Beijing AI chip startup Suanmiao Technology said on October 14 that it had completed Pre-A2, including an extension, and Series A funding rounds, bringing its total financing over the past year to nearly 2 billion yuan. The rounds were led by municipal- and district-level state-backed investment platforms in Beijing, including E-Town Capital. The company develops cloud-based inference chips for large models. Its approach uses 3D hybrid bonding to stack eight memory wafers vertically on top of a compute logic wafer, addressing the “memory wall” while sidestepping constraints on access to advanced process nodes. Its first product, the A4E, taped out in June.

Xin Dongxi sees 3D-stacked inference chips as an increasingly important route for replacing foreign computing power with domestic alternatives. But the same report also noted that the A4E remains some distance from mass-production delivery and large-scale customer procurement.

3D TokenPU’s Technical Trade-Off: 16 TB/s Bandwidth Without the Most Advanced Process Nodes

According to Suanmiao Technology, the company pioneered its 3D TokenPU architecture, which uses relatively mature domestic process nodes and supply chains. Through 3D hybrid bonding, memory wafers are stacked vertically on top of a logic wafer. Specifically, eight layers of memory wafers are stacked above the compute logic wafer, with through-silicon vias and bumps enabling micron-scale interconnects. This cuts the millimeter-scale transmission distances between conventional chips by two orders of magnitude and delivers memory-access bandwidth of 16 TB/s.

This approach shifts the battleground. Domestic AI chips generally face limited access to advanced process nodes, pushing the industry toward architectural innovation. The A4E is built on Suanmiao’s proprietary RISC-V architecture and in-house IP, and has been jointly optimized with leading large-model customers since the chip-definition stage. For hardware-software co-design, it introduces a Tile-Native design that treats tiles as the basic unit for data movement, storage and computing, allowing data to be moved once and reused multiple times. The hardware natively supports tile-level data scheduling and dynamic switching between precision formats, while the software side includes a compiler toolchain compatible with open-source ecosystems such as LLVM and Triton.

The “memory wall” is the common target across this sector. During large-model inference, model parameters must repeatedly move from video memory to compute units, meaning bandwidth often becomes the bottleneck before computing power does. In a conventional von Neumann architecture, memory and computing are separated. Moving parameters frequently between the two creates memory-access latency, bandwidth pressure and additional power consumption.

A Loongson Pedigree: Seven Years of 3D Expertise Dating Back to CAS Shenglong

Founded in November 2022, Suanmiao Technology is headquartered in Beijing’s Daxing district. The team has deep roots in the Chinese Academy of Sciences. Its central figure, Wang Fuquan, holds a doctorate from and served as a researcher at the State Key Laboratory of the Institute of Acoustics, Chinese Academy of Sciences, where he studied under underwater acoustics physicist and CAS academician Zhang Renhe. He later conducted postdoctoral research at the Institute of Computing Technology, Chinese Academy of Sciences, under Hu Weiwu, Loongson’s chief scientist. According to Xin Dongxi, Wang previously led the development of the JASMINER X4, described as the world’s first high-throughput computing power chip integrating memory and computing, and served as CEO of CAS Shenglong before founding Suanmiao Technology.

CTO Liu Ming also holds a doctorate from the Institute of Computing Technology under Hu Weiwu. He spent 15 years on semiconductor R&D with the Loongson team and has twice successfully developed 3D-stacked chips. Chief Algorithm Scientist Yan Chao has 12 years of experience developing Loongson processors. Chief AI Expert Lou Jianguang studied under CAS academician Tan Tieniu, while Chief Software Architect Wei Yongming founded the open-source MiniGUI and HVML projects.

More notable than the résumés is the team’s engineering track record. According to the company, its core team has nearly seven years of experience in 3D ICs. It previously led the creation of a wafer-on-wafer hybrid-bonding 3D IC supply chain in mainland China, mass-produced more than 10,000 3D IC wafers and generated over 1.2 billion yuan in cumulative revenue from the related business. Suanmiao’s website says the team was the first in the industry to use hybrid bonding to develop and mass-produce 3D-stacked chips. That experience is crucial to determining whether 3D stacking can move from a “technology gamble” to a viable production-line process: the hardest parts of hybrid bonding are not design, but yield, thermal management and testing.

Four Funding Rounds in One Year, With State-Backed Investors Fully On Board

The funding pace has been intense. In February, Suanmiao Technology announced that it had completed consecutive Pre-A and Pre-A1 rounds totaling nearly 1 billion yuan. The Pre-A round was co-led by Source Code Capital and Shixi Capital, with participation from semiconductor industry investors including Lenovo Capital. The Pre-A1 round was led by Xiang He Capital, with state-backed investors including CDB Capital and Beijing Shunxi participating. Including the latest Pre-A2 and Series A rounds, the company has raised nearly 2 billion yuan within a year.

The latest round was led by Beijing municipal- and district-level state-backed investment platforms, including E-Town Capital. According to Suanmiao Technology, the investors will provide funding and policy support at multiple levels while helping the company deepen ties with manufacturing, packaging and large-model application partners in the Beijing E-Town area, as well as secure supply-chain and production-capacity resources. Its investor roster now spans three groups: state-backed platforms such as CDB Capital, Beigong Investment, Beijing Shunxi and Zhongguancun Science City; leading private investment firms including Source Code Capital, Xiang He Capital and GenBridge Capital; and industry funds such as Shixi Capital, Lenovo Capital and Zhiwei Capital.

Data from the same period also reflects the sector’s momentum. China Fund News reported that the 3D chip segment is currently attracting strong interest, with roughly 50 startups already operating in the field. Funding activity has been intense: Dongfang Suanxin completed a Series A+ round at the end of April 2026, reaching a post-money valuation of 12.275 billion yuan, with participation from the National AI Industry Investment Fund; Weina Hexin raised 1 billion yuan in a Series C1 round in September; Qianhe Yibang secured more than 2 billion yuan in Series B financing; and Guangyu Xinchen raised nearly 2 billion yuan within six months.

Industry support has also been repeatedly highlighted. According to Jiwei, Huawei introduced “Tao’s Law” publicly for the first time in May 2026, further fueling interest in 3D-stacking technology. At the national level, Phase III of China’s Big Fund has made its first investment in hybrid-bonding technology.