DeepSeek Is Shaping Huawei's New AI Chips: The Ascend 950DT and 960DT Partnership

In the global AI race, the boundaries between hardware supply and software ecosystems are taking on an increasingly strategic dimension. US export restrictions on advanced semiconductor technologies have pushed China-based tech giants and AI labs toward an unprecedentedly integrated way of working. The most concrete and influential example of this transformation is the deepening partnership between telecom and hardware giant Huawei and DeepSeek, which made waves worldwide with its highly efficient large language models (LLMs). Huawei no longer just designs its next-generation AI accelerators in its own labs; it shapes them directly around DeepSeek's model architecture and compute requirements.

Joint Engineering on the Ground: Direct Input Into Chip Design

In the traditional semiconductor industry, chipmakers would first develop the hardware, and AI companies would then try to optimize their algorithms for the capacity that hardware offered. The collaboration between Huawei and DeepSeek has completely reversed this one-way process. According to information from industry sources, Huawei is running a simultaneous R&D process by stationing expert hardware and systems engineers directly in DeepSeek's offices.

The bottlenecks DeepSeek researchers hit while training huge-parameter models and running inference are analyzed in real time by Huawei engineers. Feedback from critical metrics in particular, such as matrix multiplications, routing traffic in Mixture-of-Experts (MoE) architectures and interconnect bandwidth between chips, forms the core building blocks of Huawei's next-generation Ascend architecture. For the first time, an AI lab's algorithmic vision is steering a semiconductor maker's silicon design this directly.

Ascend 950DT and Ascend 960DT: The Next-Generation Hardware Roadmap

The first big fruits of this tight integration between the two companies can be seen in Huawei's next-generation Ascend series accelerators. According to the roadmap that has emerged, DeepSeek is preparing to actively deploy Huawei's Ascend 950DT processors for both model training and large-scale inference operations.

A Massive 160,000-Chip Cluster in Inner Mongolia

Industry reports reveal that DeepSeek plans to bring at least 160,000 Ascend 950DT chips online to handle model serving and inference workloads at the huge data center campus it is building in the Inner Mongolia region. An installation of this scale would be one of the largest compute clusters China has built with domestic AI accelerators.

The Ascend 960DT Timeline Moved Forward

Another critical link in the collaboration is the next generation, the Ascend 960DT. In line with heavy demand and strategic goals from the domestic ecosystem, led by DeepSeek, Huawei has pulled forward the production and launch timeline of the Ascend 960DT series. It is reported that DeepSeek will be given priority in the first shipments of Ascend 960DT chips, which are planned to enter mass production in the first quarter of next year. Equipped with an improved high-bandwidth memory (HBM) architecture and increased chip-to-chip communication capacity, this new series directly targets the complex compute loads of training frontier models.

A Software Move Against NVIDIA CUDA's Dominance: TileLang

The biggest force keeping NVIDIA at the top of the AI hardware market is considered to be not just the raw compute capacity of its GPUs, but the CUDA software ecosystem that has matured over more than twenty years. Changing developers' habits and moving models to a different chip architecture usually leads to serious performance losses and engineering costs.

To get past this software wall, Huawei and DeepSeek are pursuing a joint strategy centered on the open source world:

• TileLang Support and an Open Source Toolchain: DeepSeek has released a version of TileLang, the open source programming language that makes it easier to write high-performance kernels across different hardware architectures, that is fully compatible with the Huawei Ascend 950 series, together with the related communication libraries.
• CANN and Compiler Optimizations: Compatibility and optimization problems previously seen on CANN (Compute Architecture for Neural Networks), Huawei's own software stack, are being largely resolved through low-level code work by DeepSeek engineers.
• A Simplified Migration Process: Thanks to the published open source compiler and inference frameworks, the goal is to move complex models that previously ran only on NVIDIA GPUs to Huawei Ascend clusters with minimal code changes.

Cluster Stability and the Engineering Hurdles to Overcome

Although the Huawei and DeepSeek partnership is a revolutionary step in China's vision of technological independence, serious technical challenges that must be overcome at global scale remain. Training frontier models requires thousands of chips to work uninterrupted and in sync as a single supercomputer for weeks or even months.

Hardware experts point out that what matters is not a single chip's on-paper TFLOPS figures but cluster reliability and fault tolerance in huge clusters where tens of thousands of chips are connected together. A hardware failure or communication delay on a single node during training can stall the whole training run and cause costly time losses. In addition, because of restrictions on advanced manufacturing techniques (lithography), keeping production yield high and optimizing energy consumption are among the most important engineering tests Huawei faces.

A New Era in the Global AI Ecosystem

The combination of DeepSeek's expertise in algorithmic efficiency and Huawei's hardware engineering capacity is accelerating the era of “co-evolution of hardware and models” (co-design) in the AI industry. Vertical integration similar to Google designing its own TPU chips together with its Gemini models in the Western ecosystem, or the custom silicon projects of Meta and Microsoft, is becoming a standard in China that spreads to the entire domestic ecosystem through the Huawei and DeepSeek partnership. This strategic synergy looks set to redefine not only regional supply chains but also the balance of global AI hardware competition in the coming period.