▲ 'Kimi k3'
As 'Kimi k3,' a massive open-weight model released by Chinese AI startup Moonshot AI, gains industry attention, forecasts suggest that consumers will need a substantial amount of memory semiconductors to utilize this model.
Bloomberg reported on July 20 (local time) that while stock prices of semiconductor-related companies are plummeting following the release of 'Kimi k3'—much like when DeepSeek's 'R1' was unveiled in early 2025—this reaction may overlook a crucial difference.
The explanation is that 'Kimi k3' is a model that focuses more on memory than on computing (processing).
DeepSeek was previously credited with achieving breakthrough results in reducing the training and operational costs of AI models.
As concerns spread that the semiconductor resources required for AI would be less than initially expected, the market capitalization of Nvidia, often called the bellwether of AI stocks, once evaporated by nearly 600 billion dollars in a single day.
Kimi k3 has 2.8 trillion parameters, making it the largest model to emerge from China.
Its 'sparsity ratio,' an indicator of computing efficiency, is also at an all-time high.
The higher this ratio, the fewer parameters are activated for each task relative to the total size of the model.
However, for this model to run, all 2.8 trillion parameters must be stored in memory.
This means that even if the model is compressed using a low-precision data format, it requires 1.4 terabytes of memory.
Therefore, Bloomberg explained that running this model directly would require an AI rack system on the scale of Nvidia's 'Blackwell GB300' system.
Bloomberg assessed, "With 2.8 trillion parameters and a 1-million-token 'context window,' k3 demands significantly more memory capacity than previous generations," adding, "This suggests sustained demand in a sector dominated by a few players, such as SK Hynix and Samsung Electronics."
Stanley Tang, a senior portfolio manager at Sumitomo Mitsui DS Asset Management, predicted that considering the limited number of suppliers, memory companies will likely remain in the most advantageous position in the semiconductor industry. He added that if models like Kimi gain popularity, they could accelerate the adoption of AI agents, meaning overall demand is unlikely to decrease.
Moonshot plans to release the model weights for Kimi k3 on July 27.
(Photo: AP, Yonhap News)
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.