Tsinghua chip veteran’s US$1.8bn 3D AI chip startup targets China’s GPU gap

Wei Shaojun’s new venture targets the AI training gap left by NVIDIA’s H100 export ban, leveraging 3D stacking to bypass traditional scaling limits. This move pressures domestic EDA vendors to rapidly support TSV and hybrid bonding flows, while pushing SMIC to prioritize 2.5D/3D packaging capacity. However, if the U.S. BIS extends controls to advanced packaging tools, yield ramp and cost targets become highly vulnerable. Competitors like Cambricon may pivot to inference specialization, while NVIDIA could flood the market with downgraded H20 chips to choke high-end alternatives. Within 18 months, China’s AI chip sector will face a make-or-break validation phase: without achieving >70% 3D integration yield, investor enthusiasm will evaporate despite current lofty valuations.