Future of Tech: Can China Achieve AI Supremacy?
Summaries are public. PDF access requires an active membership; all members have the same access. Sign in
Bernstein examines whether China can match or exceed United States AI compute capacity by 2035. Presently, US AI compute stands at 35 ZFLOPs (FP8), while China registers only 5 ZFLOPs, representing merely 15% of US capacity. However, China possesses a formidable structural advantage in power infrastructure. The nation adds over 500GW of generation capacity annually, roughly ten times the US rate.
The analysis suggests China could overcome its semiconductor efficiency deficit through sheer scale. If domestic chip efficiency improves from the current 25% of US levels to over 50% by 2035, parity becomes feasible. Achieving this requires expanding AI data center capacity from 14GW to 214GW, alongside sustained annual generation additions of 385GW. Furthermore, AIDC capital expenditure must scale to $974 billion by 2035, representing a 32% compound annual growth rate.
The gap between SMIC and TSMC remains approximately six years, though TSMC performance gains are decelerating as Moore's Law slows. China maintains dominant positions in solar, wind, nuclear, and battery component manufacturing, with costs roughly one-third of Western equivalents. If power represents the sole bottleneck, China could theoretically achieve 1,936 ZFLOPs by 2035, dwarfing US projections of 511 ZFLOPs.
Beneficiaries include CATL, rated Outperform with a 530.00 CNY target; Sungrow, Outperform at 260.00 RMB; SMIC-A, Outperform at 150.00 CNY; SMIC-H, Outperform at 100.00 HKD; Cambricon, Outperform at 2,000.00 CNY; and Hygon, Outperform at 280.00 CNY. Key verification milestones include 2030 chip performance and power generation scalability.