Customer evaluations of Huawei’s latest AI accelerator, the 950PR, have yielded positive feedback and large Chinese technology companies including ByteDance and Alibaba intend to place orders, according to two people familiar with the situation and a third person who has knowledge of those plans. The responses from prospective customers mark a notable development for Huawei as it seeks broader adoption of domestically developed AI chips.
Industry insiders said the new 950PR is seen as more usable for many Chinese tech firms because it offers greater compatibility with Nvidia’s CUDA software ecosystem and improved response speeds compared with Huawei’s prior flagship, the Ascend 910C. That combination of software compatibility and faster responses reportedly made potential buyers more willing to deploy the chip more widely across their operations.
Two people familiar with Huawei’s plans told sources that the company intends to ship around 750,000 950PR units this year. They said sample units were provided to customers in January and that mass production is expected to begin next month, setting up full-scale shipments to take place in the second half of the year. The people who provided these details were not authorised to speak publicly and declined to be named. Huawei, ByteDance and Alibaba did not reply to requests for comment.
Strategic and technical context
The 950PR uses traditional DDR memory in its base configuration and will be priced at about 50,000 yuan per card, the people said. A higher-end model equipped with faster HBM memory is expected to be sold for around 70,000 yuan. Huawei first referenced the new chip last September when it outlined its longer-term semiconductor strategy and said it planned to introduce some of the world’s most powerful computing systems.
Huawei has previously relied primarily on its proprietary CANN software stack, while many Chinese technology firms have historically used Nvidia’s software ecosystem. The 950PR’s enhanced support for CUDA is intended to make it easier for developers to migrate existing models and workflows that were built for Nvidia systems over to Huawei’s hardware.
Performance profile and market demand
Compared with the Ascend 910C, the 950PR reportedly offers only a modest uplift in raw computational throughput. However, the chip has been tuned to perform particularly well on inference workloads - the stage in which trained AI models are executed to answer queries or carry out tasks. Sources said demand for inference computing in China is rising rapidly as the technology sector shifts focus from model development to practical deployment, a trend that has been accelerated by the swift adoption of the open-source AI agent OpenClaw.
That pivot toward inference use cases is one reason Huawei’s new design has attracted attention: even modest improvements in latency and compatibility can meaningfully affect the economics and integration effort for large-scale AI deployments in production environments.
Implications for Nvidia and supply restrictions
The 950PR’s arrival comes amid a challenging market for Nvidia in China. Many Nvidia artificial intelligence chips have been restricted from sale in China by the U.S. government over concerns that the technology could enhance the Chinese military’s capabilities. Last year, the Trump administration approved sales of Nvidia’s H200 chips under a set of conditions that could limit volumes. The H200 has since received approval from Chinese authorities, but it remains unclear when the product will be allowed into the country.
Sources and limitations
The details in this report are based on conversations with people close to the matter who were not authorised to speak publicly and declined to be identified. Where information was limited in the accounts provided, this article reflects those limitations rather than adding unconfirmed detail.