The Zhitong Finance App learned that Bank of America's latest research report indicates that the recent release of next-generation open source models, including Kimi K3 launched on Dark Side of the Moon, has further strengthened its long-term bullish logic on the AI storage market and Micron Technology (MU.US). Boosted by positive news, storage concept stocks surged collectively on Tuesday. Micron Technology (MU.US) rose more than 8%, SK Hynix (SKHY.US) rose more than 9%, and SanDisk (SNDK.US) rose more than 10%.
Bank of America analysts led by Vivek Arya said that China's large open source model API pricing is extremely competitive, and the price is 5 to 350 times lower than the Western model, but this is more a reflection of business model choices rather than a sharp drop in hardware costs.
Analysts pointed out that although the new model reduced GPU computing power requirements through architecture optimization and improved inference efficiency, as the model parameter scale and activation parameters continued to grow, the demand for storage resources such as high-bandwidth memory (HBM) and DRAM did not decrease; on the contrary, it further increased. Furthermore, every open source model download means that customers need to deploy the model themselves, thereby adding storage requirements such as HBM, DRAM, and NAND, which are not available in the closed source model.
Last week, Dark Side of the Moon, a Chinese AI startup invested by Alibaba (BABA.US), officially unveiled the Kimi K3. This large model with 2.8 trillion parameters is called the world's largest open source weighting model by the company, and its performance is close to Anthropic's latest flagship model, Fable.
Bank of America believes that Chinese memory chip manufacturer Changxin Storage does not currently pose a substantial threat to Micron, and reaffirms Micron's “buy” rating while maintaining a target price of 1,550 US dollars.
Analysts said that although Changxin Storage is actively expanding production capacity, it currently accounts for a low single digit of about 10% of global DRAM wafer production capacity. It mainly focuses on the consumer grade and standard DRAM markets, and has not entered the high-end AI storage field such as HBM3E and HBM4. Furthermore, there is still great uncertainty about whether US original equipment manufacturers (OEMs) can obtain government approval to purchase Changxin storage products in the short term.
Bank of America also indicated that the stock repurchase restrictions imposed by Micron after receiving government subsidies under the Chip Act are expected to expire around December 2026. With the restrictions lifted, the company's free cash flow is expected to reach more than 12 billion to 13 billion US dollars per year in the next few years. If implemented according to the 40% capital return policy, it is expected that about 5 billion to 6 billion US dollars of stock repurchases can be implemented every year, which is equivalent to 5% to 6% of its current market value of about 1 trillion US dollars.
Regarding the low price strategy of the Chinese AI model that the market is concerned about, Bank of America said that this does not mean that AI infrastructure costs are falling at the same time. For example, Tencent's mixed-element model API input price is only $0.06 per million tokens, while Anthropic Claude Opus 4.8 is about $15; Kimi K3 charges about $3, but the performance is close to the leading international model.
However, analysts emphasized that API prices reflect more of a business strategy than hardware costs. Taking Kimi K3 as an example, each inference instance still requires about 1.4 TB of HBM to run; in contrast, the OpenAI open source model “oss-120b” parameter size is only about one-third that of Kimi K3, and it still requires about 63 GB of model weight memory to run.
Analysts said that the API price is usually still closely related to the model's weight scale and number of activation parameters, even for China's big open source model. “Open source” only means open model weights, not reduced deployment costs. Customers still need to purchase storage hardware such as HBM, DRAM, and NAND to complete local deployment.
Furthermore, Bank of America believes that the price advantage of China's AI models also comes from improved model architecture efficiency and cost advantages. Analysts expect that China has a certain leading edge in model architecture efficiency, and in terms of infrastructure costs such as electricity, labor, and land, it has an advantage of about 1.5 to 2 times compared to overseas. The remaining price differences may come from state capital subsidies and financial support and free cash flow subsidies provided by cloud service providers such as Alibaba, Tencent, and Baidu (BIDU.US).