Kimi K3’s Massive Size Sustains Strong Demand for AI Memory and Processors

When DeepSeek’s R1 model debuted in early 2025, Nvidia’s market value dropped by nearly $600 billion in a single day amid fears that artificial intelligence (AI) would require less computing power than anticipated. Moonshot AI’s launch of Kimi K3 on Friday triggered a similar reaction, contributing to a sharp decline in semiconductor stocks.

However, this comparison may overlook a key distinction. DeepSeek’s breakthrough focused on reducing the costs of training and running AI models. While Kimi K3 also improves computational efficiency, it does so much larger model that places greater demands on memory infrastructure. This could continue to drive demand for SK Hynix’s high-bandwidth memory, Nvidia’s latest AI systems, and the advanced chip manufacturing services of Taiwan Semiconductor Manufacturing Co. (TSMC).

Kimi K3 features 2.8 trillion parameters—the largest model in China to date—raising the sparsity ratio (a measure of computational efficiency) to a record level, according to data compiled by Bloomberg from information released by the model’s creators. A higher ratio implies that fewer parameters are activated for each task relative to the model’s total size.

Nevertheless, each of those 2.8 trillion parameters must be stored in memory. Even after compressing the model using lower-precision data formats, Kimi K3 occupies approximately 1.4 terabytes of memory. Consequently, their deployment requires AI processor clusters with high memory capacity, such as Nvidia’s Blackwell GB300 systems.

Another major milestone is approaching. On July 27, Moonshot plans to release the weights for the Kimi K3 model, enabling companies to run the model themselves rather than accessing it cloud service. Although the software will be available for free, organizations wishing to deploy it at scale will still require substantial AI hardware infrastructure. Meanwhile, Alibaba Group Holding also unveiled a preview version of Qwen 3.8-Max over the weekend—a 2.4-trillion-parameter model offering performance comparable to leading state-of-the-art models—though it has not yet released benchmark results. As Chinese AI developers increasingly tailor their models to domestic chips, the country’s hardware ecosystem could stand to benefit. This optimism boosted China’s ChiNext index—which focuses on technology companies—on Monday, driving it up by as much as 3.6%.

Taken together, these launches suggest that competition in the advanced AI model space is expanding beyond leading U.S. labs like OpenAI and Anthropic PBC. This could put pressure on the pricing power of U.S. model providers while continuing to drive investment in the infrastructure needed to run increasingly capable AI systems.