Strategic Shift in AI Training: Huawei Ascend Enters Core Development Arena

The competitive landscape of AI infrastructure is evolving beyond surface-level applications. Recent reports from supply chain sources reveal that DeepSeek's upcoming V5 model may be trained on Huawei's Ascend computing platform, marking a significant expansion in the chipset's role within advanced AI development.

From Inference to Training: Expanding Responsibilities

Ascend chips are already familiar components within DeepSeek's existing technology stack, but their function has been primarily limited to inference tasks—applying trained models to real-world scenarios rather than building them from scratch.

"Training demands fundamentally different capabilities," notes an AI hardware engineer familiar with both platforms. "The computational intensity, memory requirements, and distributed coordination needed for model development represent a higher threshold. Moving into this space signals substantial technical progress."

Technical Foundations Laid Earlier

The groundwork for this potential transition became visible last September, when DeepSeek open-sourced a suite of infrastructure components specifically designed for Ascend's computing architecture. These included:

  • TileLang Compilation Tools: Enabling more efficient utilization of Ascend's computational resources
  • Optimized Computing Libraries: Core algorithms tailored for Ascend's architecture
  • Distributed Communication Libraries: Critical components for multi-chip training coordination

Significantly, these open-source tools mirrored existing components developed for NVIDIA platforms, creating parallel support systems that facilitate potential migration between hardware ecosystems.

Broader Implications for AI Infrastructure

Should DeepSeek proceed with Ascend-based training for V5, it would represent a notable advancement for domestic AI chips in the most computationally intensive phase of model development. This decision extends beyond a single company's technical preferences, potentially influencing how the broader industry evaluates computing alternatives.

With the global AI training market still concentrated among few hardware providers, diversification of training platforms could reduce supply chain vulnerabilities while fostering more balanced competition. Technical transitions of this scale naturally require extensive validation—from performance tuning to system stability—but the direction itself opens new possibilities for AI infrastructure evolution.