[News] DeepSeek R2 Model Launch Reportedly Delayed Amid Huawei Ascend Chip Hurdles

According to Financial Times, DeepSeek has delayed releasing its new model, reportedly due to unsuccessful attempts to train it on Huawei’s chips—highlighting the limits of Beijing’s drive to replace U.S. technology.

The report, citing sources, says DeepSeek has faced ongoing technical hurdles in training its R2 model with Ascend chips, leading it to use NVIDIA chips for training while relying on Huawei’s for inference. Huawei has reportedly dispatched a team of engineers to DeepSeek’s office to assist in developing the R2 model on its AI chip, the report adds.

However, DeepSeek has yet to complete a successful training run on the Ascend chip and is still working with Huawei to ensure the model is compatible with Ascend for inference, the report says, citing sources.

China’s Push for Domestic Chips

Following the January release of its R1 model, DeepSeek was urged by authorities to use Huawei’s Ascend processor instead of NVIDIA’s systems, according to sources cited in the report.

Chinese authorities have been urging tech companies to adopt homegrown technology. Similarly, according to Reuters, sources say the Cyberspace Administration of China this week summoned companies including Tencent, ByteDance, Baidu, and several smaller firms over their purchases of NVIDIA’s H20 chips, asking them to justify the transactions.

DeepSeek R2 Delays Continue as Usage Rate Plummets

DeepSeek-R2, originally expected in May this year, has faced multiple delays, as noted by Chinese media outlet Mydrivers. A Reuters report from June adds that DeepSeek has yet to set a launch date for the R2 model, as CEO Liang Wenfeng remains dissatisfied with its performance.

Regarding the continued delay, Mydrivers, citing industry sources, explains that R1 was built using OpenAI’s existing, globally sourced, and certified training data, whereas R2 requires a much larger volume of training data. The quality of domestically available data for R2, however, is relatively low. Another factor is the severe shortage of high-end GPUs, which has significantly reduced DeepSeek’s training efficiency and, in turn, slowed the progress of R2’s development.

As the R2 model faces delays, DeepSeek’s usage rate has also reportedly declined. According to Mydrivers, its usage rate has dropped from 50% at the beginning of the year to just 3% at present.

Read more

(Photo credit: DeepSeek’s X)