Kimi K3 and Other Open-Source LLMs Boost AI Memory Demand, BofA Reiterates Buy Rating on Micron
China's open-source LLM APIs are priced extremely competitively, with costs ranging from 5 to 350 times lower than comparable Western models. However, this pricing advantage primarily reflects differences in commercial strategy rather than a significant reduction in hardware costs. Although newer models have improved inference efficiency and reduced GPU compute requirements through architectural optimizations, the continued growth in total and active model parameters has not reduced demand for high-bandwidth memory (HBM), DRAM, and other memory solutions. Instead, memory demand is expected to increase further.
Moreover, every download of an open-source model requires customers to deploy the model locally, creating additional demand for HBM, DRAM, and NAND storage. This deployment-driven demand does not exist for proprietary cloud-based models. Bank of America believes that Chinese memory manufacturer CXMT (ChangXin Memory Technologies) is unlikely to pose a material competitive threat to Micron in the near term. The bank therefore reiterated its Buy rating on Micron while maintaining its US$155 target price.
Last week, Moonshot AI, a Chinese AI startup backed by Alibaba, officially unveiled Kimi K3. The model features 2.8 trillion parameters and is described by the company as the world's largest open-weight language model. According to Moonshot AI, its performance approaches that of Anthropic's latest flagship model, Fable.
Although CXMT is aggressively expanding production capacity and currently accounts for a low-single-digit to roughly 10% share of global DRAM wafer capacity, its focus remains on consumer and standard DRAM products. The company has yet to enter the high-end AI memory market, including HBM3E and HBM4. In addition, uncertainty remains over whether U.S. original equipment manufacturers (OEMs) will receive government approval to procure CXMT products in the near term.
Bank of America also noted that Micron's share repurchase restrictions under the U.S. CHIPS and Science Act are expected to expire around December 2026. Once those restrictions are lifted, the company could generate annual free cash flow exceeding US$12–13 billion over the coming years. Assuming Micron maintains its policy of returning approximately 40% of free cash flow to shareholders, annual share repurchases could reach US$5–6 billion, equivalent to roughly 5%–6% of the company's current market capitalization of approximately US$100 billion.
The report further emphasized that API pricing is primarily a reflection of commercial strategy rather than underlying hardware costs. For example, each inference instance of Kimi K3 still requires approximately 1.4TB of HBM to operate. By comparison, OpenAI's open-source oss-120b model contains only about one twenty-third the number of parameters of Kimi K3, yet still requires around 63GB of memory to store its model weights during deployment.
Market Insight:
API pricing remains closely tied to model size and the number of active parameters, including for Chinese open-source LLMs. Open-source availability simply means that model weights are publicly accessible—it does not reduce deployment costs. Organizations deploying these models locally must still invest in HBM, DRAM, NAND, and other memory hardware, providing continued structural support for AI memory demand.
