In Q1 2025 Alibaba, together with ByteDance and Tencent, pre-ordered more than $16 billion of NVIDIA H20 chips. When the US put H20 under export licensing in April, those orders were effectively blocked; Alibaba accelerated substitution with its in-house T-Head chip and announced RMB 380 billion of AI infrastructure investment over three years. A year on, the return distribution across this chain points one way: the spender fell, the recipients rose. Alibaba is -24% in 2026, while equipment names Naura (+69%) and AMEC (+86%), designer Cambricon (+66%) and memory name GigaDevice (+95%) lead.
A forced conversion
Export controls did not remove demand. They converted it, by force, from imports into self-sufficiency investment. Unable to buy, cloud chip budgets flow down to domestic design (Cambricon, Hygon, T-Head), domestic foundry (SMIC, Hua Hong), and the domestic equipment that fills those fabs (Naura, AMEC). Hua Hong exploding +412% in 2025 then +59% in 2026, while SMIC cooled to +5.6% after +170% in 2024, signals propagation moving one layer at a time from foundry to equipment within this chain.
The shape mirrors the US in 2023-24, when the hyperscalers spending on AI lagged NVIDIA and the supply layer receiving that spend. The Chinese difference is the trigger: policy rather than market, which ties this clock to a regulatory calendar.
Risks to this clock
The weakness of forced conversion is efficiency. Even if domestic chips match H20-class performance as reported, the software-ecosystem and yield gap is a variable returns have not priced. And the moment controls loosen, the clock runs backward: resumed H20-class supply drains the self-sufficiency premium first from the most stretched layer, which today is design. If controls tighten instead, the next propagation moves down to materials and components that have risen least.
The summary as of now: the reliable quantity in China's AI chain is not cloud earnings but cloud spending, and policy has pinned that spending at home.