DeepSeek has been forced to abandon training its next-generation R2 artificial intelligence model on Huawei processors after repeated technical failures, reverting to Nvidia hardware and pushing back its launch schedule. The Chinese AI developer had initially sought to use Huawei’s Ascend chips in a move aligned with Beijing’s push for technological self-sufficiency, but persistent instability and software issues ultimately derailed the plan.
Technical hurdles undermine domestic hardware push
According to company sources, attempts to train the R2 model exclusively on Ascend hardware failed repeatedly, even with Huawei engineers on-site to troubleshoot. Problems included inter-chip communication bottlenecks, unstable operation under high workloads, and an immature software ecosystem compared with Nvidia’s well-established CUDA platform. While Huawei’s processors will still be used for inference tasks, the compute-heavy training has now shifted back to Nvidia’s H20 GPUs.
Launch delay weakens competitive position
The R2 model had been scheduled for release in May, but the hardware issues, combined with slower-than-expected data labelling, have postponed its debut. This setback has allowed rivals such as Alibaba’s Qwen series and Baidu’s ERNIE to consolidate their lead in China’s AI market, further intensifying competition in a sector where innovation cycles are measured in months.
Geopolitical backdrop complicates technology choices
DeepSeek’s initial decision to use Ascend chips was widely seen as a politically strategic choice, supporting China’s ambition to reduce dependence on U.S. technology. However, the company’s retreat to Nvidia highlights the current gap between Chinese and U.S.-made AI chips, particularly in training performance and developer ecosystem maturity. The episode also underscores the difficulty of balancing political directives with commercial realities in a fast-moving global AI race.
China’s chip challenge remains unresolved
The incident raises questions about how quickly China can close the performance gap with global leaders like Nvidia. While Huawei’s Ascend line has made progress, industry analysts say significant advances in software optimisation, chip interconnect speeds, and reliability are still required before they can compete at scale in training frontier models. For now, DeepSeek’s compromise—using Ascend for inference while relying on Nvidia for training—reflects a pragmatic approach in a highly competitive market.
REFH – Newshub, 15 August 2025
Recent Comments