DeepSeek R2 model release reportedly held back by faulty Huawei chips

China’s leading artificial intelligence startup DeepSeek Ltd., which took the industry by storm earlier this year, was reportedly forced to delay the release of its upcoming R2 model after struggling to train it using chips supplied by Huawei Technologies Co. Ltd.

The delay highlights the ongoing struggles China faces as it pushes domestic AI companies to move away from their reliance on U.S. technology, Reuters reported today.

Three anonymous sources told Reuters that DeepSeek had been encouraged to use Huawei’s Ascend-branded graphics processing units, rather than Nvidia Corp.’s hardware, to develop the R2 model. The startup previously rocked the AI industry with the release of its flagship R1 large language model, claiming it trained the algorithm at a cost of just a few million dollars, in contrast to the billions of dollars spent by U.S. AI firms such as OpenAI and Google LLC.

Chinese authorities reportedly told DeepSeek to use Huawei’s chips in the wake of a decision by U.S. President Donald Trump to prohibit the export of Nvidia’s popular H20 GPU to China. The ban came into effect immediately following the April announcement.

However, the startup faced numerous and persistent technical issues when trying to train R2 on Huawei’s Ascend chips, and ultimately went back to using the Nvidia chips that were available to it. It did, however, continue to use the Ascend chips for inference, the sources said. AI training involves teaching models to learn using large datasets, while inference refers to using already-trained models to power AI applications, such as chatbots and image generators.

DeepSeek had originally hoped to launch R2 in May, but it has not yet done so, and is seen to have lost ground to its U.S. rivals, which have since debuted AI models that surpass the performance of R1.

The difficulties experienced by DeepSeek illustrate how China’s domestic chipmaking industry still lags behind that of the U.S., hampering the country’s efforts to become self-sufficient in technology. It also explains why China was so keen to secure a trade deal with Washington, which included the provision that it be allowed to resume buying Nvidia’s H20 chips.

Despite being allowed to buy Nvidia’s chips again, China has insisted that any local AI developers justify such orders, which will inevitably come at the expense of domestic chipmakers, the Financial Times said in a report earlier this week. The country is still keen to promote the adoption of alternatives from Huawei and Cambricon Co. Ltd. where possible, the report said.

Last month, Huawei debuted its most advanced AI server, the CloudMatrix 384 system that’s powered by 384 Ascend 910C GPUs, positioning it as an alternative to Nvidia’s GB200 NVL72 system. At the time, it said the CloudMatrix 384 surpassed Nvidia’s server in terms of pure petaflops performance, while also providing more memory and greater bandwidth, albeit while using significantly more power.

But although some Western analysts praised the CloudMatrix 384 system, others believed that the Ascend chips remain plagued by stability issues and slower chip-to-chip connectivity than Nvidia’s products. Ritwik Gupta, an AI researcher at the University of California at Berkeley, told the Financial Times that the software provided with Huawei’s chips is thought to be inferior to Nvidia’s.

Reuters’ sources said Huawei sent a crack team of engineers to try and assist DeepSeek in training the R2 model, yet even with them onsite, the company struggled to conduct a successful training run. The engineers did have more success in getting the Ascend chips to power inference, though.

According to Gupta, Huawei appears to be facing “growing pains” in terms of using Ascend for AI training, but he expects the company to solve whatever challenges are holding it back. “Just because we’re not seeing leading models trained on Huawei today doesn’t mean it won’t happen in the future,” he said. “It’s a matter of time,”

DeepSeek founder Liang Wenfeng has reportedly told employees that he’s disappointed with the progress of R2, and wants to spend more time enhancing the model so it can unseat its American rivals. However, Chinese media reports suggest that R2 may finally make its debut in the coming weeks.

Image: SiliconANGLE/Meta AI