DeepSeek founder Liang Wenfeng announced that the company will increase its reliance on AI accelerators from Huawei and other domestic suppliers. The firm expects to receive a new batch of Huawei hardware by the fourth quarter of 2026 to support its model training efforts.
The company is currently developing an AI model with 2 trillion parameters, surpassing its existing V4 flagship, which features 1.4 trillion parameters. Future development plans include building a model with 8 trillion parameters, though success depends on the availability of memory chips and power supply.
While DeepSeek intends to shift toward local hardware, it acknowledges that a full transition from imported chips is not immediately feasible. For instance, training an OpenAI-level model would require 50,000 Nvidia GB300 units, whereas using Huawei Ascend 950 chips would necessitate 200,000 units.
Despite current supply constraints faced by Huawei due to Western sanctions, DeepSeek leadership anticipates that Huawei's AI chips will match Nvidia's performance within a few years. The company expects to receive Ascend 950DT accelerators by year-end, with the Ascend 960DT model slated for early 2027.
