China’s new AI chip startup uses 3D stacking to bypass US restrictions

A New Entrant in China's AI Silicon Race

A previously stealth Chinese semiconductor startup, Dongfang Suanxin, has surfaced with a strategy that reflects the broader constraints shaping China's domestic chip industry: rather than chasing the most advanced transistor nodes directly, the company is betting on 3D stacking technology to extract more AI compute performance from chips built on less cutting-edge manufacturing processes. The move is a direct response to years of tightening US export controls that have cut off Chinese firms from top-tier lithography tools and advanced foundry access abroad.

Why 3D Stacking Matters as a Workaround

US restrictions have primarily targeted the most advanced process nodes and the equipment needed to produce them, notably extreme ultraviolet (EUV) lithography systems. Without reliable access to these tools, Chinese chipmakers face a ceiling on how small and dense they can make individual transistors. 3D stacking offers an alternative path: by layering multiple dies vertically and connecting them with dense interconnects, companies can effectively multiply compute density and memory bandwidth without needing the latest node. This is conceptually similar to techniques used by leading AI silicon designers globally, including advanced packaging methods that pair logic dies with high-bandwidth memory stacks.

If Dongfang Suanxin can execute on this approach at scale, it would suggest that packaging and integration innovation may partially offset China's disadvantage in front-end fabrication — though it's important to frame this as a promising direction rather than a proven equivalent to leading-edge Western or Taiwanese silicon.

Implications for the AI Datacenter Buildout

China's hyperscalers and state-backed AI initiatives have been aggressively pursuing domestic alternatives to Nvidia GPUs, driven both by export restrictions and by a strategic push for supply-chain self-sufficiency. Startups like Dongfang Suanxin add to a growing ecosystem of custom AI silicon efforts — alongside firms such as Huawei's HiSilicon, Biren, and Moore Threads — all racing to fill the gap left by restricted foreign hardware. A viable stacking-based architecture could feed directly into China's datacenter expansion plans, offering an alternative inference and training substrate that doesn't depend on smuggled or gray-market GPUs.

Cost and Custom Silicon Considerations

For the broader custom AI silicon and TPU-style hardware conversation, this development underscores a pattern already visible in the West: packaging and system-level integration increasingly matter as much as raw transistor scaling. If 3D stacking proves cost-effective, it could lower the barrier for more players to build competitive inference hardware without waiting on next-generation fabs. That said, yield challenges, thermal management, and packaging costs at scale remain significant open questions that will determine whether this approach meaningfully changes AI inference economics in China.