Yunqi Capital Leads Index Technology's Nearly RMB 100 Million Pre-A Round: Hardware Design Enters Its "Vibe Coding" Moment | Yunqi Capital

While AI coding is reshaping software engineering, hardware R&D remains stuck in a "classical era" — still heavily dependent on experience and repetitive manual labor.

Could there be a way for hardware design to make a similar leap in efficiency, something like "Vibe Coding"? What if AI could generate design blueprints directly, turning the work from heavy lifting into genuine decision-making and creativity? Yunqi Capital's latest investment, Zhishu Technology, is working on exactly that.

Recently, Zhishu Technology announced the completion of a nearly RMB 100 million Pre-A funding round led by Yunqi Capital. The company's founder, Qunsong Ye, also sat down with An Yong Waves, a deep-dive tech and VC media outlet, to explain why hardware needs its own "technological democratization" — and what happens to the engineering world when AI compresses hardware design timelines from weeks down to one or two days. This edition of Yunqi Partners brings you their conversation.

➤ ➤ Yunqi Capital's Investment Thesis

"Zhishu Technology is establishing a new paradigm for electronic design using AI. Unlike traditional engineering tools, what they're building is an intelligent system that 'understands intent and produces results' — with human efficiency improved by over 20x compared to conventional methods. This ground-up approach not only redefines the entry point for industrial electronics design using AI, but also demonstrates significant potential for industrial migration and global market expansion."

The following content is adapted from An Yong Waves.

While software engineers have grown accustomed to pair programming with Cursor, the world of hardware engineering still feels distinctly classical —

If you want to build an app, AI can probably generate a runnable code framework for you in minutes. But if you want to design a simple piece of smart hardware — say, an eye-care lamp or an electronic toy — you still have to endure a lengthy process: component selection, schematic drawing, PCB (Printed Circuit Board) design, prototyping, soldering, debugging, and more.

"A consumer electronics product of moderate complexity typically requires a circuit board with up to 300 components. For an experienced engineer, going from scratch to completion usually takes 20 to 30 days," Qunsong Ye, founder of Zhishu Technology, told An Yong Waves. What he's working on, he says, "can compress that timeline to one or two days."

An Yong Waves has learned exclusively that AI-powered hardware design generation platform Zhishu Technology recently completed a nearly RMB 100 million Pre-A funding round, led by Yunqi Capital, with Yuzun Capital and Shangshi Capital participating. Earlier, in 2024, shortly after its founding and with only a demo in hand, Zhishu Technology had already secured an angel round jointly led by QF Capital and Huagai Capital.

What Zhishu Technology aims to do can be summed up in one sentence: use AI technology to generate electronic circuit design blueprints and accompanying embedded code, achieving "unattended" hardware R&D.

Founder Qunsong Ye graduated from the University of Electronic Science and Technology of China in 2008. Early in his career, he worked at Rigol Technologies on instrumentation; he later joined Honeywell, where he led the innovation lab — an "innovation tree" program seeking out novel electronics innovations from across the country.

In 2019, Ye chose to start his own company, founding Miji Technology, focused on AI-powered educational hardware and "AI companion reading" products. It was once a glamorous track — the team secured partnerships with leading clients like Zuoyebang and New Oriental. But just two years later, the "double reduction" policy on private tutoring landed.

"The education clients we'd ranked in our top tier basically all collapsed. What survived were the ones we'd considered third-tier — companies doing traditional hardware upgrades." Correspondingly, the demands Miji Technology faced became extremely fragmented and technically trivial — a flood of repetitive tasks like "make the light turn on" and "add a button interaction" poured onto the "ambitious" R&D engineers in Ye's team.

"People from R&D backgrounds mostly don't defer to each other easily — nobody wants to do boring work." This extreme hunger for efficiency, combined with the pressure of being forced to find a way out amid "low-end repetition," drove Ye to begin exploring automation internally. He divided employees' time into three parts: one portion for survival business, one for repetitive labor, and whatever remained of the day had to yield two hours for "crazy experiments" in automated generation.

It was this survival crisis that unexpectedly incubated the earliest form of Zhishu Technology in 2023. Ye chose a path of "changing lanes to overtake" — rather than trying to rebuild an EDA (Electronic Design Automation) system within existing rules, he would use AI large models and a black-box approach to generate results directly.

Simply put, what he wants to build is an AI company that lets ordinary people "hard control" Huaqiangbei — Shenzhen's legendary electronics market.

In Ye's ultimate vision, everyone in the future will have something like J.A.R.V.I.S., the AI butler from Iron Man, able to translate their needs and ideas into real hardware products with ease. Ye believes that when AI drives the barrier to hardware design close to zero, human creativity will no longer be shackled by tedious engineering details.

Recently, An Yong Waves sat down with Ye —

Building the "New Energy Vehicle" of Hardware Design

An Yong: EDA (Electronic Design Automation) has existed and evolved for decades. How is what Zhishu Technology does different?

Ye: The current EDA "Big Three" — Synopsys, Cadence, and Siemens — control roughly 80% of global market share. They've constructed an enormously complex rule system. If you try to build a "better automation tool" within their rules, that's a dead end. Because they wrote the rules. If automation could be solved within that framework, they would have done it decades ago.

Yet engineers' pain points remain unaddressed. Existing EDA tools are essentially still "drawing" software — engineers still spend 80% of their time on low-value repetitive tasks like "make the light turn on" and "connect this wire," work with zero intellectual challenge.

Competing along traditional EDA lines offers absolutely no chance. So we chose to "change lanes." After founding Zhishu Technology, we proposed AI For Hardware (AFH) — not relying on traditional EDA rules, but using AI large models and a black-box approach to generate results directly.

An Yong: So you want to use AI to disrupt the traditional EDA giants?

Ye: I wouldn't say disrupt. Let me use the auto industry as an analogy: if China's auto industry tried to catch up with Germany and Japan on internal combustion engines, that would be extremely difficult. But "new energy vehicles" offer a chance to overtake on a curve — a completely different track, not competing on engines but on batteries and intelligence. You could say we're building the "new energy vehicle" of electronic design.

An Yong: Specifically, what is Zhishu Technology's current product? What level can it achieve?

Ye: Our core product is an AI-based automated design platform that can generate circuit board design blueprints and accompanying embedded code directly from requirements. Currently, we mainly focus on boards with four layers or fewer, which covers the vast majority of consumer electronics products.

To give a concrete example: a board with under 300 components. For a human engineer, going from receiving requirements to component selection, schematic drawing, and PCB layout takes roughly 20 to 30 days. Using our platform, it now takes just one to two days.

More critically, hardware R&D has one massive pain point: EMC testing (note: a comprehensive evaluation of electronic products' electromagnetic interference and immunity, one of the most important quality indicators for hardware products). This is every electronics designer's nightmare. A human-designed board might function fine, but fail radiation testing in the lab — then it's back to revising layout, rerouting traces, another half month gone. Typically a product goes through three to five iterations to pass. But our product, trained on massive amounts of historical successful data, generates boards with very high first-pass rates on radiation tests, usually requiring fewer than two iterations.

An Yong: This sounds quite similar to Vibe Coding, which has been very hot in software design lately.

Ye: Exactly. Vibe Coding dramatically accelerates software development by not memorizing rules, just vibe-based conversational programming. I think hardware should have its own Vibe Design or Vibe PCB concept. Right now, engineers' R&D workflows are very rhythmic and bounded, but that also limits the speed at which innovation can emerge. We want to reconstruct the entire electronics design R&D process, letting engineers become true decision-makers without spending excessive time on tedious procedures.

An Yong: Many vertical AI platforms face the question of whether they'll still exist after general-purpose large models iterate. Do you think what you're building has high enough barriers?

Ye: Current general-purpose large language models are fundamentally probability-based "replication" and "assembly" — they don't truly understand the intent behind circuit design or the rules of physics.

At the intersection of AI and electronics, two enormous barriers already exist:

First, data is extremely fragmented and dirty. After decades of rapid development, the electronics industry has matured considerably, but there aren't many open-source, high-quality schematics and PCB source files available online. Some leading manufacturers in electronics have accumulated data for single or multiple product categories over many years, but because AI and electronics technology developed relatively independently for so long, this data isn't suited to AI's "appetite." The foundational corpus we're building for the electronics industry is specifically designed to "feed" AI — requiring massive cleaning and synthesis work.

Second, there's a lack of evaluation baselines. GitHub has oceans of code, but very little embedded code specifically tied to hardware. For pure software development — say, AI writing webpage code — you run it and know if it's right or wrong. But embedded code errors get burned into chips; you can't evaluate them directly in the cloud. We're building evaluation benchmarks for embedded code, which is a blank space in the industry. Our current architecture combines reinforcement learning with MoE, integrated with knowledge graphs. This isn't just about drawing diagrams — it's about understanding the constraints of the physical world.

An Yong: If AI dramatically improving efficiency becomes reality, would hardware giants like DJI or Bambu Lab do this themselves?

Ye: Companies like DJI and Bambu Lab, with their extremely strong product definition capabilities, are actually our best potential customers. They use traditional EDA tools to design boards so they can quickly get products to market. They won't build their own EDA just because the tools are frustrating — that makes no business sense; it's a completely different track.

I believe that once our product has enough validation, they'll prefer to partner with us rather than build their own tool.

An Yong: Zhishu Technology isn't your first startup. Did your previous experience influence this one?

Ye: My previous company was called "Miji Technology," doing AI companion reading hardware for education. In 2021, things were looking up — we'd secured commitments from major clients like Zuoyebang and New Oriental. But on July 24, 2021, when the "double reduction" policy hit, these top-tier clients basically collapsed instantly, business completely retracted. What rose instead were third-tier clients — traditional hardware manufacturers looking to transform.

To survive, we started taking their orders. But the problem was, most of this work was extremely tedious, technically trivial "dirty work." The team grew frustrated. Everyone came from R&D backgrounds, from major corporate labs — they wanted to build cool things, not spend their days drawing simple boards and tweaking basic features. Everyone was pent up.

At the time, I told the team: since everyone thinks this is simple, why can't we solve these uninteresting tasks through automation? So we made a rule: divide each day into three parts — one for core business survival, one for boring grunt work, and whatever time remained had to yield two hours dedicated to "banging heads together" on automated generation.

What Zhishu Technology is doing now was essentially "forced out" by this pressure.

Achieving Per-Capita Revenue of Hundreds of Millions of Dollars

An Yong: Who are your actual customers right now?

Ye: Primarily B2B clients. Because our core advantage is speed. Many clients come to us when they're facing tight product launch deadlines or when their existing solutions have problems. Even if we quote above market average, they're willing to pay because we solve urgent needs. For example, we have an eye-care lamp client who initially gave us just one or two products to try — now 80% of their products are produced through us. Their R&D team basically only needs to handle acceptance. To date, we have about forty to fifty enterprise users, with annual orders exceeding RMB 80 million.

An Yong: From a capital markets perspective, B2B in China isn't considered particularly "sexy." You're either stuck on willingness to pay or trapped in customization services. How do you solve this?

Ye: B2B is primarily for accumulating data — B2B clients provide large volumes of real data. Next, in 2026, we plan to extend our reach to large numbers of engineers, letting them dramatically improve efficiency and focus more energy on innovation.

We plan to enter through maker communities — places like Elecfans, JLCPCB's community, and overseas open-source hardware communities. These people love to tinker and have the highest tolerance for new tools. Get them using it to generate simple modules: filters, small appliance control boards. Through these makers, we'll gradually penetrate to professional engineers and system architects. When an engineer discovers our tool can compress two weeks of work into two hours, they'll naturally bring it into their company.

Ultimately, we want this platform to be B2C-capable. Once the model has been trained on data from enterprise and professional clients, error rates will drop further — then ordinary people will be able to generate hardware products through conversation.

An Yong: Will that many ordinary people really want to design hardware themselves?

Ye: I don't think everyone wants to be an inventor — that's too hard. But when I ran the innovation tree at Honeywell in 2016, I received innovative ideas from all over the country every day.

Put simply, everyone has an idea — "I want to make..." That idea ultimately comes down to clothing, food, housing, transportation, or software and hardware products. Software is already convenient; hardware is slow because the barrier is too high. Many innovative ideas are subconsciously killed. I believe that if the barrier is low enough, creative demand will emerge.

An Yong: A kind of technological democratization?

Ye: I believe so.

An Yong: Hardware is very different from software — software just needs to run on a computer or phone, but hardware has to be a physical object.

Ye: We're seeing companies like Bambu Lab mature in 3D printing, even printed circuit board technology. And on our platform, users don't need to worry about these things. Through our downstream supply chain layout, what we can definitely achieve is: user inputs requirements, blueprint generates automatically, user refines it, enters shipping address, and then just waits for the delivery.

An Yong: Whether for professional users like engineers or B2C ordinary people, the blueprints generated could be wildly varied. Can today's manufacturing actually produce so many non-standard products?

Ye: Decades of smart manufacturing development have largely solved this. Rapid production according to design blueprints is basic competence for many manufacturing enterprises.

Take JLCPCB and HQPCB, with whom we have deep partnerships — they're already among the top players in "produce-to-blueprint." Thanks to their AI panelization technology, they can complete production of thousands or even tens of thousands of blueprints from different customers within 24 hours. So at this stage, manufacturing is no longer the key bottleneck limiting electronics product creation speed (though it once was).

Only when technologies or platforms like AFH solve R&D efficiency will electronics product creation speed potentially become constrained by manufacturing — but I believe by then there will certainly be faster ways to complete manufacturing. We've also seen overseas companies researching how to complete PCB bare board manufacturing, PCBA soldering, and testing from entirely new blueprints within two hours.

An Yong: What's your current team size? If you move toward B2C in the future, won't you become a labor-intensive company?

Ye: We're currently under 50 people. I've set a strategic rule for the company: headcount never exceeds 100. I truly believe that with AI augmentation, a company with per-capita revenue of hundreds of millions of dollars is possible. I still write code myself — tools I can build in a day might have taken months in the past. The company's human resources are mainly invested in R&D and tool optimization, not linearly scaling headcount with order volume.

An Yong: Layering manual optimization on top of AI-generated blueprints is one pricing strategy. Wouldn't more users increase human resource demands?

Ye: That actually depends on business model. If we only did B2B, we'd indeed need more people for delivery and communication — but that's exactly the part we'll strategically "retreat" from. B2B business currently serves for cash flow and data, but we won't infinitely stack people just to grow revenue.

For B2C or B2Professional business, human resource needs won't scale linearly with user volume. Because reviewing blueprints requires understanding the other party's design logic from scratch, which is extremely time-consuming. As model capabilities improve, the proportion requiring human intervention will keep decreasing.