Analyst: Although Kimi K3 improves computing efficiency, high memory requirements still support demand for AI chips

Deep Tide TechFlow news: On July 20, according to analyst Sunny Bangia’s view, with DeepSeek R1 expected to be released in 2025, the launch of Kimi K3 has also sparked concerns that AI computing power may fall short of expectations. However, this comparison may overlook an important difference. Although models like Kimi K3 are designed to make more efficient use of compute resources, they still require a large amount of memory during runtime, a characteristic that may continue to support demand for companies such as SK hynix, TSMC, and NVIDIA.

Kimi K3 has 2.8 trillion parameters, bringing sparsity to a new high. Higher sparsity means that, for each task, the number of parameters activated is relatively smaller than the model’s total size, thereby achieving higher computational efficiency per token. Parameters are only numbers stored in memory. Even if the model is compressed using lower-precision data formats, Kimi K3’s parameters still take up about 1.4 terabytes of memory. Adopting K3 itself creates the need to upgrade high-memory silicon chips, such as NVIDIA’s Blackwell GB300. (Jin10)