Chinese AI company DeepSeek announced that it will launch its next-generation DeepSeek-R2 large language model between August 15 and August 30, 2025. This move, which comes on the heels of the launch of ChatGPT-5, is seen as a strategic step towards reducing dependence on Western technology.
What will DeepSeek-R2 offer?
DeepSeek-R2 promises a significant leap forward in architecture by utilizing an advanced Mixture of Experts (MoE) structure and a more intelligent gate network. The model’s most striking feature is that it is trained entirely on Huawei’s Ascend 910B chips. Huawei’s processing cluster reportedly achieves 91% of the performance of Nvidia’s A100 cluster. This is interpreted as a critical threshold for reducing China’s reliance on US-made AI hardware.
The model is expected to scale up to 1.2 trillion parameters, nearly double the capacity of its predecessor, DeepSeek-R1.
According to reports, DeepSeek-R2’s training cost was 97% less than GPT-4 thanks to local hardware and optimization techniques. Analysts predict this cost advantage will be directly passed on to the end user. By offering API access at significantly lower prices, DeepSeek is expected to disrupt the current pricing models dominated by giants like OpenAI and Anthropic.
This expectation has already generated excitement in Chinese technology markets. Shares of AI chipmaker Cambricon have increased by 20%, pushing its market value to over $49.7 billion.
Simultaneously with DeepSeek’s launch, Huawei introduced a new AI inference framework called Unified Cache Manager (UCM). UCM, which optimizes memory layers (HBM, DRAM, SSD) to accelerate the model’s inference process, achieved up to a 90% reduction in latency and a 22-fold increase in efficiency in tests conducted with China UnionPay. Huawei plans to open-source this groundbreaking technology in September.
The launch of DeepSeek-R2 and Huawei’s ICM framework mark a major shift in China’s global AI ambitions. These developments highlight how far the country has come to build a self-sufficient, high-performance AI ecosystem without the need for Western chips or software tools.
