DeepSeek has released a suite of open-source programming tools for Huawei's Ascend AI chips, co-developed with Huawei with full engineering support, aimed at making it easier and cheaper to run AI models on Chinese-made processors.
The release includes two core libraries - one for AI-specific compute kernels and another for high-throughput chip-to-chip communication - designed to run efficiently on Ascend hardware without requiring Nvidia's CUDA drivers. It also adds Ascend support for TileLang, a high-level programming language that abstracts hardware details so developers can write code once and compile it across accelerator architectures.
The two companies also optimized the stack for a super-node configuration built around 128 Ascend 950 accelerators, tackling both per-chip compute utilization and inter-chip data-movement latency for large-scale model workloads.
The initiative targets a key obstacle to replacing Nvidia hardware: migrating to domestic chips requires adapting software and development tools to a different computing architecture, not just swapping processors. By lowering that software barrier, the open-source stack could broaden Ascend adoption among organizations seeking to diversify away from Nvidia for cost, supply-chain or geopolitical reasons.
Fitch Ratings, in an October 7 note, said China's AI stack - spanning models, chips, software and infrastructure - may support expansion into cost-sensitive 'global South' markets over the long term, though US leaders still lead on advanced chips, hyperscale infrastructure and cloud platforms.
Image: Efficiency of AI-related computer chips. Source: Wikimedia Commons (CC BY-SA 4.0).
