DeepSeek V4's Shift and Huawei Hardware

The global artificial intelligence landscape is witnessing a significant evolution, as China intensifies efforts to reduce its reliance on US technology. A recent report by The Information, cited by Reuters, indicates that DeepSeek's upcoming AI model, V4, may operate on Huawei chips, rather than the NVIDIA hardware that currently powers most Large Language Models (LLM) and large-scale AI systems. This potential change is not an isolated incident but reflects a broader transition in Chinese AI infrastructure.

DeepSeek's decision to adapt parts of its model to work with Huawei's Ascend chips is a clear signal of this direction. In parallel, major Chinese technology companies such as Alibaba, ByteDance, and Tencent have reportedly placed orders for hundreds of thousands of Huawei chips in anticipation of the V4 release. The scale of these orders suggests a strategic reorganization in infrastructure planning, with an emphasis on securing supply chains and reducing dependence on foreign suppliers.

The Technical Context and the Challenge to NVIDIA's Dominance

For years, companies engaged in building complex LLM and AI models have relied on NVIDIA GPUs. The CUDA software stack, along with chips like the A100 and H100, has set the industry standard for AI model training. However, DeepSeek is taking a different route. Reuters reported that DeepSeek withheld early access to V4 from US chipmakers such as AMD and NVIDIA, instead granting more time to Chinese suppliers, including Huawei, to optimize their software.

Compatibility between AI models and hardware is not a given: a model optimized for one type of chip does not always run efficiently on another. To overcome this challenge, DeepSeek is reportedly rewriting parts of V4's code, in collaboration with Huawei and Cambricon, to ensure the model can operate effectively on Huawei chips. Although a 2025 Reuters report indicated that Huawei and other Chinese chipmakers had struggled for years to match the performance of high-end NVIDIA chips for training, adapting the model to the hardware could help narrow this gap. The real-world performance of AI systems depends on a complex interaction between hardware, software, and data.

Implications for the AI Ecosystem and TCO

This evolution suggests the formation of two distinct AI ecosystems. One is centered on US technology, with NVIDIA hardware and software at its core. The other is taking shape around Chinese companies, with Huawei chips and local software stacks. Export controls imposed by the United States have limited access to more advanced NVIDIA chips, fueling demand for domestic alternatives and accelerating the development of local solutions.

The DeepSeek V4 model could become a crucial test for the viability of this second system. Should the model demonstrate competitive performance, it could encourage more companies to follow the same path. This shift could also influence the structure of AI costs. Reuters linked DeepSeek's earlier claims regarding the lower costs of its models to investor concerns about the high spending levels of some US AI firms. For those evaluating on-premise deployments, Total Cost of Ownership (TCO) analysis becomes critical, considering not only initial hardware costs but also energy efficiency, maintenance, and data sovereignty.

Future Prospects and Trade-offs in the AI Landscape

The race to develop increasingly powerful AI models is often framed as a competition based on model size or benchmark scores. However, control over chips, software, and supply chains is a decisive factor that can shape what is technically possible and at what cost. DeepSeek's move to Huawei chips does not settle the "AI race," but it adds a new strategic dimension.

It demonstrates that high-level AI work is no longer tied to a single hardware path. Different regions may build their own technological stacks, each with its specific trade-offs in terms of performance, costs, and autonomy. If V4 delivers robust results, it could mark a turning point, demonstrating that competitive AI systems can be built effectively outside the dominant NVIDIA ecosystem. This scenario offers new opportunities for those seeking self-hosted or air-gapped solutions, where hardware choice and local software stack management are priorities.