Open-source and open-weight models are not a threat to memory demand, but are expanding memory TAM

Overview

This report focuses on whether open weights and Chinese LLMs will weaken memory demand. BofA reaches the opposite conclusion: low-priced APIs do not mean lower hardware demand, and open models expand the total addressable market for HBM, DRAM, and NAND through local deployment of model weights, self-hosting, and more inference endpoints. The report uses models such as Kimi K3, DeepSeek, Qwen, GLM, and gpt-oss as case studies, and maps the theme to the investment view on Micron Technology, Inc.

Core views

The core views are: first, open models still must keep model weights resident in fast memory, and fewer active parameters mainly reduce compute rather than total weight memory; second, the pricing advantage of Chinese models comes more from MoE sparsity, MLA/KDA attention compression, low-bit quantization, lower power/labor/land costs, and possible subsidies; third, open-weight models split the centralized memory demand of closed APIs into a large number of enterprise, government, and cloud self-hosted endpoints, thereby replicating more HBM pools and KV cache capacity; fourth, model size, context windows, and multi-agent workloads are growing faster than efficiency improvements, and the Jevons Paradox keeps total demand expanding; fifth, CXMT poses limited competitive threat in AI HBM3E/HBM4, and MU remains a primary beneficiary.

Analysis framework

The report adopts a theme-driven industry research approach, comparing open/closed model deployment patterns, model parameter scale, weight memory, API pricing, architectural efficiency, and infrastructure costs, then applying the conclusions to memory demand and MU valuation. The analysis also uses model lists, pricing and weight-memory tables, AI rankings, token usage growth, cases such as Kimi K3 and DeepSeek, as well as Micron's sum-of-the-parts valuation and buyback scenarios.

Methodology notes

• Industry chain demand frameworkOpen-weight endpoint expansion logic

  Open models diffuse weight deployment from centralized cloud environments to multiple customer self-hosted endpoints.

  Closed models share the same HBM weight pool across a small number of cloud data centers; after enterprises, governments, and cloud providers download open-weight models, each deployment requires independent HBM, DRAM, NAND, and KV cache capacity, so endpoint expansion amplifies memory demand.

• Technical efficiency analysisSeparation of MoE and weight memory

  Fewer active parameters reduce compute, but total parameters determine resident weight memory.

  Kimi K3 activates only 50B parameters per token, but MoE routing dynamically selects experts, so the full expert pool still needs to reside in nearby fast memory, making HBM capacity closer to total parameters than active parameters.

• Demand elasticity frameworkJevons Paradox

  Efficiency gains reduce unit costs, but may trigger greater total usage.

  The report believes low-cost APIs, lower model inference costs, and multi-agent workloads will drive more tokens and more deployments, with overall demand growth exceeding the efficiency gains from MoE, quantization, and KV cache compression.

• Valuation methodSum-of-the-parts valuation

  Sum-of-the-parts valuation estimates traditional cyclical memory and AI HBM businesses separately.

  BofA's US$1,550 target price is based on SOTP: the traditional cyclical memory business is valued at 3x CY28E P/B, while the AI HBM business is valued at 31x CY28E PE.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Micron Technology, Inc (MU)

  Core beneficiary target

  Strengths

  Benefits from rising demand for HBM, DRAM, and NAND; the AI HBM business is included in high-multiple valuation; large-scale buybacks may resume after CHIPS Act buyback restrictions expire.

  Weaknesses

  Still exposed to memory ASP cycles, end-demand volatility, and intensifying competition.

  Comparison

  The report believes CXMT is mainly focused on consumer and commodity DRAM and has not yet become a direct threat in HBM3E/HBM4.

  Risks

  Memory ASP declines greater than expected, competition from new Chinese entrants, loss of share to major competitors, and weakening demand in data centers, smartphones, and PCs.

• HBM / HBM3E / HBM4

  Direct beneficiary segment of open-model deployment and AI inference expansion

  Strengths

  Large open-weight models need to keep weights and KV cache resident in high-speed memory, and endpoint replication increases the number of HBM pools.

  Weaknesses

  Efficiency technologies and low-bit quantization reduce the memory-bit demand per model unit.

  Comparison

  Closed models are deployed centrally and share HBM; open models replicate weights across multiple customer clusters.

  Risks

  If open-model commercialization underperforms expectations or customers shift back to closed APIs, endpoint expansion may be slower than expected.

• Chinese open-weight LLMs：Kimi、DeepSeek、Qwen、GLM、MiMo

  Demand expansion catalyst and also a source of price disruption

  Strengths

  Low API pricing, open weights, and capability improvements approaching frontier models drive token usage and self-hosted deployment.

  Weaknesses

  Low prices may depend on subsidies, loss-leading customer acquisition, or capacity constraints, and may not reflect a sustainable cost structure.

  Comparison

  Compared with closed models from Anthropic, OpenAI, and Google, Chinese open models expand adoption through lower prices and open deployment.

  Risks

  Subscription suspensions, insufficient GPU capacity, unsustainable subsidies, and policy and supply-chain constraints.

• CXMT

  Potential competitive variable

  Strengths

  Expansion of China's domestic memory capacity, with global wafer capacity share already reaching the low double digits.

  Weaknesses

  The report believes its focus remains on consumer and commodity DRAM rather than AI HBM3E/HBM4.

  Comparison

  Compared with MU, CXMT's direct competitiveness in high-end AI memory is judged by the report to be limited.

  Risks

  If it achieves technological breakthroughs or is approved to enter U.S. OEM supply chains, competitive pressure may rise.

Key data

• Kimi K3 weight memoryabout 1.4TB HBM / 64+ accelerators per serving instanceThe report argues this demand is unrelated to Moonshot's low API pricing and reflects the hardware required to run the model.
• Kimi K3 model scale2.8T total parameters；about 50B parameters activated per tokenMoE reduces inference compute intensity, but total weights still need to reside in memory.
• Open model scale growthDeepSeek-V3 671B；Kimi K2 1T；DeepSeek-V4-Pro 1.6T；Kimi K3 2.8TThe report emphasizes that frontier open-weight model parameter scale continues to rise, directly driving weight memory demand higher.
• Chinese model API pricingcan be about 5-350x lower than Western models；Kimi K3 is US$3/M input、US$15/M outputBofA believes the low pricing reflects business models and subsidies more than a proportional decline in hardware costs.
• Architectural efficiency advantageabout 2-3x per tokenSources include MoE sparsity, MLA/KDA attention, and low-bit quantization such as FP8/MXFP4.
• Infrastructure cost advantageabout 1.5-2x per tokenPower, labor, land, and data-center-related costs in China are lower than in U.S. hyperscaler regions.
• Global token usageaverage +6% per week since 2025；average +9% per week since 2026The report states that Chinese models contributed about 70% of tokens in July 2026.
• MU target price and current pricetarget price US$1,550；current price US$865.46Implied upside is about 79%, with a Buy / C-1-7 rating.
• Potential MU buybackabout US$50-60bn/year；about 5-6% of US$1Tn market cap per yearBased on a scenario of US$120-130bn/year+ FCF outlook and a 40% payout policy.

Impact & implications

The investment implication is that open weights and low-priced Chinese models are not a negative shock for memory, but may instead expand the AI memory demand curve. Beneficiary areas include HBM3E/HBM4, server DRAM, LPDDR5X/DDR5, enterprise NAND, and AI inference-side storage tiers. For MU, the report treats open-model endpoint expansion, AI HBM demand, CXMT's limited competitiveness in high-end AI memory, and the future resumption of buybacks as the core reasons for reiterating Buy.

Risks

• Memory ASP declines more than expected.
• New Chinese entrants bring stronger competition.
• MU loses market share to major competitors.
• Demand weakens in major end markets such as data centers, smartphones, or PCs.
• The low pricing of open models may come from subsidies or loss-leading customer acquisition; if subsidies are reduced, the pace of demand expansion may change.
• If model efficiency improvements, quantization, and KV cache compression exceed endpoint growth and workload growth, unit memory demand may weaken.

What to watch

• Download volume, self-hosted deployment volume, and actual inference capacity of open-weight models such as Kimi K3, DeepSeek, Qwen, and GLM.
• Whether Chinese model API prices remain low, and whether low prices are accompanied by capacity limits or subscription suspensions.
• HBM3E/HBM4 supply-demand balance, pricing, customer qualification, and changes in MU share.
• Whether CXMT advances from consumer/commodity DRAM into AI HBM.
• The timing of expiration of MU's CHIPS Act buyback restrictions, FCF performance, and buyback policy.
• The impact of multi-agent, long-context, and reasoning workloads on tokens per query and KV cache capacity.