TLDR: Chinese AI startup DeepSeek has postponed the launch of its R2 AI model due to persistent technical difficulties encountered while attempting to train it using Huawei’s Ascend chips. This setback has forced DeepSeek to revert to Nvidia chips for the intensive training phase, highlighting the ongoing struggles in China’s push for technological self-sufficiency in advanced semiconductors.
Chinese artificial intelligence firm DeepSeek has announced a delay in the highly anticipated launch of its R2 AI model, a decision primarily driven by significant technical hurdles encountered during its training phase with Huawei’s domestically produced Ascend chips. The postponement, which pushed back the R2 model’s release from its initial target of May 2025, underscores the formidable challenges China faces in its strategic drive to reduce reliance on foreign, particularly U.S., technology for critical AI infrastructure.
Following the successful launch of its R1 model in January 2025, DeepSeek, like many of its Chinese counterparts, was reportedly encouraged by Beijing authorities to integrate Huawei’s Ascend processors into its development pipeline, moving away from Nvidia’s dominant systems. This push aligns with broader national efforts to bolster indigenous technological capabilities amidst escalating U.S.-China tensions and export restrictions on advanced semiconductors.
However, DeepSeek’s attempts to train the R2 model on Ascend chips were met with ‘persistent technical issues,’ according to sources familiar with the matter. These difficulties necessitated a strategic pivot, with DeepSeek opting to use Nvidia chips for the computationally intensive training process, while still aiming to utilize Huawei’s hardware for the less demanding inference tasks—where a trained AI model generates responses. Despite Huawei dispatching a team of engineers to DeepSeek’s offices to provide on-site assistance, a successful training run on the Ascend chip could not be completed.
Industry insiders point to a discernible performance gap, noting that Chinese-made chips, including Huawei’s, still lag behind Nvidia’s offerings in crucial areas such as stability, inter-chip connectivity, and software ecosystem maturity. This disparity is reflected in market preferences; in 2024, Chinese firms reportedly purchased approximately 1 million Nvidia H20 chips, significantly more than the 450,000 Huawei Ascend 910B chips acquired, despite official encouragement to buy domestic alternatives. Furthermore, Huawei is projected to produce only 200,000 AI chips in 2025, highlighting production capacity constraints. Experts estimate Chinese firms are approximately two years behind in chip design and five generations behind in semiconductor manufacturing equipment.
The delay has allowed competitors to gain ground, putting pressure on DeepSeek. Internally, DeepSeek’s CEO, Liang Wenfeng, has reportedly expressed dissatisfaction with the R2’s progress, emphasizing the need to develop an advanced model that can maintain the company’s competitive edge in the rapidly evolving AI landscape. The R1 model, which made a significant impact on the AI sector earlier this year, was predominantly trained on Nvidia H20 chips, a testament to their continued necessity for cutting-edge AI development in China. Beijing has recently intensified its scrutiny, reportedly summoning major tech companies like Tencent, ByteDance, and Baidu to justify their orders of Nvidia H20 chips.
Also Read:
- Huawei Unveils AI Software to Reduce Reliance on Foreign Memory Chips and Boost Inference Performance
- Chinese AI Agents Surge in 2025: A Deep Dive into Their Capabilities and Global Impact
While DeepSeek continues to work with Huawei to ensure the R2 model’s compatibility with Ascend for inference, the current challenges underscore the complex path China navigates in its quest for technological independence in the global AI race.
