DeepSeek and Huawei Launch TileLang to Challenge Nvidia's Software Dominance

On September 30, 2026, DeepSeek, a Chinese AI startup, in collaboration with Huawei Technologies, launched a suite of open-source programming tools designed specifically for Huawei's Ascend AI chips. This initiative aims to directly challenge Nvidia's dominance in the AI software landscape.

The centerpiece of this launch is TileLang, a high-level programming language that offers a simpler alternative to Nvidia's CUDA framework. The software package includes six modules that enhance compute, communication, and programming capabilities for Ascend hardware, making development easier and optimizing performance.

DeepSeek plans to deploy at least 160,000 Huawei Ascend 950DT chips in a new data center in Inner Mongolia, with the chips expected to be partially operational by late 2027 or early 2028. This move signifies a strategic shift away from Nvidia hardware for model training.

DeepSeek's previous work with Huawei's Ascend architecture, particularly in optimizing its V4 AI model, has laid the groundwork for this transition. Nvidia's CEO, Jensen Huang, has expressed concerns regarding the implications of this partnership for US interests, viewing it as a competitive threat.

The competition now extends beyond hardware to developer habits. By lowering the switching costs for developers through TileLang, DeepSeek and Huawei aim to encourage adoption of their technology, potentially weakening Nvidia's hold in the Chinese market.

However, the effectiveness of TileLang in matching CUDA's performance in real-world applications remains to be seen. The full impact of this ecosystem will unfold as the Ascend 950DT chips are launched and developer engagement with TileLang is monitored.

New Developments in AI Model Competition

• Reflection AI is set to release an open-weight foundation model aimed at competing with DeepSeek and Alibaba's Qwen.
• The model is expected to be launched in October 2026, as reported by Axios.
• Reflection AI was founded in March 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou.
• The startup raised $2 billion at an $8 billion valuation in October 2025, with current valuations reportedly ranging between $8 billion and $25 billion.
• Reflection AI has secured a compute deal with Nebius worth over $1 billion and significant resources from SpaceX’s Colossus, including rentals of Nvidia servers.
• The company's first product, Asimov, is a code-comprehension agent designed to assist developers.
• Reflection AI's strategy involves combining its open weights with client data and advanced Nvidia GPU compute for customized AI deployments.
• Chinese laboratories currently lead in open-weight model performance, particularly in coding, mathematics, reasoning, and cost-efficiency.