JPMorgan sees China AI chip self-sufficiency hitting 80% by 2028

China will source 80 percent of its AI infrastructure computing from domestic chips by 2028, up from 40 percent this year, JPMorgan projects, as inference demand climbs at an 80 percent compound annual rate.

The bank's analysts, in a report on China's AI infrastructure sector, said the rapid advance of local open-source models toward the global frontier, heavier AI investment by major cloud providers and a fast-developing domestic supply chain are driving the takeoff. Overseas chip restrictions have pushed Chinese chipmakers to adopt advanced packaging, interconnect technologies and system-level optimization to close the performance gap of single chips.

Inference-driven computing demand will account for 85 percent of total demand by 2030, with AI chip unit shipments growing about 40 percent annually to roughly 14 million units. Domestic AI chip supply is projected at about 2 million units in 2026, 3 million in 2027 and 5 million in 2028, according to the report.

The localization push names beneficiaries across the supply chain, including SMIC (00981.HK), ILUVATAR COREX (09903.HK), NAURA (002371.SZ), AMEC (688012.SH) and VNET Group (VNET.US), as foundry and HBM capacity become bottlenecks by 2028.

Foundry and HBM Constraints

Local foundries are expected to add 150,000 to 200,000 wafers of advanced logic capacity per month over the next two to three years, although yield rates and technology migration remain challenging. HBM demand is projected to grow at a 60 percent compound annual rate from 2026 to 2030, and fully meeting local demand by 2030 would require HBM capacity of about 200,000 wafers per month.

The capacity squeeze benefits backend service providers and server ODM/OEM firms, while localization also supports power, cooling and energy-storage suppliers. Incremental AI-related power demand from 2026 to 2030 is expected to grow at a 50 percent annual rate, driving shifts toward liquid cooling, 800V DC and solid-state transformer architectures and prefabricated power modules.

Who Wins From Localization

JPMorgan's top picks include ILUVATAR COREX, NAURA, AMEC, CHANG ELEC TECH (600584.SH), V-TEST TECH (688372.SH), HUAQIN (603296.SH), VNET Group, SMIC, MINTH GROUP (00425.HK), WEICHAI POWER (000338.SZ) and WASION HOLDINGS (03393.HK).

Demand is being pulled by Chinese open-source models — Alibaba's Qwen and Zhipu AI's ChatGLM among them — that are closing the gap with global frontier systems and driving inference token consumption. JPMorgan expects inference to dominate total computing demand, reaching 85 percent by 2030, as cloud providers scale deployments.

The projections align with China's broader push to build an indigenous semiconductor base. Chinese firms added more chipmaking capacity in 2024 than the rest of the world combined, with 1 million more wafers a month than in 2023, according to the Information Technology and Innovation Foundation. Yet domestic equipment vendors still satisfy less than 14 percent of back-end demand, per SemiVision's supply-chain analysis, leaving room for the localization ramp JPMorgan models.

For investors, the question is whether the 80 percent self-sufficiency target is achievable given yield and equipment constraints. SMIC, a proxy for China's advanced-node ambitions, faces the steepest climb — its 7nm production relies on more expensive, lower-yield processes than TSMC's leading-edge nodes. If JPMorgan's supply estimates hold, domestic chip supply of 5 million units by 2028 would still fall short of the roughly 14 million units of demand the bank projects, implying continued reliance on imports and a persistent gap for suppliers to close.

This article is for informational purposes only and does not constitute investment advice.