Checked Nov 14, 2026Start a new check

What was checked

It is a lie that China needs Nvidia chips to train their models. They can now train entirely on Huawei chips

Assessment summary

Inaccurate

Available evidence does not support this broad claim. Huawei has trained some models on Ascend chips, but disclosed runs have been much smaller than frontier-scale training, while Nvidia remains widely used and Huawei still has performance and supply limitations. China is not exclusively dependent on Nvidia, and some models have used domestic chips. However, leading Chinese developers still rely primarily on Nvidia hardware, and Chinese vendors continue using Nvidia chips for training.

Key findings

Model assessments

Assessment summary

  • The claim is significantly misleading.
  • Current reporting says Chinese companies still use Nvidia chips for model training, while Huawei chips are a growing alternative but lag in performance, software maturity, and supply vietnam.vn afp.com samaritan-research.org.
  • Some models may be trained on domestic hardware, but that does not establish that China can now train all models entirely on Huawei chips samaritan-research.org rand.org.

Claims reviewed

China can now train entirely on Huawei chips

Contradicted
Available evidence does not support this broad claim. Huawei has trained some models on Ascend chips, but disclosed runs have been much smaller than frontier-scale training, while Nvidia remains widely used and Huawei still has performance and supply limitations samaritan-research.org rand.org.

This claim may mislead readers without added context.

China needs Nvidia chips to train their models

Needs context
China is not exclusively dependent on Nvidia, and some models have used domestic chips. However, leading Chinese developers still rely primarily on Nvidia hardware, and Chinese vendors continue using Nvidia chips for training vietnam.vn afp.com samaritan-research.org.

This claim may mislead readers without added context.

Assessment summary

  • The claim contains both accurate and misleading elements.
  • While some Chinese companies have successfully trained models entirely on Huawei chips (DeepSeek V4, Zhipu AI's GLM-Image, China Telecom's TeleChat3), the broader claim that China "can now train entirely on Huawei chips" overstates the current reality.
  • Most major Chinese AI labs still prefer Nvidia hardware when available, and performance gaps remain significant on a per-chip basis.

Claims reviewed

They can now train entirely on Huawei chips

Needs context
Some Chinese companies have demonstrated this capability (DeepSeek V4, Zhipu AI, China Telecom), but this is not universally true across China's AI industry. Huawei chips still lag behind Nvidia on a per-chip basis, and most major AI labs continue to rely on Nvidia when possible. The claim suggests this is a widespread capability when it remains limited to specific companies and use cases.

This claim may mislead readers without added context.

It is a lie that China needs Nvidia chips to train their models

Needs context
While some Chinese companies have trained models on Huawei chips, most major Chinese AI developers still prefer Nvidia hardware. As of 2024, Chinese companies bought around 1 million Nvidia H20 chips compared to 450,000 Huawei Ascend 910B chips, and many top models including DeepSeek V3 were trained on Nvidia hardware. The statement oversimplifies a complex situation where Nvidia remains the preferred choice despite alternatives.

This claim may mislead readers without added context.

Assessment summary

  • The statement mixes a kernel of truth with an overbroad claim.
  • Chinese firms have increasingly used Huawei Ascend chips to train some models, but independent analysis and government reports show Huawei hardware still lags Nvidia on many large-scale training workloads and is not yet a universal drop-in replacement.
  • The overall claim is therefore misleading rather than fully true or fully false.

Claims reviewed

They can now train entirely on Huawei chips.

Contradicted
Some Chinese companies have trained certain models on Huawei Ascend hardware, but evidence does not support the blanket claim that all (or even most frontier) models can now be trained entirely on Huawei chips; independent technical comparisons and expert reports indicate Huawei still lags Nvidia for many large-scale training workloads and requires larger clusters or other workarounds to approach parity merics.org spectrum.ieee.org uscc.gov.

This claim may mislead readers without added context.

It is a lie that China needs Nvidia chips to train their models.

Needs context
Chinese firms do not strictly "need" Nvidia chips for all training because domestic alternatives (e.g., Huawei Ascend) are being used for many training jobs theinformation.com reuters.com. However, credible analyses and US government reports show Ascend-series solutions still trail Nvidia on many high-end pretraining tasks, so saying the claim is a categorical lie overstates the situation merics.org spectrum.ieee.org.

This claim may mislead readers without added context.

Sources cited

20 sources

Share this result