Bank of America believes that China's AI value chain is developing a "dumbbell‑shaped" profit structure: value is concentrated at both ends—on strategic bottleneck segments such as AI accelerators, semiconductor equipment, foundry services, and memory, as well as on large cloud platforms that benefit from scale advantages and ecosystem network effects. Meanwhile, independent large‑model research labs and physical‑AI manufacturers situated in the middle face mounting pressure from low switching costs, rapid imitation, and weak pricing power.
China's AI localization process has evolved from a policy objective into a commercial reality, yet the gap between market‑share expansion and profit growth is widening. As domestic accelerator, storage, and hardware vendors rapidly capture market share in China, capital‑intensive inflows have fueled redundant investments and price competition, making "who is truly making money along the value chain" a far more critical investment question than "how far localization has progressed."
According to Bank of America Merrill Lynch's China AI Strategy Report released on September 17, China's AI value chain is evolving into a "dumbbell‑shaped" profit structure: value is concentrated at both ends—on strategic bottleneck segments such as AI accelerators, semiconductor equipment, foundry services, and memory—as well as on large cloud platforms that benefit from scale advantages and ecosystem network effects. Meanwhile, independent large‑model research labs and physical AI manufacturers situated in the middle face mounting pressure from low switching costs, rapid imitation, and weak pricing power.
The direct implication of this assessment for investors is that China's AI sector is not a homogenous theme, but rather a portfolio of three distinct categories with markedly different risk–return profiles. Bank of America Merrill Lynch has grouped relevant equities into three buckets—domestic‑beneficiary stocks, infrastructure exporters, and AI application firms—and warns that if U.S. AI capital spending slows, the nature and magnitude of the impact on these three groups will diverge significantly.
Localization Progress: Market share is genuine, but the cutting-edge gap remains.
The expansion of domestic AI accelerator market share has been confirmed at the revenue level.
According to estimates from IDC, company reports, and Bank of America Merrill Lynch, by 2025 China's accelerator market will see domestic vendors collectively accounting for approximately 50%, a significant increase from less than 30% in 2024, with projections indicating this share could approach 80% by 2028. Huawei enjoys a clear roadmap advantage on the training side.$Cambricon (688256.SH)$、$ILUVATAR COREX (09903.HK)$、$Moore Threads Technology (688795.SH)$Meanwhile, other vendors are gradually building competitive advantages on the inference side.
The catch-up in the storage sector has been equally pronounced but uneven. According to company reports and estimates from Bank of America Merrill Lynch, ChangXin Memory's global DRAM revenue share has risen from roughly 3%–4% in 2024 to about 10% as of the second quarter of 2026, while Yangtze Memory's NAND market share has increased from 7%–8% to approximately 14%.
In the semiconductor equipment sector, according to data from SEMI.org, China's equipment spending is projected to reach US$49.3 billion by 2025, representing a five-year compound annual growth rate of 21% compared with US$19.0 billion in 2020. China has been the world's largest equipment market for six consecutive years. Domestic suppliers currently hold about 30% of the market share, up from 22% in 2019; however, advanced lithography, ion implantation, and metrology and inspection remain key bottlenecks—domestic penetration in lithography equipment remains below 5%, and the gap with ASML is unlikely to close in the near term.
China's core strategy for addressing hardware constraints is system-level optimization rather than isolated breakthroughs: it seeks to close the performance gap of individual chips by leveraging larger-scale domestic computing clusters, faster interconnects, co‑design of software and hardware, and more efficient model architectures. This approach has already been validated in recent large‑model training and deployment efforts; however, it also implies that localized competitive advantages lie primarily at the system level, rather than in directly comparable single-chip specifications.
The Smiling Curve: Why Value Is Concentrated at Both Ends
The core framework outlined in the Bank of America Merrill Lynch report is that the profit distribution across China's AI value chain follows a "smile curve" pattern—high at both ends and low in the middle.
The moat at the strategic bottleneck stems from capital intensity, lengthy development cycles, deep technological expertise, and regulatory barriers.
In the wafer foundry sector,$SMIC (00981.HK)$The 8-inch‑equivalent blended average selling price has risen from $904 in the first half of 2025 to $965 in the first half of 2026, while HuaHong's has increased from $437 to $461. In the second quarter of 2026, capacity utilization rates reached 94% and 103%, respectively, reflecting a marked improvement in pricing discipline.
In terms of storage,$CXMT Corporation (688825.SH)$In the first half of 2026, the operating profit margin is expected to reach 79%, broadly on par with global industry leaders. However, storage profitability remains highly sensitive to new capacity additions, and cyclical risks cannot be overlooked.
On the equipment vendor side,$NAURA Technology Group (002371.SZ)$and$Advanced Micro-Fabrication Equipment Inc. China (688012.SH)$In 2025, gross margins will face pressure due to the introduction of new products; however, both companies remain profitable and have set long-term gross margin targets above 40%. The key determinant will be their ability to translate current investments into a larger installed base and sustained service revenue.
The advantages of a large-scale platform stem from diverse use cases, cross-selling opportunities, and network effects.
Alibaba Cloud's revenue growth has accelerated from 36% year over year in the December 2025 quarter to 45% in the June 2026 quarter, with AI product revenues posting triple-digit growth.$KINGSOFT CLOUD (03896.HK)$In the first half of 2026, revenue grew 34% year over year, with AI cloud‑billing revenue surging 85% and accounting for 40% of total revenue. However, cloud platforms remain capital‑intensive—$Alibaba (BABA.US)$In the latest quarter, capital expenditures increased by 75% year over year, and higher depreciation kept Kingsoft Cloud in a loss under GAAP.
The situation of the middle class, by contrast, is entirely different.
As of June 2026, China has registered over 980 generative AI services and applications—more than 15 times the number at the end of 2023—yet the pace of supply expansion far outstrips its proven monetization potential. The token‑pricing war has driven down DeepSeek V4.1 Flash's off‑peak input and output rates to $0.15 and $0.60 per million tokens, respectively, while Anthropic's Claude Fable 5.1 charges $10 and $50 per million tokens in the same periods.
Basic reasoning capabilities are increasingly commodified, leaving independent large‑model research labs under triple pressure: low switching costs, rapid imitation, and substantial R&D investment. As a result, both pricing power and profitability are structurally constrained.$NVIDIA (NVDA.US)$Its operating profit margin has long exceeded 60%, whereas Cambricon, although expected to turn profitable in 2025, is forecast by Bloomberg consensus to maintain an operating profit margin of just over 30% by 2028—still a significant gap compared with NVIDIA.
Physical AI manufacturers face a similar predicament.
China's humanoid robot market has entered a phase of "involution," with approximately 140 manufacturers set to launch over 330 models by 2025. The average price of Ubtech Robotics' humanoid robots is expected to decline by about 72% between 2023 and 2025, while the average price of Leju Robotics' Kuavo series is projected to fall by 26% by 2025.
How are risk and return distinguished?
Bank of America Merrill Lynch has categorized China's AI‑related stocks into three investment themes, each with distinct risk‑return profiles.
Local‑content beneficiaries—including semiconductor equipment (NAURA, AMEC), AI accelerators (Cambricon), memory (CXMT), and foundry services ($SMIC (00981.HK)$— among others — are highly aligned with China's policy of technological self-reliance, offering global investors diversified exposure that differs from the U.S. AI capital‑expenditure cycle. However, these stocks are already widely held by Asian and European investors, with valuations generally elevated; their forward P/E ratios for 2027 typically range from 50 to 150 times. Should U.S. AI capital spending slow, these equities could face a compression in valuation multiples, though the impact on earnings would likely be relatively muted, potentially presenting long-term investors with buying opportunities on dips.
Infrastructure exporters—including optical modules ($ZJ INNOLIGHT (03308.HK)$), PCB/copper-clad laminate ($VGT (02476.HK)$), liquid cooling, AI power supplies, and the like—have largely benefited from the capital‑expenditure cycle of U.S. hyperscale data centers. Over the past three to four years, their stock prices have risen markedly, with valuations remaining relatively reasonable, typically trading at P/E ratios between 20 and 40 times. However, their earnings and share‑price performance are closely tied to the global AI‑spending cycle; should AI‑related capital spending decelerate, these stocks will face dual pressure on both earnings and valuations.
AI application‑related stocks—covering large language models, humanoid robots, EVs/AVs, and consumer and enterprise AI applications—represent the final monetization layer of the value chain. While they hold substantial long-term potential, their near‑term path to profitability remains unclear, and no "killer app" has yet emerged to drive widespread adoption and sustainable profitability. Under a scenario of slower AI growth in the U.S., deeply undervalued, high‑quality stocks—such as major internet platforms—may offer relative defensive appeal, given their light investor holdings, modest valuations, and limited capital‑expenditure pressures.
Verification Window and Key Risks
The establishment of the aforementioned framework hinges on several verifiable conditions.
At the results level, it will be important to monitor whether the domestic accelerator's revenue share progresses along a trajectory from roughly 50% in 2025 toward nearly 80% by 2028, whether contract‑manufacturing average pricing and capacity utilization continue to improve, and whether ChangXin and Yangtze Memory's market share and profitability remain sustainable. At the institutional level, key indicators for assessing whether midstream pricing power can improve will include whether equipment suppliers' gross margins can return to their long‑term target of over 40% as new products undergo validation and installed base expands, and whether the pricing anchor on the large‑model side can shift from a per‑token price to a cost‑per‑completed‑task metric.
Counterfactual conditions also warrant close attention: if the slowdown in U.S. AI capital spending inflicts substantial damage not merely on valuations but on earnings, the rationale for buying on dips among domestic beneficiaries will no longer hold; and if independent large‑model labs and physical AI manufacturers experience meaningful improvements in switching costs and pricing power, the "smile curve" thesis would need to be revised.
Starting in the third quarter of 2026, the earnings season will see a concentrated release of gross margin and capacity utilization data from contract manufacturers, equipment suppliers, and memory chipmakers. Coupled with the year-end wave of IPOs and tender disclosures for domestically produced accelerators, this period will serve as a critical window for validating the aforementioned assessments. For investors, before increasing exposure to China's AI sector, it is first essential to clearly identify the specific types of risks they are willing to assume.
Edited by melody