DeepSeek R2 Faces Delays Due to Failing Huawei Chips

deepseek

The development of the new DeepSeek model is not going smoothly. The model is being trained on Huawei Ascend chips that are proving to be unreliable.

Earlier this year, DeepSeek surprised the world with its efficient R1 model, but successor R2 is keeping us waiting. Difficulties during the training phase have forced DeepSeek to postpone the release of the model. Under pressure from the Chinese government, DeepSeek is using Huawei chips, but they are failing.

DeepSeek R2 should have been launched in spring. The launch has been postponed because the Chinese company struggled to obtain powerful Nvidia hardware due to export restrictions. DeepSeek was forced to switch to Huawei.

Failing Chips

Training on the Huawei Ascend GPUs has been difficult, confirm multiple sources to ArsTechnica. DeepSeek barely managed to run stable training cycles on the chips. To such an extent that DeepSeek continues to actively search for Nvidia chips for training and would only want to run inference on Huawei equipment.

However, the Chinese government is also interfering and wants Chinese technology companies to stop purchasing chips from American companies. Huawei sent an intervention team on-site to fix the problem. Chinese media remain convinced that the launch of DeepSeek R2 won’t be delayed much longer.

Caught between Two Fires

DeepSeek’s difficulties illustrate the position of Chinese technology players, who are caught between two fires. Export restrictions make it difficult for them to obtain foreign chip technology, but they’re also still waiting for the first Chinese “wonder chip”.

Not only are Chinese chips less powerful than what Nvidia offers, they also lack support for the Cuda ecosystem, Nvidia’s secret weapon. This makes a “China chips first” strategy unattractive for a company like DeepSeek that wants to compete on the global AI stage. Nvidia itself doesn’t want to let go of the Chinese market and made a “reconciliation offer” to the US government.

The new model risks losing relevance compared to new models like GPT-5. An additional difficulty for DeepSeek is that it is viewed with suspicion by European authorities.

read also

GPT-5 Tested: Structured and “No-Nonsense” Thinker