Founder of Delphi Labs: My observations and feelings about the AI ecosystem in China in two weeks

Original: Delphi Labs Cofounder, ++@ZeMariaMacedo++

Compiled by: Big Pliers | PANews Lobster

I flew to China with high expectations—thinking I would see severely underestimated geniuses changing the world at a fraction of the price. Two weeks later, I returned with a more complex answer:

Chinese hardware is quietly winning a war that Westerners are not even aware has begun; however, in the software sector, valuation bubbles and homogenized founder profiles are turning what should be an explosive ecosystem into another game of hot potato.

I spent two weeks in China, visiting founders, venture capitalists, and CEOs of publicly listed companies in the AI ecosystem. Before entering, I held an optimistic view of this ecosystem, hoping to find world-class AI talent making great strides at valuations far lower than those in the West.

When I left, my perspective became more nuanced: I am more optimistic about hardware than expected, but more pessimistic about software, and I formed some views on Chinese founders that surprised even me.

Founder Issues

The outstanding founders I have invested in share a common trait: independent thinking, a rebellious spirit, intense focus, and near-paranoia. They do not follow the crowd, constantly ask "why," and refuse to accept second-hand wisdom. Their decisions may seem inexplicable to outsiders, but to them, they are self-evident. They possess a strong drive that comes from deep within, often manifested in a life filled with obsession and excellence. Their life trajectories have a sense of "sharpness," easily recognizable among the high-IQ individuals I encounter as a VC.

Many of the Chinese founders I met, however, belong to another type—this surprised me.

They are exceptionally talented—coming from top universities, having worked at ByteDance or DJI, published papers in Nature, and holding multiple patents. These achievements, which only top technical talents in the West can attain, are merely entry tickets here. Their level of diligence almost surpasses anyone I have ever met. We met at any time, on weekends, and across cities. One founder even came to meet us on the day his wife gave birth!

But independent thinking, rebellious spirit, and a vision from zero to one—these are hard to find. The backgrounds of different founders are highly similar, their business plans are more conservative and stable, and their ideas are often refined versions of existing things rather than truly original bets. Given that China has cultivated such a large pool of technical talent, I expected to encounter more people proposing ideas I had never heard of.

My judgment is: China's education system can cultivate excellence, but it does not leave enough room for "deviation." It produces founders who can execute known problems to the extreme, rather than those who come with a "problem that no one has realized is a problem."

Venture Capital Reinforcing This Model

Interestingly, local investors are actively amplifying this trend.

Most Chinese funds base their entire investment logic on investing in the best alumni from ByteDance or DJI—favoring prestigious backgrounds over individuality, prioritizing credentials over judgment. The composition of VCs is also similar: most come from big companies, consulting, or investment banking backgrounds, mirroring European VCs from a decade ago.

Ironically, the best Chinese founders in history—those who truly established generational enterprises—never worked in big companies. Jack Ma was an English teacher who failed the college entrance exam twice; Ren Zhengfei founded Huawei after retiring from the military at 43; Liu Qiangdong started from a street stall and founded JD; Wang Xing dropped out of a PhD program and started from scratch. The recent Liang Wenfeng had never worked anywhere before founding DeepSeek; he only stayed at his own company. These individuals are outliers, lacking impressive resumes—precisely the type of people that the current system tends to overlook.

Finding these individuals is where true excess returns lie, and in my view, almost no one is looking in that direction right now.

Shenzhen and the Hardware Ecosystem

In China, what amazed me the most was not any startup pitch.

It was Shenzhen's "hardware underground world"—where engineers systematically purchase high-end Western products, disassemble them piece by piece, and conduct reverse engineering with meticulous precision. When I left, I truly wasn't sure if most Western hardware founders really understood what they were competing against. The network effects here are not just theoretical; they are real, inseparable, and built over decades.

The entrepreneurs we met validated this: over 70% of hardware raw materials come from the Greater Bay Area, nearly 100% from within China—this allows for iteration cycles far beyond the capabilities of Western hardware companies.

Most of the founders I met are following DJI's playbook: doing consumer-grade hardware in a niche area—electric wheelchairs, robotic lawn mowers, next-generation fitness equipment—achieving revenues in the 8 or 9 figures, and then expanding into adjacent categories using their customer base or underlying technology. Some of these companies are already far larger than you can imagine. The most impressive company I encountered was Bambu, a 3D printing company that most Westerners have never heard of, reportedly achieving annual profits of $500 million and doubling each year.

Pessimistic About Chinese Software

When I left, I was even more skeptical about opportunities in the Chinese software sector than when I arrived.

At the model level, China's open-source achievements are indeed impressive—but closed-source models still show significant gaps compared to the strongest models in the West, and these gaps are likely to continue widening. The gap in capital expenditure is enormous, GPU acquisition remains constrained, and Western labs are intensifying their crackdown on model distillation. Revenue data already illustrates the issue: reports indicate that Anthropic achieved $6 billion in just one month in February. The best models in China still have ARR in the tens of millions of dollars.

In terms of software startups, the mainstream profile consists of former product managers and researchers from ByteDance, developing intelligent agents or environmental perception consumer software aimed at Western markets.

The talent is real, but most of these products fall within the functional range that large labs would natively integrate—potentially rendered obsolete by a single product release. I am also shocked by the overall lack of large-scale, high-growth private software companies in China.

In the West, aside from model companies, several startups have already achieved 9-figure or even 10-figure ARR with astonishing growth—Cursor, Loveable, ElevenLabs, Harvey, Glean. This level of breakthrough private software companies basically does not exist in China—while a few exceptions, such as HeyGen, Manus, and GenSpark, once they find breakthroughs, ultimately choose to leave.

Valuation Bubble

Despite the poor situation in the software sector, the bubble is very real—both in early and late stages.

In the early stage, although the costs of top talents from ByteDance, DeepSeek, and Moonlight are still significantly lower than their American counterparts, median valuations have converged. It has become commonplace for consumer startups pre-product to be valued at $100 million to $200 million, and seed rounds exceeding $30 million are not unusual.

In the late stage, the numbers become even harder to justify. MiniMax's valuation in the public market is about $40 billion, while its ARR is less than $100 million—about 400 times its sales. Zhizhu AI is valued at about $25 billion, with revenues around $50 million. In contrast, OpenAI's peak valuation was about 66 times its ARR, and Anthropic's was about 61 times.

Private model companies like Moonshot are using these publicly listed companies as benchmarks, raising funds at valuations of $6 billion, $10 billion, and $18 billion within just a few months. People in the crypto circle should be familiar with this dynamic—investors are comparing private valuations with pre-unlock public stock prices.

Moreover, part of the reason supporting Zhizhu and MiniMax in maintaining their current valuations is that they are currently the only way to gain exposure to the Chinese AI narrative, which carries a premium in itself. However, as more companies go public and this scarcity is diluted, the situation will change. Finally, the IPO window has a tendency to close quickly and without warning—before you complete your arbitrage, the benchmarks you reference may have already changed, and there is no guarantee of that.

The humanoid robot sector is facing a similar situation. China has about 200 humanoid robot companies, around 20 of which have raised over $100 million, with several valued at billions—almost all are in the pre-revenue stage, most planning to list on the Hong Kong Stock Exchange in 2026 or 2027. If this market is real, China's hardware dominance makes the long-term trajectory clear. However, commercial realization is likely to be much slower than the current fundraising pace suggests, and I am deeply skeptical about whether the Hong Kong market can accommodate the many billion-dollar humanoid robot companies currently in the queue. I choose to wait and see for now.

Asymmetry Worth Noting

One thing I did not anticipate: almost every founder I met is prioritizing global markets over the Chinese market. They use Claude Code, pay attention to Dwarkesh, and are well-versed in the entrepreneurial ecosystem of San Francisco, often knowing more than Western investors who have not been paying close attention.

The hostility of the West towards China far exceeds China's hostility towards the West. Chinese founders do not feel any contradiction in combining China's engineering execution and hardware depth with the West's market development capabilities and product vision. When this combination materializes in the right founding teams, it will give rise to some truly remarkable companies.

Finding those founders—those who do not fit the "credential template" optimized by the local VC ecosystem—is precisely where we focus.

Special thanks to ++@woutergort++ for opening up his amazing network in China to us, thanks to ++@PonderingDurian++ for organizing this trip, and thanks to Claude for patiently helping me organize my fragmented thoughts on the plane.

Note: This article was primarily compiled by AI.