Open models expand memory demand; MU remains a Buy
AI summary card
Open models expand memory demand; MU remains a Buy
Bank of America believes that low API prices for Chinese open-weight large models do not imply a corresponding decline in hardware costs; instead, larger model weights, greater HBM capacity, and more customer-side deployment endpoints expand the memory-chip TAM.
- Chinese open-weight large-model APIs such as Kimi K3 are inexpensive, but a single-instance deployment still requires approximately 1.4 TB of HBM and 64+ accelerators; lower prices do not eliminate memory demand.
- MoE, quantization, and KV-cache compression primarily reduce computational intensity or improve throughput; they do not change the fact that total model weights must remain resident in high-speed memory.
- Open-weight models allow every enterprise, government, or cloud customer to replicate a local memory footprint, whereas closed-source API models centrally share model weights across a limited number of data centers.
- Bank of America reiterates its Buy rating and $1,550 target price for MU, believing CXMT is primarily a near-term threat to commodity DRAM rather than to HBM3E/HBM4.
Report interpretation
Overview
This report discusses memory-chip investment opportunities in the U.S. semiconductor industry, focusing on why Chinese open-weight large models and low-priced APIs will not weaken memory demand but may instead expand total HBM, DRAM, and NAND demand through larger model weights, longer context windows, and more self-hosted deployment endpoints. The report maps this thesis to Micron Technology, Inc. (MU) and maintains a Buy rating.
Core views
The core view is that open models lower the barrier to use and expand inference workloads, while hardware memory demand remains determined by total parameter count, resident weights, KV cache, and the number of deployment endpoints. The pricing advantage of Chinese models is driven more by MoE sparsity, MLA/KDA and other attention optimizations, low-bit quantization, lower power and labor costs, subsidies, or share-taking pricing than by a proportional decline in HBM/DRAM/NAND costs. As open-weight models such as Kimi K3 and DeepSeek continue to scale, the open-source ecosystem will replicate memory demand previously concentrated in closed-source APIs across customer-side hardware, creating incremental upside for the memory industry.
Analysis framework
The report analyzes the issue across four dimensions: model economics, model architecture, hardware-memory mapping, and deployment models. It first compares API pricing between Chinese and Western models and then breaks down the sources of cost advantages. It subsequently uses total parameters, active parameters, quantization precision, and HBM capacity to show that computational efficiency and memory capacity are not the same variable. Finally, it compares the memory-demand implications of centralized closed-source API deployment with multi-endpoint self-hosted deployment of open-weight models, linking the conclusions to MU, CXMT, and the competitive landscape for AI memory.
Methodology notes
Open-weight endpoint replication
Closed-source models are typically hosted centrally by a small number of cloud data centers, while open-weight models are downloaded and deployed on the hardware of enterprises, governments, and cloud customers. This causes the same model weights to remain resident across multiple HBM pools, thereby amplifying memory demand.
Separation of total and active MoE parameters
MoE can reduce the number of experts activated per token and the computational workload, but routing dynamically selects experts. Therefore, the full expert pool still needs to remain resident near the compute units, making HBM capacity more closely related to total parameters than to active parameters.
Jevons Paradox
Efficiency gains and low-priced APIs reduce the cost per use, thereby stimulating more inference, agentic, and long-context workloads. Demand growth may exceed the savings generated by efficiency improvements.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Micron Technology, Inc (MU)Core beneficiary
- Strengths
- Exposure to HBM, DRAM, and NAND; benefits from multi-endpoint memory replication driven by open-weight models. The expiration of CHIPS Act buyback restrictions could enable large-scale repurchases.
- Weaknesses
- The traditional memory business remains cyclical and is affected by ASP fluctuations and end-market demand.
- Comparison
- The report believes CXMT is primarily focused on consumer/commodity DRAM in the near term and is not yet a direct competitor in HBM3E/HBM4.
- Risks
- Memory ASP declines exceeding expectations, intensified competition from new Chinese entrants, share losses, and weakening data-center, smartphone, or PC demand.
- CXMTPotential competitor
- Strengths
- Aggressive expansion of domestic Chinese memory capacity, reaching a low-double-digit share of global wafer capacity.
- Weaknesses
- Primarily focused on consumer and commodity DRAM; its AI HBM capabilities and approval to supply U.S. OEMs remain uncertain.
- Comparison
- Compared with MU, the report believes CXMT poses limited direct near-term threat to AI memory.
- Risks
- If technological upgrades or regulatory approvals accelerate, competition in the memory industry could intensify.
- HBM / DRAM / NANDDirectly beneficiary asset classes in the value chain
- Strengths
- Resident model weights, KV cache, long contexts, and multi-endpoint self-hosting all increase memory content.
- Weaknesses
- Demand timing depends on AI deployment intensity, the continued scaling of models, and capital expenditure.
- Comparison
- Compared with standalone GPU compute demand, memory demand is more directly influenced by total parameter count and deployment replication.
- Risks
- If model compression, quantization, and architectural optimization exceed expectations, the incremental memory demand per model could decline.
Key data
- Kimi K3 weight size2.8T parametersThe report states that Kimi K3 was released on July 16, 2026, and was the largest open-weight model at that time.
- HBM requirement for a single Kimi K3 instanceApproximately 1.4 TB HBM, 64+ acceleratorsEven with relatively low API prices, deploying the model still requires large-scale high-bandwidth memory.
- OpenAI oss-120b weight memoryApproximately 63 GBThe report uses this for an approximate comparison with Kimi K3's MXFP4 precision, illustrating that weight memory broadly scales with total parameter count.
- Global token usage growthAverage +6% per week since 2025, average +9% per week since 2026The report cites OpenRouter API data and states that Chinese models accounted for approximately 70% of token usage in July 2026.
- Sources of Chinese model cost advantagesApproximately 2-3x architectural efficiency; approximately 1.5-2x infrastructure cost advantageIncludes MoE sparsity, MLA/KDA, FP8/MXFP4 quantization, and advantages in power, labor, and land costs.
- MU target price$1,550Based on sum-of-the-parts valuation: traditional cyclical memory businesses at approximately 3x CY28E P/B and the AI HBM business at approximately 31x CY28E P/E.
- Potential buyback capacityApproximately $50-60bn per yearThe report assumes MU will generate $120-130bn or more in annual FCF in the coming years; a 40% payout policy would imply annual buybacks equivalent to approximately 5-6% of market capitalization.
- CXMT capacity positionLow-double-digit share of global wafer capacityThe report believes CXMT is primarily focused on consumer/commodity DRAM and is not a direct near-term threat to HBM3E/HBM4.
Impact & implications
If the report's assessment is correct, the adoption of open-weight large models will spread AI infrastructure investment from a small number of closed-source API providers to more enterprise, government, and cloud self-hosting scenarios, increasing memory content in HBM, DRAM, LPDDR5X/DDR5, and enterprise NAND. In terms of asset mapping, memory suppliers—especially MU—will benefit from growth in AI HBM and multi-tier memory demand. The GPU, server, and cloud-services ecosystems will also benefit from expanding inference workloads, although low API price competition could compress model-service-provider profitability in the near term.
Risks
- Memory ASP declines more than expected.
- Competition from new Chinese memory entrants is stronger than expected.
- MU loses market share to larger competitors.
- Demand weakens in major end markets such as data centers, smartphones, or PCs.
- Low pricing for Chinese AI models may be driven by subsidies or loss-making share capture; if unsustainable, the pace of demand expansion could fluctuate.
- Continued rapid advances in model compression, quantization, and KV-cache optimization could partially offset memory demand per deployment.
What to watch
- Parameter sizes, download volumes, and self-hosted deployment volumes of open-weight models such as Kimi K3, DeepSeek, GLM, Qwen, Hunyuan, and MiMo.
- HBM3E/HBM4 supply, pricing, customer qualification, and changes in MU's market share.
- Whether Chinese model API pricing is sustainable, and whether subscription pauses, subsidies, or substantial price cuts continue to occur.
- CXMT's technological progress in HBM or high-end DRAM and U.S. OEM procurement approvals.
- MU's capital return policy after CHIPS Act buyback restrictions are lifted around December 2026.
- The growth rate of token consumption from long-context, agentic, and multi-agent workloads.