DeepSeek is rumored to launch its R2 AI model, which could feature a hybrid MoE architecture and double the parameters of its predecessor, R1, reaching 1.2 trillion. The model is expected to offer a staggering 97% lower cost per token compared to GPT-4, making it highly competitive in the AI market.Additionally, R2 is said to utilize Huawei's Ascend 910B chip cluster, achieving 512 PetaFLOPS at FP16 precision, indicating a significant vertical integration of the AI supply chain. While these details are speculative, they suggest a potential disruption in the AI landscape.