 384 Ascend 910C processors across 16 racks, has recently demonstrated competitive performance against Nvidia's GB200 NVL72 system, delivering approximately 300 petaFLOPS of BF16 computing power—almost double that of Nvidia's flagship system. This advancement has already attracted at least ten Chinese tech companies to adopt the CloudMatrix 384 system, signaling growing market acceptance despite its greater power consumption (approximately 3.9 times that of Nvidia's equivalent). The launch of the CloudMatrix 384 was strategically timed to coincide with new US export restrictions on Nvidia's H20 chips to China, creating an immediate market opportunity that Huawei is actively filling. According to industry analysts, Huawei is expected to ship approximately 1 million Ascend 910C chips in 2025, with increased orders following the Trump administration's export restrictions on Nvidia's products. This trajectory positions Huawei to potentially capture a significant portion of China's AI acceleration market, which has been valued at billions of dollars annually.
Product
Huawei's Ascend AI chip portfolio includes multiple generations and variants designed for different computing scenarios, with the Ascend 910 series serving as the flagship for high-performance data center applications and the Ascend 310/320 series designed for edge computing deployments. The latest iterations include the Ascend 910B, 910C, and the upcoming 910D and 920 models, each offering progressively improved performance characteristics. The Ascend chips are based on Huawei's proprietary Da Vinci architecture, which features three primary computing units: Cube Unit (for matrix operations), Vector Unit, and Scalar Unit, creating a flexible and scalable framework for AI workloads from edge devices to large-scale training clusters. Independent testing has indicated that the Ascend 910B delivers approximately 80% of the efficiency of an Nvidia A100 when training large language models, with Huawei claiming it outperforms the A100 by up to 20% in some specific workloads. The newer Ascend 910C reportedly achieves approximately 60% of the performance of Nvidia's H100 in inference tasks, while the CloudMatrix 384 system uses scale to overcome individual chip limitations, incorporating 384 Ascend 910C processors in an all-optical mesh network to deliver up to 300 petaFLOPS of BF16 computing power.
The upcoming Ascend 910D is currently in testing with Chinese tech companies, with first samples expected by late May 2025, and is designed to more directly compete with Nvidia's H100 in standalone performance. The Ascend 920, announced in April 2025 with mass production expected in the second half of 2025, will utilize SMIC's 6nm process technology and is projected to deliver over 900 teraFLOPS of BF16 performance with 4 terabytes per second of memory bandwidth. Huawei's AI chips utilize the company's proprietary Da Vinci architecture and CANN (Compute Architecture for Neural Networks) software stack, along with the MindSpore AI framework that supports deployment across cloud, edge, and device scenarios while offering compatibility with popular frameworks like TensorFlow and PyTorch. The competitive landscape for AI accelerators includes Nvidia's H100, A100, and H20 GPUs, AMD's Instinct MI300 series, Google's TPU v4/v5, Intel's Gaudi 2/3, and various Chinese competitors such as Biren's BR100, though Nvidia remains the performance leader with the most comprehensive software ecosystem.
Bottom Line
Organizations operating in China's domestic market should strongly consider Huawei's Ascend AI chips for their AI inference workloads, particularly as US export restrictions continue to limit access to Nvidia's high-performance options like the H100. The Ascend 910C offers the best balance of performance and availability for Chinese enterprises deploying large language models and other AI applications, as demonstrated by its adoption by companies like DeepSeek for inference operations. Global enterprises with operations in China will likely need to implement dual-vendor strategies, using Nvidia accelerators where available and Huawei Ascend chips for China-based operations, which may increase development complexity due to different software stacks. Despite individual chip performance that still lags behind Nvidia's flagships, Huawei's improved production yields (now reaching 40%, up from 20% a year ago) and system-level innovations like the CloudMatrix 384, which delivers 300 petaFLOPS of BF16 computing power, indicate that the Ascend line has reached commercial viability and is increasingly cost-effective.
Organizations should closely monitor the upcoming Ascend 910D, expected to begin testing with first samples as early as May 2025, and the Ascend 920, slated for mass production in the second half of 2025, as these chips may deliver performance comparable to Nvidia's Blackwell architecture GPUs and could represent significant convergence in capabilities. The 910D is specifically being positioned to challenge Nvidia's H100 performance, while the Ascend 920 is designed to fill the void left by the US ban on Nvidia's H20 chips in China. While software ecosystem limitations remain a barrier to wider adoption, Huawei's integration of its AI accelerators into complete systems like the CloudMatrix 384 and investments in developer support through its MindSpore AI framework indicate a strategic commitment to building a comprehensive AI hardware and software stack. As organizations evaluate their AI infrastructure strategies for 2025-2026, Huawei's Ascend products deserve serious consideration for deployments in China and potentially other markets where cost-effectiveness is prioritized over raw performance or where US export restrictions limit access to leading alternatives.