DeepSeek, a prominent Chinese AI company, faced unexpected delays in releasing its R2 model due to persistent technical difficulties encountered with Huawei’s Ascend chips. Chinese authorities had encouraged DeepSeek to utilize domestic processors over Nvidia’s H20 systems.
Despite on-site assistance from Huawei engineers, DeepSeek was unable to complete a successful training run using Ascend chips. Consequently, the company relied on Nvidia hardware for training processes while deploying Ascend chips for inference tasks. This dual-hardware approach underscores the challenges in fully transitioning to domestic AI infrastructure for complex operations.
The R2 model’s launch, initially scheduled for May 2025, was postponed. This delay stemmed from the unresolved technical obstacles associated with the Ascend chips and extended data labeling requirements for the updated training dataset. DeepSeek founder Liang Wenfeng reportedly articulated dissatisfaction with the model’s development progress, indicating the necessity for additional time to ensure the model could maintain DeepSeek’s competitive standing.
Competitors have leveraged DeepSeek’s development setbacks. Alibaba’s Qwen3, for instance, has integrated core training algorithms similar to DeepSeek’s, while simultaneously enhancing efficiency and flexibility in its own systems. This development highlights the rapid evolution within AI ecosystems and the potential for rivals to capitalize on the difficulties encountered by individual startups.
Beijing’s broader strategic initiative for AI self-sufficiency has intensified pressure on domestic firms to adopt locally produced hardware. This push aims to reduce reliance on foreign technology, particularly from companies like Nvidia. However, the implementation of this strategy has revealed existing disparities in stability, inter-chip connectivity, and software maturity when comparing Huawei’s chips with Nvidia’s established products.
Nvidia has emphasized the strategic importance of maintaining access for Chinese developers, noting that restrictions on technology adoption could adversely affect economic and national security interests. Chinese AI companies are navigating a complex landscape, balancing governmental directives to use domestic hardware with the practical realities of developing and deploying large language models, which often require advanced and reliable processing capabilities.
Despite these challenges, DeepSeek’s R2 model may still see a release in the coming weeks. The model’s eventual performance will likely undergo scrutiny, particularly when evaluated against rivals that have been trained using more mature and established hardware. This situation exemplifies the ongoing tension between national political objectives, available technical capabilities, and the practical demands of real-world AI deployment.
