Chinese Open LLMs are no threat to memory demand, BofA says
Investing.com -- Bank of America told clients in a note Tuesday that the latest wave of Chinese open-source AI models, including Moonshot's Kimi K3, reinforces rather than threatens its bullish thesis on memory demand, reiterating a Buy rating on Micron with a $1,550 price target.
Analyst Vivek Arya said aggressive Chinese open-model API pricing, up to 5-350 times cheaper than Western models, "reflect[s] business-model choices, not necessarily reflective of hardware costs."
The firm noted these models can reduce compute and GPU intensity through efficiency techniques, but "require the same or more memory as their model weights and active parameters increase."
Using Kimi K3 as an example, BofA noted the model still requires approximately 1.4 TB of HBM per serving instance, comparable to OpenAI's smaller open model when adjusted for parameter count, concluding that API pricing "actually scales with total memory weight and active parameters, even for open Chinese LLMs."
BofA added that "'Open' refers to weight distribution, not deployment cost," since customers still need to purchase HBM, DRAM and NAND to run these models on their own hardware.
BofA attributed the pricing advantage of Chinese models largely to architectural efficiency techniques and lower Chinese electricity, labor and land costs, along with potential state-linked capital subsidies.
On competition, BofA believes China's CXMT is not a threat to AI memory, as it "primarily addresses the underserved consumer/commodity DRAM segment, not HBM3E/HBM4."
The firm also highlighted Micron's CHIPS Act buyback restrictions expiring around December 2026, potentially opening the door to $50-60 billion in annual repurchases.
