Source Code Capital Co-Leads Pre-A Round in 3D AI Chip Startup Suanmiao Tech, Which Raised Nearly 1 Billion Yuan Across Two Consecutive Rounds

Suanmiao Technology, an early-stage company long focused on R&D of 3D compute chips, recently closed two consecutive funding rounds totaling nearly RMB 1 billion. The proceeds will be used for R&D and mass production of domestic 3D compute chips. Source Code Capital was a co-lead investor in the company's Pre-A round.

Zhang Xun, Managing Director at Source Code Capital, said: "Against the backdrop of explosive growth in multi-agent systems and video models, compute consumption in inference scenarios is beginning to surge. Suanmiao Technology's high-bandwidth, low-cost solution enabled by 3D IC architecture is precisely the key to solving this core pain point. Source Code's AI industry layout has achieved systematic positioning across the entire industrial chain, spanning over 20 companies from semiconductor materials (ESWIN Material), equipment (Suzhou Xihang Semiconductor Technology Co., Ltd., Acceleration Era), and chips (Biren Technology, Gpixel, Suanmiao Technology, Zhihe Computing) to models and applications (Moonshot AI, Lovart, sand.ai, meshy), and on to AI+robotics scenarios (Unitree, Galaxy Universal, hillbot, Woan Robotics, Booster Robotics).

We believe that with its deep accumulation in the 3D IC field and profound understanding of AI algorithms, Suanmiao Technology has the capability to pioneer a new paradigm in AI chips, leading the industry from the 2.5D era into the 3D era. Source Code Capital will accompany innovative companies like Suanmiao Technology with a long-term mindset through cycles, jointly driving the transformation and breakthrough of China's AI compute infrastructure."

Recently, An Yong Waves interviewed Suanmiao Technology founder Wang Fuquan for an candid discussion of why he firmly chose the direction of 3D AI compute chips, the team composition, technical moats, and current achievements. The full interview follows:

---

An under-the-radar 3D AI chip company completed two funding rounds in four months.

Exclusively learned by An Yong Waves: Suanmiao Technology recently closed two consecutive rounds totaling nearly RMB 1 billion. The Pre-A round was co-led by Source Code Capital and Stone River Capital, with follow-on investment from Lenovo Capital and Incubator Group and other core semiconductor industry players. The Pre-A1 round was led by Xiang He Capital, with additional backing from state-backed capital including China Development Bank Capital and Beijing Shunxi. Proceeds will be used for R&D and mass production of 100% domestic 3D compute chips.

Suanmiao Technology is a company long focused on R&D of 3D compute chips, with core products being customized 3D chips for AI large model inference.

"The biggest problem with existing AI chips isn't computation but memory. When running AI inference, up to 70% of NVIDIA's H100 compute units sit idle, waiting for data to be moved from memory. Over the past 20 years, Moore's Law has driven a 60,000-fold increase in compute capability, while memory bandwidth has only grown 100-fold."

Suanmiao Technology founder Wang Fuquan told An Yong Waves that his goal is to solve the "memory wall" constraint on AI large model computation through computer architecture innovation and the 3D IC supply chain that his team has built over years of collaboration with domestic semiconductor core industry partners. Current 3D DRAM bandwidth can reach 32 TB/s, equivalent to 4x that of NVIDIA's B200. Suanmiao Technology's R&D focus is to convert that high bandwidth into tangible inference performance.

Suanmiao Technology provided An Yong Waves with A4 Paladin simulation data showing that on mainstream overseas open-source large models like Llama and Mixtral, A4's inference throughput (tokens/s) reaches 1.26x to 2.19x that of NVIDIA's H200.

Wang Fuquan, 51, was formerly a PhD and researcher at the State Key Laboratory of the Institute of Acoustics, Chinese Academy of Sciences, studying under Academician Zhang Renhe. After graduation, he entered the Institute of Computing Technology, Chinese Academy of Sciences for postdoctoral research in computer architecture, with Hu Weiwu — chief scientist of the renowned domestic CPU "Loongson" — as his collaborating advisor.

Currently, Suanmiao Technology's core scientists mostly graduated from the Institute of Computing Technology, Institute of Acoustics, Institute of Automation of the Chinese Academy of Sciences, and Tsinghua University. Among them are both veteran entrepreneurs who have spent years in the semiconductor industry and chief researchers who have been doing frontier AI exploration at Microsoft Research Asia — a team with strong immediate combat capability and high complementarity.

After 2019, full-industry-chain domesticization became the mainstream narrative in the chip industry. But the Suanmiao Technology team says they are not an opportunistic bet on the "domestic substitution" trend. Behind this startup is the story of a group of technology workers who were among the earliest to devote themselves to domesticization, navigating between state will and market economy.

Twenty years ago, after completing his PhD at the Chinese Academy of Sciences, Wang Fuquan naturally became a deep participant in the domestic CPU chip "Loongson." In 2009, he founded Sunsonic, and for nearly the following decade, did almost one thing — various industrialization exploration and development around Loongson. In that era that worshipped brands and globalization, sticking with Loongson was an extremely lonely journey; the lack of software ecosystem soil constrained the market competitiveness of domestic general-purpose processors.

In January 2018, Wang Fuquan was invited to a year-end party of an encryption compute chip design company and received "the most direct shock" — "A 10-person team, several hundred million in annual revenue, over 100 million in profit." This was because encryption algorithms demand extreme hardware design (ASIC) and face no software ecosystem barriers. This also provided a "fair competition" stage for domestic chip companies to challenge international chip giants.

At age 44, wanting to "find a new direction," Wang Fuquan quickly entered a "midlife rebellion." He shut down all Loongson-related business at Sunsonic, took RMB 10 million+ in angel funding and a "5.5-person" R&D team, and with the dream of becoming a world-class chip design company, plunged headfirst into the dark forest of encryption compute.

Entering the completely unfamiliar encryption compute field, Wang Fuquan's team didn't choose the most mainstream Bitcoin compute chip, but instead selected Ethereum — technically and commercially one of the most challenging options.

Just entering the encryption chip field, Wang Fuquan's team discovered that unlike Bitcoin, which had long entered the ASIC chip mining era, on the Ethereum blockchain, nearly everyone was still using NVIDIA and AMD graphics cards to mine. Unlike the Bitcoin consensus algorithm, which is very ASIC-friendly, Ethereum's consensus mechanism algorithm (ethash) fully exploits the so-called "memory wall" problem and was designed to be "ASIC-resistant" — a typical memory-hard algorithm that deliberately bottlenecks compute on memory access bandwidth. To achieve extreme compute release, there was only one path: beyond standard bus memory (like DDR, HBM), one had to find extreme memory bandwidth technology.

To crack this problem, Wang Fuquan's team rejected existing technologies to find new paths at three "last-minute" junctures, and finally locked onto the then-nascent "3D stacking" architecture by end of 2019. In Q4 2021, its high-throughput compute chip JASMINER X4 launched globally, using mature 40nm process to achieve 20x better encryption compute power efficiency than NVIDIA's 7nm flagship graphics card, eliminating NVIDIA and AMD graphics cards from Ethereum mining in one stroke. In the final year before Ethereum's transition to POS consensus, this single chip alone brought the team RMB 800 million in revenue, with JASMINER becoming the world's top brand in Ethereum mining.

Unexpectedly, ChatGPT emerged at the end of 2022. Wang Fuquan saw that behind AI large model computation lay a bottleneck extremely similar to Ethereum mining — compute being completely blocked by the "memory wall." Through the team's long-term focused work, 3D stacking had been proven the best practical solution for such memory-hard problems, and the explosive growth of AI large models provided an extremely grand compute scenario. Suanmiao Technology was thus born.

Today, with GPU capital stories already dazzling at home and abroad, Wang Fuquan believes that Suanmiao Technology's key to survival and development is "the new opportunity brought by the leap in compute paradigm in the AI large model era."

"At Suanmiao we rarely mention concepts like 'domesticization' or 'domestic substitution,' because what we do is already the best globally. Our goal remains to become an internationally competitive chip company, contributing new solutions with Chinese advantages to global AI large model computation, alleviating the global compute crisis and compute energy crisis," Wang Fuquan said. His confidence comes from the team's years of accumulated R&D experience in the 3D IC field, and successful large-scale commercialization practice in the encryption compute market.

Suanmiao Technology's two funding rounds have brought together national industrial capital, 3D IC core supply chain industrial capital, and top-tier market-oriented funds. At least at the capital level, this ambitious dream now has its most basic support.

Also worth noting: Source Code Capital, which completed fundraising in the second half of 2025, was the earliest to lead investment in Suanmiao Technology. Looking across Source Code's portfolio, this long-low-key investment firm has quietly completed systematic positioning across the entire AI industrial chain. From semiconductor materials to equipment to chips — ESWIN Material, Biren Technology, Seeya Technology, Gpixel, Suzhou Xihang Semiconductor Technology Co., Ltd., Acceleration Era — to AI+robotics scenarios like Unitree, Galaxy Universal, Hillbot, Booster Robotics, Woan Robotics, and on to models and applications like Moonshot AI, Lovart, sand.ai, meshy, and AI for science with DeepWise — over 20 companies in total. Clearly, Source Code wants to miss none of this AI era rushing toward us.

In the winter of 2025, at Suanmiao Technology's headquarters in Beijing's Zhongguancun, An Yong Waves met Wang Fuquan. This middle-aged entrepreneur who emerged from the Chinese Academy of Sciences laid out a grand vision and strong confidence facing the current market. From his account, you can see how a scientist-entrepreneur with "national team" background embraced the market, believed in the market, and then with a new mindset returned to the path of "all-domestic."

The interview follows:

01

The "Top-Tier" Chip Architecture That NVIDIA Won't Build

An Yong: Your simulation data shows A4 using 12nm process "defeating" NVIDIA's H200 which uses TSMC 4nm process on inference performance. Is this reasonable?

Wang Fuquan: Counterintuitive, but it conforms to the physical nature of large model computation. Large model inference is a typical "memory-bound" task. Simply put, the bottleneck isn't that the brain doesn't spin fast enough (insufficient compute cores), but how quickly data can be brought into the brain (memory bandwidth).

Imagine a compute chip as a factory, with data in memory chips as raw materials. NVIDIA's H200 factory has very complete equipment (4nm process) and can produce all kinds of products (general processor), but the problem now is how to quickly transport raw materials into the factory. The 2.5D architecture chip (CoWoS) approach is to build wider conveyor belts, ultimately constrained by the factory gate width (shoreline). NVIDIA's H200 chip size is already at the limit of mass-producible dimensions, with memory bandwidth reaching 4.8 TB/s.

The 3D architecture chip is another solution approach. We stack memory chips directly on top of the compute cores (shortening transport distance) and build hundreds of thousands of vertical elevators, so raw materials can be quickly transported to all corners of the factory, no longer limited by gate width, achieving 16-32 TB/s bandwidth. Meanwhile, the factory we design (compute chip) is specifically for inferencing AI large models (dedicated chip), so we can achieve higher inference performance with less equipment (12nm process).

An Yong: If 3D architecture is so powerful, why doesn't NVIDIA do it itself?

Wang Fuquan: NVIDIA is a great company. Its moat is built on the CUDA ecosystem and general GPU architecture, but this also means its hardware architecture innovation must be subordinate to software ecosystem compatibility. Its hardware architecture needs to accommodate graphics rendering, scientific computing, AI training, and various other scenarios — it must be an "all-around champion." Meanwhile, 3D stacking architecture innovation brings entirely new challenges to hardware. Suanmiao chose a customized ASIC approach, sacrificing the generality not essential for large model computation in exchange for extreme inference performance. If NVIDIA did this, it would be dismantling its own GPU empire. This is precisely the startup's opportunity — not burdened by the giants' past baggage, we can operate based on first principles.

An Yong: That's also my next question — why only do large model inference and not training? Are you avoiding the frontal battlefield?

Wang Fuquan: This isn't avoidance, but strategic focus.

From a technical perspective, large model training requires not just chip design capability but also engineering capability for interconnecting tens of thousands of cards and complex software stack ecosystems. So the training market isn't an appropriate entry point for a startup.

From a market perspective, 90% of future AI compute demand will occur on the inference side — large model inference compute demand will far exceed training compute demand. Large model training will eventually converge; in the future, everyone's phones and every company's servers will mainly be doing large model inference. At that point, inference cost (TCO) will be the only consideration. Customers won't care whether you're a GPU or not, only how much money and electricity it takes to generate one million tokens. This is precisely the battlefield where ASIC excels.

Moreover, I believe this is an opportunity for Chinese people.

An Yong: Is this nationalist sentiment?

Wang Fuquan: This isn't sentiment, but my observation over the years of differences in engineering thinking between China and the United States.

American engineers' strength is "abstract thinking" and software. Look at CUDA, Windows, iOS — they excel at abstracting the complex world into layers of standard interfaces, building ecosystems. This is what American engineers are capable of.

But ASIC dedicated chips are different — they are the ultimate in "concrete thinking." They require you to arrange transistors with Swiss watch-like precision in an extremely small physical space, repeatedly polishing to save a tiny bit of power or squeeze out a bit more performance, even "hacking" memory dies. This kind of meticulous refinement "making a temple in a snail shell" is precisely what Chinese engineers excel at.

The history of encryption compute chips has already proven this: although Europe and America invented Bitcoin and Ethereum first, the companies that ultimately dominated Bitcoin encryption compute chips were Bitmain, and the king of Ethereum compute chips was Sunsonic — both are homegrown Chinese companies. In AI inference, which similarly demands extreme efficiency, I believe history will repeat.

An Yong: Will customers accept this chip layout model with separate training and inference?

Wang Fuquan: Training and inference are two different scenarios/customer groups. Only a very few top players can sustain large model training; the inference customer base is much larger. Our first product focuses on large model inference, allowing our chip architecture to be extremely streamlined — only needing to accommodate limited generality within inference scenarios, while dedicating more resources to enhancing inference computation itself. This brings lower cost and lower power consumption. Suanmiao doesn't want to replace all of NVIDIA, but to carve out the biggest piece of its future incremental cake.**

An Yong: How big is this cake?

Wang Fuquan: Just large model inference alone — globally, this is already a compute market of hundreds of billions of dollars. In China alone, it's already a compute market of hundreds of billions of RMB, and both are growing rapidly, beyond most people's expectations.**

An Yong: Domestically there are GPU "N Little Dragons," plus big tech companies eyeing this aggressively. With so many competitors, you're still quite confident.

Wang Fuquan: We don't do GPUs, that's not our strength either. For six or seven years, our team — driven by the market, together with our supply chain partners — pioneered China's 3D stacking chip field and has already initially formed significant advantages globally. We focus on R&D, mass production, and global sales of 3D chips. As a startup, we've invested over RMB 1 billion in this field, concentrating our earned money and raised money into this area, building a complete team with rich experience and combat effectiveness in 3D chips and large model compute software and hardware. Of course we have reason to remain confident. At the same time, we must always maintain a treading-on-thin-ice mentality, carefully and steadily walking the broad road we've pioneered. We believe 3D chips represent the future of computing, and believe ASIC is the correct path for solving large model inference compute. GPU is just a transition. But in any case, we must thank GPU companies represented by NVIDIA for birthing this great AI large model era.**

02

What Is the Correct Answer for Domesticization?

An Yong: That 2018 year-end party completely changed you. Why was its impact so great?

Wang Fuquan: It was my friend's company party. They also did chips, but encryption mining chips that we "regular troops" looked down on at the time. That night, what shocked me tremendously wasn't just those numbers — 10 people, RMB 300 million revenue, RMB 100 million net profit — but the brutality and directness of that business model.**

They didn't need to "beg" customers, didn't need to write tedious application materials. As long as the chip's compute power was strong enough, customers would line up with cash in hand. That was the first time I intuitively felt "compute is currency." This market force was too raw and powerful. Based on our experience and understanding, I firmly believe that China's chip industry's future lies in market forces. China's technological market forces are clearly severely underestimated.

An Yong: Was it that easy to have a "midlife rebellion"?

Wang Fuquan: I developed a profound self-doubt at the time: if technology can't be converted into real money in the market, where does its value lie? If I wanted to build a great chip design company, it had to withstand market testing.**

So I decided to "go into business," letting go of all Loongson-related business. Many old friends didn't understand at the time, thinking I had "degraded" from building national heavy equipment to making encryption mining machine chips. But looking back now, without those years of extreme survival training in the global market, without fighting in that arena that was purely about PPA (performance, power, area), we couldn't possibly have developed the team we have today, let alone developed together with domestic 3D IC core supply chain partners.

An Yong: Why did you choose Ethereum at the time?

Wang Fuquan: Our team's instinct is to find big jobs to do. Ethereum was the second-largest cryptocurrency network at the time. More importantly, Ethereum mining chips were the most challenging chip in that field. Because the entire chip's algorithm bottleneck was stuck at "memory bandwidth." To double the chip's compute power, you had to double the effective bandwidth. No shortcuts possible. This "memory wall" problem has been a traditional problem plaguing the entire computing field.**

An Yong: How did you crack this hard problem?

Wang Fuquan: The team had just five and a half people at the time; one member was still part-time. We determined to take the "pure ASIC" path. The first thing we tried was the HBM solution — at the time, the highest bandwidth memory solution based on DRAM storage medium. But as research deepened, due to supply chain and cost-performance considerations, we quickly abandoned it. We were a team that saw through HBM memory very early. The second solution was to not use external memory at all, but use the limited SRAM inside the chip, trading computation for space. Despite the small space, use repeated computation to make up for it. This solution took 18 months; the chip was even designed. At that juncture, we found a third solution.**

An Yong: 18 months of work, and you just put it down?

Wang Fuquan: Abandoned it. Because we found something better. The third solution was strong interconnect, using SerDes to complete interconnection of multiple chips to solve the low SRAM capacity problem. Power consumption was good, but the challenge was the board became extremely complex. We worked another 3 months, until end of 2019, when we encountered what we believed was the ultimate solution: the 3D architecture solution.**

An Yong: What was the process of pivoting to AI chips? What opportunity did you see?

Wang Fuquan: When I first encountered AI large models in early 2022, ChatGPT hadn't yet broken through, and I was quite hesitant. On one hand, before this, AI had always been much praised but little purchased — no chip company had really made money on AI chips. On the other hand, in the previous small model era, various model architectures emerged endlessly, forcing the design of general processors to accommodate various models, which went against our team's DNA.**

In the first half of 2023, I was in Silicon Valley, personally experiencing this wave of AI large models. Scaling law-driven improvement in AI large model intelligence far exceeded my expectations; the Turing test was in fact broken — this was the real big deal in the AI field. Meanwhile, our team's research on the Transformer algorithm had reached considerable depth, and our accumulated 3D architecture technology happened to be the most promising solution for solving the memory bottleneck in large model computation. So we quickly pivoted to AI large model chips.

AI compute power in the future will be like today's water, electricity, and gas — infrastructure for the new era. The core competitiveness of the AI era lies in compute power, and the future of compute power lies in architecture innovation. We firmly believe that 3D stacking architecture and ASIC extreme optimization design philosophy are the optimal solutions for AI large model compute power in the next 5-10 years.

An Yong: I know you're very actively recruiting now. How do you persuade top engineers already at big chip companies to join you?

Wang Fuquan: Simple — first-rate talent should do first-rate work, and what we're doing is the top-tier work globally in the large model compute field. And our team has already undergone long-term, thorough commercialization training in global markets, with deep understanding of the compute business.**

Additionally, we don't stint on compensation — whether equity incentives or salary, it's definitely top-tier.

An Yong: Starting from Loongson's "national team," you went to the encryption compute circle to be a "grassroots hero." Now you're back on the "national heavy equipment" main track of AI compute chips. Looking back, how do you define yourself now?

Wang Fuquan: I think I've become a "realistic idealist."**