China's top AI models still run on Nvidia chips despite local push
Synopsis
Despite intense political pressure to go local, China's top AI labs — including those behind DeepSeek, Kimi K3, and Alibaba's models — are still training on Nvidia chips. One researcher says switching to Huawei's Ascend would add at least 50% in time and costs, exposing a software-ecosystem gap that hardware advances alone cannot fix.
Key Takeaways
China's leading LLM developers, including those behind DeepSeek , Kimi K3 ( Moonshot AI ), LongCat , Alibaba , and Meituan , continue to train models on Nvidia chips as of August 2026 .
Nvidia 's CUDA platform remains the industry standard; code written for CUDA cannot run directly on Huawei 's Ascend chips without extensive rewriting.
Migrating to Huawei 's Ascend chips and the CANN software stack could add at least 50 per cent in time and costs, according to James Wang , an AI researcher at a Shanghai -based university institute.
The bottleneck is primarily a software ecosystem problem, not a hardware capability problem — Huawei 's chips have advanced, but compiler and tooling maturity lags significantly behind CUDA .
US export controls have restricted access to Nvidia 's top-tier chips, yet the cost of switching to domestic alternatives remains a stronger deterrent for most developers.
China's most advanced large language models (LLMs) continue to be trained on Nvidia chips, according to sources at major Chinese AI developers, as the prohibitively high cost of migrating to domestic semiconductors stalls Beijing's self-sufficiency drive. The disclosure, surfacing on 10 August 2026, underscores a stubborn dependency that persists even as homegrown hardware makers race to close the gap.
The Software Lock-In Problem
At the heart of the delay is a deep-rooted software ecosystem challenge. Nvidia's Compute Unified Device Architecture (CUDA) platform has long been the de facto standard for AI model development worldwide. Switching away from it is not simply a matter of swapping hardware — it demands a wholesale rewrite of training pipelines, tooling, and optimisation logic.
'Training LLMs on Nvidia chips for now remains the norm among Chinese AI developers,' said a person familiar with the industry. The statement reflects a consensus that has quietly persisted even as geopolitical pressure mounts on Chinese firms to reduce reliance on US technology.
Huawei's CANN vs Nvidia's CUDA
Huawei Technologies' alternative compute platform — Compute Architecture for Neural Networks (CANN), designed for its Ascend chip series — requires developers to rewrite and optimise large amounts of existing code, according to an AI researcher involved in model development. This is not a minor adjustment; it represents a fundamental re-engineering of workflows that teams have spent years building on CUDA.
James Wang, who develops AI models at a research institute affiliated with a Shanghai-based university, put it plainly: 'Our existing training pipelines are reliant on CUDA. CUDA code cannot run directly on Ascend and requires extensive rewriting.' Wang estimated that migrating existing workflows to Huawei's Ascend chips could add at least 50 per cent in time and costs for his team.
Who Is Affected
The dependency spans some of China's most prominent AI players. Developers behind models including DeepSeek, Kimi K3 by Moonshot AI, LongCat, and products from Alibaba Group Holding and Meituan are all navigating the same bottleneck, according to industry sources. For these companies, the engineering overhead of a full chip migration is not merely inconvenient — it is commercially prohibitive in the near term.
The Competitive Backdrop
The situation highlights a structural asymmetry in the global AI chip race. While Huawei's Ascend series has made measurable hardware progress, software maturity — compilers, libraries, debugging tools, and community support — remains years behind CUDA's entrenched ecosystem. Domestic alternatives lack the breadth of third-party optimisations that Nvidia's platform has accumulated over more than a decade.
The US export control regime has restricted China's access to Nvidia's most advanced chips, yet Chinese developers continue to work with available Nvidia hardware rather than pivot to local silicon, signalling that regulatory pressure alone has not been sufficient to force a transition.
What's Next
The pace at which Huawei and other domestic chipmakers can mature their software stacks will be the decisive variable. Until CANN and comparable platforms can absorb CUDA-native workflows with minimal friction, China's frontier AI development will remain tethered to Nvidia infrastructure — a dependency that both complicates Beijing's technology ambitions and exposes leading AI developers to ongoing supply-chain risk.
Point of View
But they cannot instantly dissolve a software moat built over fifteen years. Beijing's self-sufficiency push has focused heavily on hardware — fab capacity, chip design, packaging — while the harder problem of replicating CUDA's vast library of optimised kernels, third-party integrations, and developer tooling has received comparatively little public attention. What mainstream coverage often misses is that the 50 per cent cost premium cited by researchers is not a temporary friction but a compounding disadvantage: every month that Chinese AI labs delay migration, their CUDA-native codebases grow larger and harder to port. Huawei's real challenge is not building a chip that benchmarks well — it is convincing time-pressured AI teams that the migration pain is worth absorbing before a competitor pulls ahead.
NationPress
10 Aug 2026
Frequently Asked Questions
Why are China's AI companies still using Nvidia chips?
China's leading AI developers continue to use Nvidia chips primarily because their training pipelines are built on Nvidia's CUDA software platform, and migrating to domestic alternatives like Huawei's Ascend requires extensive code rewrites. According to one AI researcher, the switch could add at least 50 per cent in time and costs, making it commercially prohibitive in the near term.
What is the difference between CUDA and Huawei's CANN?
CUDA (Compute Unified Device Architecture) is Nvidia's long-established software platform for AI development and is the global industry standard. Huawei's CANN (Compute Architecture for Neural Networks) is the equivalent platform for its Ascend chips, but code written for CUDA cannot run directly on Ascend and requires significant rewriting and optimisation before it is compatible.
Which Chinese AI models are affected by this chip dependency?
Developers behind some of China's most prominent AI models — including DeepSeek, Kimi K3 by Moonshot AI, LongCat, and AI products from Alibaba Group Holding and Meituan — are all reportedly navigating the same Nvidia dependency, according to industry sources familiar with the matter.
How do US export controls affect China's AI chip situation?
US export restrictions have limited China's access to Nvidia's most advanced chips. However, Chinese AI developers are still using available Nvidia hardware rather than switching to domestic alternatives, indicating that supply-chain risk has not yet outweighed the engineering cost of migration to local silicon.
When will Chinese AI developers switch to Huawei Ascend chips?
No firm timeline has been established. The transition depends on how quickly Huawei and other domestic chipmakers can mature their software ecosystems — particularly compiler support, optimised libraries, and developer tooling — to a level where migration becomes cost-effective for AI labs operating under competitive pressure.
