TechFlow reports, on July 20, according to analyst Sunny Bangia, similar to the launch of DeepSeek R1 in 2025, the launch of Kimi K3 has also triggered concerns that AI compute will be lower than expected. However, this comparison may overlook an important distinction. Although models like Kimi K3 are designed to utilize compute resources more efficiently, they still require significant memory during operation, 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 the sparsity rate to a new high. A higher sparsity rate means that the number of parameters activated in each task is smaller relative to the total model size, thereby achieving higher computational efficiency per token. Parameters are just numbers stored in memory. Even if the model is compressed using low-precision data formats, Kimi K3's parameters will still occupy about 1.4 terabytes of memory. Adopting K3 itself will generate demand for high-memory silicon chip upgrades, such as NVIDIA's Blackwell GB300. (Jin10)
