China's AI Hardware Localization Accelerates as Low-Cost Models Build Moats at Both Ends of the Value Chain
nashnova research
BofA Securities argues that China's AI value chain will concentrate long-term value at hardware bottlenecks and cloud-platform ecosystems, while mid-layer model labs face structural pricing-power pressure.
Domestic AI chip share is set to nearly triple in three years — who leads?
Chinese domestic AI accelerator revenue share rose from under 30% in 2024 to ~50% in 2025; BofA projects it will near 80% by 2028.
Huawei leads the domestic camp at 35.5%; Nvidia still holds 47% overall, but the gap is narrowing steadily.
This means → at this pace, domestic chips shift from understudy to main act within three years, compressing Nvidia's pricing power in China in reverse.
How far has the catch-up gone in memory and equipment?
CXMT's global DRAM revenue share climbed from 3%–4% in 2024 to ~10% by Q2 2026; the company has entered mass production of low-power LPDDR6.
YMTC's NAND share rose from 7%–8% to ~14% over the same period, pushing beyond 232 layers toward 260+.
In plain terms = memory chips are the ammunition depot for AI training — CXMT and YMTC are swapping imported rounds for domestic ones, faster than the market expected.
On the equipment side, domestic share rose from 22% in 2019 to ~30% in 2025; etch and deposition already exceed 40%–50%, but lithography, ion implant, and metrology remain clear weak spots.
Foundry prices are rising — what does that signal?
SMIC's 8-inch-equivalent wafer ASP rose from $904 in H1 2025 to $965 in H1 2026; Hua Hong's moved from $437 to $461 over the same period.
Both ran Q2 utilization at 94% and 103%, respectively.
This means → foundry capacity is demand-pulled, not cost-pushed — downstream AI chip orders are competing for fab slots.
Chinese models are cheap, but are they strong enough?
The core advantage is inference cost-efficiency, not frontier training capability. U.S. hyperscaler capex for 2026 is projected at ~$648 billion versus ~$120 billion for Chinese peers — the gap is still widening.
Zhipu GLM-5.3 scores 45 (7th globally); Kimi K3 scores 44 (9th). Claude Fable 5.1 and GPT-6 Astra lead at 53.
In plain terms = Chinese models achieve a global top-ten ranking on roughly one-fifth the spending, but the ceiling is still visibly above them.
Who gets the most real-world usage?
Per OpenRouter, DeepSeek V4 Flash 0731 led August volume at 31.6 trillion tokens; Tencent Hunyuan 3 followed at 26.2 trillion; Xiaomi MiMo-V2.5 came third at 19.1 trillion.
On Vercel AI, DeepSeek averaged 29.7% of August token volume, above Anthropic's 24.7%.
This reflects a shift from cost advantage to real market share — "cheap and capable" is converting into actual adoption at scale.
Hard at both ends, soft in the middle — where should investors look?
BofA's core thesis: hardware (chips, equipment, foundry, memory) is protected by high capital barriers + long development cycles + Beijing's regulatory shield — a more durable moat.
Cloud platforms and internet ecosystems control the user gateway and carry their own scale advantages.
Mid-layer independent model labs and physical-AI companies face low switching costs, fierce price wars, and limited pricing power.
In plain terms = chipmakers and platform operators can capture margin; those stuck in the middle doing models alone have yet to show anyone can set prices.
市场有风险，内容仅供研究参考，不构成投资建议。