TileLang vs CUDA: DeepSeek, Huawei Debut 128-Chip Open-Source AI Stack

For AI and machine learning practitioners, the central question is whether TileLang can match CUDA's developer experience and performance while unlocking Huawei's Ascend hardware. DeepSeek says TileLang was created precisely to provide a universal, easy-to-program high-level language that still reaches full hardware performance. If true, the open-source release could reshape which hardware AI teams choose for training and inference in China.

On September 30, 2026, Chinese AI startup DeepSeek and Huawei Technologies announced a strategic partnership to build open-source programming infrastructure for Huawei's Ascend AI chips, according to reports by Reuters and statements published on DeepSeek's WeChat account. The announcement represents a deliberate move by two of China's most influential technology players to reduce reliance on Nvidia's CUDA software ecosystem and its hardware, as U.S. export controls continue to tighten access to advanced semiconductors. DeepSeek confirmed the release of open-source compute and communication libraries for Huawei's Ascend platform, and introduced TileLang, an open-source high-level programming language specifically designed for AI chips. DeepSeek said TileLang offers a simpler programming model than Nvidia's proprietary CUDA platform, improving development efficiency and simplifying code logic.

“

DeepSeek said TileLang offers a simpler programming model than Nvidia's proprietary CUDA platform, improving development efficiency and simplifying code logic.

The partnership did not emerge in isolation. Huawei had recently unveiled its next-generation AI processors and supernode computing systems, projecting wide adoption of its computing systems for model training in the coming year. DeepSeek and Huawei jointly progressed a supernode configuration built on 128 Ascend 950 chips, refining computational performance and communication links across the clustered system. Huawei provided full technical backing throughout the development of the shared programming infrastructure, according to DeepSeek. By pairing DeepSeek's expertise in large-language-model architecture and software optimization with Huawei's silicon and systems engineering, the two companies are directly targeting the high-margin software moat that has made Nvidia's GPU ecosystem so difficult to displace.

Industry context is essential here. Nvidia's CUDA platform has long been more than a programming language; it is an entrenched developer ecosystem, a library of optimized kernels, and a training ground for AI model builders. China's domestic chipmakers, including Huawei, have made meaningful progress on hardware performance, but software tooling has remained a weak point. Without a mature programming environment, domestic AI chips struggle to attract developers or match the productivity of Nvidia-based workflows. DeepSeek's open-source TileLang is therefore strategically significant: it attempts to solve the software problem that has historically constrained Chinese AI chip adoption. The collaboration also reflects the broader national imperative to create independent, self-controlled semiconductor supply chains amid tightening U.S. export controls.

The market implications are substantial. For Nvidia, the partnership reinforces the risk that China's AI market will increasingly move to domestic alternatives, even if Nvidia retains significant global performance leadership. For Huawei, the Ascend platform gains a credible software story and a high-profile AI partner, which could accelerate enterprise and cloud adoption. For DeepSeek, the collaboration lowers its own training and inference costs by optimizing models for domestic hardware, while aligning it with strategic national priorities. The joint decision to release the programming infrastructure as open source could also attract a community of developers beyond China, although U.S. sanctions and geopolitical concerns may limit international adoption.