China's DeepSeek is teaming up with Huawei to advance the infrastructure around Huawei's Ascend chips, a move aimed at reducing its reliance on Nvidia hardware, NDTV Profit reports.

The partnership pairs China's most closely watched AI startup with the country's leading domestic chip designer at a moment when access to Nvidia's data-center GPUs has become the central constraint on Chinese model training. US export controls have progressively cut off Chinese buyers from Nvidia's most advanced accelerators, pushing cloud providers and AI labs toward domestic alternatives — above all Huawei's Ascend family.

For DeepSeek, the calculus is straightforward. The company built its reputation on extracting headline performance from constrained hardware budgets, and its training runs depend on a supply of accelerators that Washington can restrict at any time. Anchoring its infrastructure to Huawei's Ascend lineup gives the startup a domestic supply chain that export controls cannot throttle, and gives it a stake in improving the software and systems ecosystem around those chips.

That ecosystem gap is the crux. Nvidia's moat extends well beyond silicon into CUDA, its mature developer toolchain and the years of optimization work baked into major training frameworks. Huawei has pushed its own alternative — the CANN software stack and the MindSpore framework — but adoption outside China remains thin, and even domestic developers frequently note the extra engineering effort required to port workloads trained on Nvidia hardware onto Ascend parts.

A collaborator of DeepSeek's profile changes that equation. If the startup's engineers help harden Ascend's software layer for frontier-scale model training, Huawei gains a reference customer whose models are scrutinized worldwide, while DeepSeek gains early access to an accelerator platform it can help shape. The arrangement follows a broader pattern across China's AI sector, where leading labs and cloud operators — from Baidu to ByteDance and Alibaba — have ramped up work on domestic silicon even as they continue to use Nvidia hardware where it remains legally available.

Huawei's Ascend roadmap has advanced despite US sanctions that cut the company off from leading-edge foundry capacity. Its Ascend 910B became the workhorse of Chinese inference deployments after Washington tightened controls, and the company has since worked to expand output through domestic manufacturing partners, including SMIC, even under lithography constraints that limit yields and volumes relative to TSMC-fabricated Nvidia parts. Capacity, not design capability, remains the binding constraint on how far Ascend can scale.

The commercial stakes run in both directions. For Huawei, DeepSeek's involvement could accelerate the maturity of an ecosystem it needs if Ascend is to compete for workloads beyond China's borders. For DeepSeek, deeper Ascend integration insulates its training pipeline from the next round of export restrictions — a risk that US policymakers have signaled repeatedly, including proposals to tighten rules further on Nvidia's China-market parts.

The deal also sharpens the competitive pressure on Nvidia inside what was once one of its largest markets. Chinese customers accounted for a significant share of Nvidia's data-center revenue before successive export-control rounds rolled that back, and every domestic stack that reaches production maturity narrows the wedge Nvidia can reclaim if restrictions ease. DeepSeek's models already demonstrated that frontier-adjacent results are achievable on constrained hardware; extending that work onto Huawei's silicon would turn the demonstration into infrastructure.

How quickly that happens depends on engineering throughput on both sides — the pace of Ascend software maturation, and how fast DeepSeek can adapt its training pipeline without sacrificing efficiency. The partnership signals that China's AI leaders now treat domestic accelerators not as a stopgap but as a strategic baseline. If Ascend's ecosystem matures fast enough to carry DeepSeek's next generation of models, the dividing line in AI compute will harden along geopolitical rather than purely technical grounds.