China's artificial intelligence (AI) startup "Moonshot AI" released an open-weight (open-weights) model, "Kimi K3," that appears to require a massive amount of memory during operation.
Bloomberg reported on the 20th (local time) that although the share prices of semiconductor-related corporations have fallen since the Kimi K3 was unveiled, the market may be overlooking the model's heavy reliance on memory.
The market is comparing the arrival of Kimi K3 with DeepSeek's AI model "R1," unveiled early last year. DeepSeek was credited with a breakthrough in reducing the expense of training and operating AI models. As concerns spread that fewer semiconductors might be needed for AI than initially expected, Nvidia's market capitalization at the time lost nearly $600 billion in a single day.
Unveiled on the 17th, Kimi K3 is one of the world's largest open-weight AI models, with 2.8 trillion parameters. It also boasts a record-high "sparsity ratio," a metric that indicates computing (processing) efficiency. The higher this ratio, the fewer parameters are activated for each task relative to the model's overall size.
However, to run the model, all parameters must be stored in memory. Even if the model is compressed into a low-precision data format, it means 1.4 terabytes of memory are required.
To run the model directly, users would need large-scale AI server equipment comparable to Nvidia's "Blackwell GB300"-based systems, the report said.
The 1 million-token context window supported by Kimi K3 is another factor that increases memory usage. A context window refers to the amount of information an AI model can remember or process at one time.
Bloomberg said, "With 2.8 trillion parameters and a 1 million-token 'context window,' K3 demands far more memory than prior generations," adding, "This suggests sustained demand in a segment dominated by a few players, including SK hynix and Samsung Electronics."
Stanley Tang, senior portfolio manager at Sumitomo Mitsui DS Asset Management, said that given the limited number of memory suppliers, they are likely to remain in the most advantageous position in the semiconductor industry going forward. He also said that if models like Kimi gain popularity, they could accelerate the spread of AI agents, so overall demand is unlikely to decline.
Meanwhile, Moonshot plans to release Kimi K3's model weights on the 27th.