Following the release of ChatGPT-5， the much-anticipated major update to China’s homegrown large language model， DeepSeek， appears to be on the horizon. Based on user inquiries directed at DeepSeek， the model has indicated that DeepSeek-R2 is expected to be launched between August 15th and August 30th， 2025.
This development has coincided with a notable surge in the stock prices of several Chinese companies within the domestic computing power supply chain. For instance， Cambricon saw its stock price hit the daily limit (20% increase) today， reaching a historical high with a market capitalization exceeding 355 billion yuan. This suggests a growing market confidence in domestically developed AI infrastructure and capabilities.
Previous information indicated that DeepSeek-R2 would leverage a more advanced Mixture-of-Experts (MoE) architecture. This approach integrates a sophisticated Gating Network designed to optimize performance for high-load inference tasks， a critical factor for efficient AI deployment.
Analysts anticipate that DeepSeek-R2’s pricing strategy could be significantly more competitive than comparable offerings from OpenAI. This could potentially disrupt the current pricing models for AI services and make advanced AI more accessible.
Furthermore， there are reports suggesting that DeepSeek-R2’s operational costs are projected to be 97% lower than GPT-4. Critically， the model has been trained on Huawei’s Ascend chips， emphasizing a commitment to comprehensive， end-to-end domestic control and self-reliance in the AI ecosystem.
Sources close to the project have revealed that DeepSeek-R2 may boast a total parameter count of 1.2 trillion， nearly doubling the 671 billion parameters of DeepSeek-R1. This substantial increase in parameters is a key indicator of enhanced model capacity and potential for improved performance.
DeepSeek-R2 has been trained on Huawei’s Ascend 910B chip clusters. In FP16 precision， the training achieved a computing power of 512 PetaFLOPS with an impressive chip utilization rate of 82%. This level of performance， from domestically developed hardware， is a significant achievement in building a self-sufficient AI infrastructure.
According to Huawei’s internal laboratory statistics， this performance is approximately 91% of that achieved by NVIDIA’s previous generation A100 training clusters. While the authenticity of these precise figures requires further observation， the reported capabilities indicate a substantial leap forward in both the power and independence of China’s AI development.
