Suanmiao Technology is building a specialized processor for companies that deploy large AI models, handling the inference work that turns a user’s prompt into a stream of tokens. ChinaVenture reports that the Beijing startup recently completed a Pre-A2 and Series A led by Yizhuang Guotou and other state-backed Beijing investors, taking its reported financing over the past year to nearly $280 million.
The first processor, A4E, puts eight memory wafers above a logic wafer using 3D hybrid bonding. The design is intended to feed model data to the computing layer with less movement than a conventional, separately packaged setup, and Suanmiao is targeting 16 terabytes per second of memory bandwidth. Its software stack is designed to help developers map inference workloads onto that hardware.
That is the company’s answer to the “memory wall”: processors can sit idle while model weights and intermediate data travel to and from memory. Suanmiao is not pitching A4E as a general-purpose graphics processor. Founder Wang Fuquan told ChinaVenture that the chip is optimized for token-sequence processing across different large models, rather than tied to one model.
The product is still on the road to production. A4E taped out on June 15, meaning its design moved into fabrication; Suanmiao says volume production is planned for 2027. The company also says it has spent about a year developing the chip with unnamed leading model developers, but neither report identifies a shipping customer or commercial deployment.
Suanmiao was founded in November 2022, but its core team had already developed and mass-produced an earlier 3D-stacked computing chip. ChinaVenture says the team produced more than 10,000 3D integrated-circuit wafers before shifting the architecture toward AI inference. That experience makes the manufacturing plan more concrete, though it does not show that A4E itself has reached volume production.
The new capital is earmarked for research hiring, core-technology work and A4E production, according to ChinaVenture. One distinction matters: the nearly $280 million figure is cumulative across several financings in the past year, not a disclosed total for the newly reported Pre-A2 and Series A. Their standalone sizes remain undisclosed.
Published versions remain available when an article is updated or corrected.
- Revision 1 · Initial publication
Initial publication