According to Bloomberg, China is considering a plan to invest approximately $295 billion over the next five years in data center infrastructure to accelerate AI development. Beijing reportedly intends for state-owned telecom operators to build and operate much of this capacity while sourcing at least 80% of the underlying technology—including AI accelerators—from domestic suppliers.
If implemented, the initiative would represent a significant escalation of China’s AI ambitions. Current estimates place 2026 capital expenditures by Chinese hyperscalers—including Alibaba, Baidu, Tencent and ByteDance—at slightly above $100 billion, still well below the $700–800 billion expected from their U.S. counterparts. However, assuming roughly $60 billion of annual government-led investment, a substantial increase in ByteDance’s spending (reportedly as high as $100 billion next year, following an estimated $50–70 billion this year), and continued capex growth across other hyperscalers, total Chinese AI and cloud infrastructure investment could surprise to the upside in 2027 and reach $250–300 billion, implying annual growth of more than 100%!
The timing is particularly noteworthy. Recent developments surrounding DeepSeek’s V4 model have demonstrated a potentially viable path toward a domestic AI hardware ecosystem. By successfully adapting the model to run on Huawei’s Ascend accelerators rather than relying exclusively on Nvidia hardware, DeepSeek provided evidence that advanced Chinese AI models can increasingly be trained and deployed on domestically developed infrastructure. If frontier Chinese models can operate effectively on local hardware, hyperscalers, startups, and government-backed projects will have stronger incentives to adopt domestic accelerators, creating a virtuous cycle of scale, investment, and technological improvement for China’s semiconductor industry.
While Chinese AI chips generally remain manufactured on less advanced process nodes than leading U.S. semiconductors, hardware limitations are increasingly being offset through software optimization. Huawei’s CANN ecosystem, for example, serves as a domestic alternative to Nvidia’s CUDA software stack, while advances in compilers, distributed training frameworks, model architectures, and quantization techniques help maximize performance from available hardware. Although software cannot fully eliminate the advantages of leading-edge manufacturing, it can materially narrow the performance gap for many real-world AI workloads.
Taken together, these developments point to a powerful growth cycle for China’s semiconductor and AI infrastructure ecosystem. Beneficiaries span the entire value chain, from foundry and equipment leaders such as SMIC and NAURA Technology, to AI accelerator developers such as Huawei, Cambricon and Moore Threads, as well as memory producers like ChangXin Memory and optical networking suppliers such as Zhongji Innolight.
Against this backdrop, forecasts that China’s semiconductor self-sufficiency could rise from roughly 41% in 2025 to near 90% by 2030 appear increasingly plausible. While such targets remain ambitious and significant technological gaps persist in several critical areas, the combination of state support, hyperscaler investment, and a rapidly maturing domestic AI stack suggests that China’s semiconductor ecosystem may be entering a new phase of accelerated development.
