Huawei is reportedly making significant strides in the AI market by collaborating with DeepSeek to integrate its Ascend series with budget-friendly AI models. This strategic partnership aims to disrupt the market dominated by NVIDIA, promising substantial cost savings and competitive performance.
GPU Capabilities and Compute Power
The centerpiece of this collaboration is an ‘all-in-one machine’ powered by Huawei’s Ascend 910B and 910C chips. The Ascend 910B is built on a 7nm process, while the Ascend 910C utilizes SMIC’s N+2 process. Reports indicate that the Ascend 910C can deliver up to 320 TFLOPS of FP16 performance, achieving 60-70% of the performance of NVIDIA’s H100. This could be a game-changer for enterprises looking for high-performance AI solutions without the hefty price tag.
Advanced Integration Techniques
According to industry sources, the Ascend 910C matches NVIDIA’s H100 performance by packaging two 910B chips together using advanced integration methods. This innovative approach not only enhances performance but also underscores Huawei’s engineering prowess in the AI chip sector.
Distributed Architecture and Storage
The machine boasts a distributed architecture, pairing either the Ascend 910B or 910C with the Kunpeng 920 CPU. It is equipped with NVMe SSD memory and offers up to 16TB of storage per unit. This robust configuration is designed to meet the demanding requirements of AI workloads, ensuring quick data access and ample storage capacity.
Cost and Pricing Advantage
Huawei’s all-in-one machines are available in two main product lines: the inference-focused Atlas and the FusionCube A300 DS Edition. The Atlas units, preloaded with DeepSeek’s R1 models, start at around RMB 300,000 to 500,000 for the 32B model. The high-end 671B version ranges from RMB 3 to 5 million.
For those handling both training and inference, prices for the FusionCube A300 DS Edition start at approximately RMB 2 million, reaching up to RMB 10 million. Despite these figures, they still offer a 60–70% cost saving compared to NVIDIA’s H100 solution, which is priced at about RMB 20 million.
API and Promotion Strategy
In terms of API pricing, DeepSeek’s V3 model input is priced at just RMB 1 per million tokens, while the R1 model output costs RMB 16 per million tokens. This is significantly lower than OpenAI’s pricing, which stands at RMB 60 per million tokens. To encourage adoption among small and medium-sized enterprises (SMEs), DeepSeek is offering free versions as part of its initial promotion, making their AI solutions more accessible.
Impact on Domestic Localization
The introduction of this all-in-one machine highlights China’s progress in domestic localization. Key players like SMIC, Hua Hong, and YMTC are crucial contributors to the supply chain, reinforcing China’s capabilities in producing competitive AI technologies. This initiative not only supports local industries but also positions Huawei as a formidable competitor in the global AI landscape.
Read More
– Huawei Reportedly Set to Test New Ascend 910D AI Chip as Early as May, Aiming to Challenge NVIDIA
– Huawei’s Ascend 920 Reportedly Set for 2H25: Concerns Grow Over SMIC’s Capability and HBM Access
For more tech updates, follow us on aitechtrend.com.
Note: This article is inspired by content from https://www.trendforce.com/news/2025/04/29/news-decoding-huaweis-deepseek-all-in-one-machine-60-70-of-nvidia-h100-performance-at-an-appealing-price/. It has been rephrased for originality. Images are credited to the original source.