DeepSeek Reportedly Plans to Pretrain V5 on Huawei Ascend 950 Chips, With Full Training Run in 2027

Chinese artificial intelligence startup DeepSeek is reportedly preparing to train its next-generation V5 model using Huawei’s Ascend 950 AI processors, marking a potential shift away from its historical dependence on Nvidia hardware for large-scale model training. According to reports citing Chinese industry sources, the company intends to use domestic chips for the computationally intensive pretraining stage, with the full training project expected to take place in 2027.

If the plan materialises, V5 could become an important test of China’s ability to develop advanced AI models using locally manufactured processors. DeepSeek and Huawei have already been collaborating on software and hardware optimisation, but the reported V5 training plan has not been officially confirmed by either company. Details about the number of chips required, the exact Ascend 950 variant and the model’s specifications remain undisclosed.

DeepSeek’s V5 Could Be Its First Major Model Pretrained on Huawei Chips

The reported plan represents a significant change in how DeepSeek approaches AI infrastructure.

Until now, Huawei’s Ascend processors have primarily supported inference workloads for DeepSeek models. Inference is the process through which an already-trained AI model generates responses, processes information and completes user requests.

Pretraining is considerably more demanding. It involves feeding a model enormous quantities of data so that it can learn patterns, language structures, reasoning capabilities and other foundational skills.

According to reports from Chinese media outlets, DeepSeek wants to use Huawei’s Ascend 950 series for this core training process. If successful, V5 could become the company’s first major model pretrained natively on Huawei’s latest generation of AI accelerators.

The reported timeline points to 2027 for the full training run. However, this is an industry-reported target rather than an official release commitment from DeepSeek.

The company has not disclosed whether V5 will use the Ascend 950PR or Ascend 950DT variant.

Why DeepSeek Is Turning Towards Huawei

DeepSeek’s reported decision is closely connected to China’s efforts to reduce its dependence on advanced American semiconductor technology.

Nvidia has traditionally supplied the high-performance GPUs used by leading AI developers to train large models. Its hardware and CUDA software ecosystem have become important components of the global AI industry.

However, US export restrictions have limited Chinese companies’ access to some of Nvidia’s most advanced processors.

This has encouraged Chinese technology companies to develop alternatives using domestic semiconductor suppliers, including Huawei.

For DeepSeek, relying more heavily on Huawei could provide greater control over its computing infrastructure and reduce exposure to foreign supply restrictions.

The challenge is that developing competitive AI models requires more than simply obtaining processors. Companies also need software frameworks, communication systems, memory capacity and reliable infrastructure capable of connecting thousands of chips.

DeepSeek’s collaboration with Huawei is intended to address several of these challenges.

DeepSeek and Huawei Are Already Working on AI Chip Software

The reported V5 plan follows a series of technical collaborations between the two companies.

In September 2026, DeepSeek announced that it was open-sourcing programming infrastructure optimised for Huawei’s Ascend processors. The work included computing and communication libraries designed to help AI developers run workloads on domestic chips.

The companies also advanced a supernode system built around 128 Ascend 950 processors, focusing on improving communication and computation across the hardware.

One important component of this effort is TileLang, an open-source programming language designed to simplify the development of AI computing kernels.

TileLang provides an alternative approach to writing low-level programs directly for Nvidia’s CUDA ecosystem. It can help developers adapt computational workloads to different hardware platforms without rebuilding every component from scratch.

This is particularly important when moving AI training systems between different chip architectures.

DeepSeek has also released Ascend-compatible versions of its TileLang programming infrastructure, helping make its software more suitable for Huawei’s processors.

These developments suggest that the reported V5 training effort is part of a longer-term technical strategy rather than an isolated hardware purchase.

Training a Large AI Model on Domestic Chips Is Difficult

Moving a major AI training workload to a new processor ecosystem presents substantial engineering challenges.

Training requires thousands of processors to work together continuously. They must exchange large quantities of information, coordinate calculations and maintain stable performance throughout long-running computing operations.

A problem with one part of the infrastructure can slow down the entire training process.

Huawei’s Ascend processors therefore need to perform reliably not only as individual chips but also as part of large interconnected computing systems.

Memory availability is another major challenge. AI accelerators require high-bandwidth memory to move data quickly during training. Restrictions on advanced semiconductor manufacturing and memory supply can limit how many powerful processors a company can obtain.

Software compatibility is equally important. DeepSeek must ensure that its training frameworks, communication libraries and model architecture work efficiently across Huawei’s hardware.

The companies’ work on TileLang and Ascend supernodes is intended to improve this compatibility, but successfully training a large model remains a much more demanding test than running an existing model.

Huawei’s Ascend Roadmap Becomes More Important

The reported DeepSeek V5 project also highlights the importance of Huawei’s upcoming semiconductor roadmap.

The Ascend 950 family is designed to support demanding AI workloads, including training and inference. Huawei is also preparing newer processors, including the Ascend 960 series, as part of its effort to expand domestic AI computing capacity.

The Information reported in October that Huawei engineers were working closely with DeepSeek to incorporate the startup’s requirements into future chip designs.

According to that reporting, Huawei was also expected to prioritise DeepSeek for supplies of its next-generation Ascend 960DT training processors.

The collaboration could help Huawei improve its hardware and software ecosystem based on the requirements of a major AI developer.

For DeepSeek, access to a reliable domestic supply of accelerators could make it easier to plan future model development without depending entirely on Nvidia.

However, manufacturing capacity remains a major constraint. Producing advanced AI processors at scale requires sophisticated semiconductor manufacturing, packaging and memory supply chains.

What the V5 Plan Means for Nvidia

The reported move could increase competitive pressure on Nvidia in China, although it does not mean DeepSeek is abandoning Nvidia hardware entirely.

Nvidia remains an important supplier for advanced AI training, and DeepSeek has continued exploring access to its processors.

The more immediate development is that DeepSeek is attempting to make its models and software increasingly compatible with Chinese hardware.

This could gradually reduce the importance of Nvidia’s CUDA ecosystem for some Chinese AI developers.

If DeepSeek successfully trains V5 on Huawei processors, other Chinese AI companies may gain greater confidence in using domestic accelerators for their own large-scale training projects.

For Huawei, securing a major AI developer as a hardware and software partner could strengthen the position of its Ascend platform.

The competitive impact will ultimately depend on performance, reliability, training costs and the availability of processors at scale.

The Bigger Picture

DeepSeek’s reported V5 plans highlight how the global AI race is increasingly becoming a competition between complete technology ecosystems rather than individual AI models.

Building a competitive model requires access to processors, memory, software frameworks, data centres and engineering expertise. China’s restrictions on access to advanced foreign chips have accelerated efforts to develop these capabilities domestically.

The collaboration between DeepSeek and Huawei could become an important example of how AI developers and semiconductor companies work together to build an alternative infrastructure.

However, the real test will be whether domestic processors can support large-scale training with the reliability and efficiency required for frontier AI models.

Looking Ahead

DeepSeek’s reported 2027 training target will depend on Huawei’s ability to supply sufficient Ascend processors and the companies’ progress in optimising the software required for large-scale model training. The exact chip configuration, computing capacity and V5 development schedule have not been announced. A successful training run would represent a significant milestone for both companies and China’s domestic AI semiconductor industry.

For Nvidia, the development could signal a gradual shift in China’s AI infrastructure towards locally developed alternatives. However, replacing an established ecosystem built around Nvidia hardware and CUDA will take time. DeepSeek’s V5 project could provide an important indication of whether Huawei’s Ascend processors are ready to move beyond inference and support the full development of large-scale AI models.