Delphi Labs Founder: After Two Weeks Deep in China's AI Scene, Shenzhen's Hardware Amazed Me, Software Valuations Terrified Me

Observations of Chinese Founders Challenge My Previous Understanding

Author: José Maria Macedo, Co-founder of Delphi Labs

Translation: Deep潮 TechFlow

Deep潮 Guide: Delphi Labs founder spent two weeks intensively visiting China’s AI ecosystem, meeting numerous founders, investors, and publicly listed company CEOs.

His conclusion was surprising: more optimistic about hardware than expected, more pessimistic about software than expected, and his observations of Chinese founders challenged his previous perceptions.

The article also covers hot topics such as valuation bubbles, humanoid robot tracks, and information asymmetry between China and the West.

Full Text:

I spent two weeks in China, meeting a large number of founders, VCs, and CEOs of listed companies in the AI ecosystem. Before going, I was bullish on this ecosystem, expecting to see world-class AI talent working at valuations far below Western standards.

By the end, my view changed—becoming more specific: hardware is stronger than I imagined, software is weaker, and some observations about Chinese founders surprised me.

The Founders’ Issue

The excellent founders I’ve invested in share common traits: independent thinking, rebelliousness, extreme focus, and obsession. They don’t follow instructions. They keep asking “why,” refusing to accept secondhand wisdom. Their decisions may seem inexplicable to outsiders, but they see them as natural. They possess an internal, unstoppable intensity, often expressed as long-term obsession and excellence. As a VC meeting many smart people daily, I can spot these individuals easily because their life trajectories have a distinct “sharpness.”

Many founders I met in China are of a different type, which surprised me.

They are extremely talented—top universities, backgrounds at ByteDance or DJI, published papers in Nature, multiple patents. In the West, these achievements are only held by top-tier technical talent; in China, they are entry tickets. They are also more driven than almost anyone I’ve seen. We held meetings at various times, even on weekends, rushing between cities. One founder even came to see us on the day his wife was giving birth.

But independent thinking, rebellious spirit, and a vision from 0 to 1—these are harder to find. The backgrounds of founders are highly similar, their pitches more conservative, many ideas are upgrades of existing products (impressive V2), rather than truly original bets. Given China’s large pool of technical talent, I expected to meet more people with “ideas I’ve never heard of.”

My interpretation is: China’s education system cultivates excellence but doesn’t leave enough space for deviation. The output is top-tier executors skilled at solving known problems, not those who come with a “problem no one knows exists.”

VCs Reinforcing This Pattern

Even more interestingly, local investors are intensifying this trend.

Most Chinese funds’ investment logic is based on one premise: investing in the best people emerging from companies like ByteDance or DJI. They look at resumes, not brilliance; backgrounds, not beliefs. The VC profile is similar—former employees of big companies, consulting firms, or investment banks, much like European VCs a decade ago.

Ironically, many of China’s truly great company founders never worked at large corporations. Jack Ma was an English teacher, admitted twice to university. Ren Zhengfei founded Huawei at 43, after serving in the military. Liu Qiangdong started selling goods at markets. Wang Xing dropped out of college to start a business. Recently, Liang Wenfeng, who founded DeepSeek, has never worked outside his own company. These are outliers—people without “standard resumes”—precisely the kind the current investment system tends to overlook.

Finding such people offers real alpha, but few are currently seeking them.

Shenzhen and the Hardware Ecosystem

What impressed me most in China wasn’t startup roadshows.

It was Shenzhen’s underground hardware workshops—engineers systematically reverse-engineering Western high-end products, disassembling each component with extreme rigor. When I left, I wasn’t sure if most Western hardware founders truly understood what they’re competing against. The network effects here are not theoretical—they are physical, dense, built over decades.

Data from the entrepreneurs I met confirms this: over 70% of hardware investments come from the Greater Bay Area, nearly 100% from China—meaning iteration cycles are unmatched by Western hardware companies.

Most founders I met are adopting DJI’s approach: developing consumer hardware in niche areas—electric wheelchairs, lawnmowers, next-gen fitness equipment—reaching 8 to 9 figures in revenue (USD), then leveraging customer base or underlying technology to expand into adjacent categories. Some companies are much larger than you might think. The strongest I saw was Bambu Lab, a 3D printing company most Westerners haven’t heard of, reportedly earning $500 million annually, doubling each year.

Pessimism Toward Chinese Software

When I left, my doubts about Chinese software opportunities deepened.

On the model front, China’s open-source efforts are strong, but closed-source models still lag significantly behind the best Western ones—and the gap may be widening. Capital expenditure differences are huge. Access to GPUs remains limited. Western labs are increasingly cracking down on distillation. Revenue figures tell the story: Anthropic reportedly generated $6 billion ARR just in February. China’s top model companies have ARR in the tens of millions of dollars.

In software startups, the mainstream profile is PMs and researchers from ByteDance, developing agentic or ambient consumer software for Western markets. Talent is indeed strong, but many of these products fall within the scope of features originally developed in large labs—each release could render them redundant. I was also surprised by China’s lack of large, fast-growing private software companies. In the West, aside from model companies, there are startups with ARR in the hundreds of millions or even billions—Cursor, Loveable, ElevenLabs, Harvey, Glean—growing rapidly. Such breakthrough private software companies are almost nonexistent in China—except for a few like HeyGen, Manus, GenSpark, which have all eventually left.

Valuation Bubble

Despite the poor outlook for software, bubbles are real—both early and late stage.

Early-stage, top talent from ByteDance, DeepSeek, and Moonlit Face is indeed cheaper than comparable US talent, but median valuations are converging. Consumer startups with no product often reach $100-200 million valuation. Pre-seed and Series A rounds exceeding $30 million are common.

Late-stage valuations are even harder to justify. MiniMax is valued at around $40 billion in the public market, with less than $100 million ARR—about 400x revenue. Zhipu is around $25 billion with $50 million ARR. To compare: OpenAI’s highest valuation rounds are roughly 66x ARR, Anthropic about 61x.

Private model companies like Moonlit Face raise funds using these public market benchmarks—rising from $6 billion to $10 billion, then to $18 billion within months. Crypto investors are familiar with this pattern: comparing private valuations to “pre-unlock” public market prices. Additionally, Zhipu and MiniMax maintain these levels partly because they are currently the only ways to gain exposure to the “China AI narrative,” which carries a premium. But as more companies go public, this premium will diminish. Finally, IPO windows are unpredictable—shutting suddenly without warning. No one can guarantee you’ll have time to exit before the benchmark prices shift.

Humanoid robot tracks are similar. China has about 200 humanoid robot companies, around 20 with over $100 million in funding, several valued in the billions—most with no revenue, planning IPOs in Hong Kong in 2026 or 2027. If this market is real, China’s hardware advantage could make the long-term landscape clearer. But commercialization may lag far behind current fundraising pace, and I doubt Hong Kong can sustain dozens of humanoid robot companies valued in the billions waiting to IPO. I’m staying away for now.

Asymmetry of Information to Watch

One unexpected thing: nearly every founder I met is initially targeting global markets before China. They use Claude Code, watch Dwarkesh’s podcasts, and are well-versed in San Francisco’s startup scene—often more than Western investors who haven’t been paying close attention.

Western hostility toward China is clearly greater than China’s toward the West. Chinese founders see no contradiction in combining China’s engineering execution and hardware depth with Western go-to-market and product thinking. When this combination forms within a capable founding team, truly remarkable companies can emerge.

Finding these founders—those who don’t fit the “standard resume template” optimized by local VCs—is what we are working on now.

Special thanks to @woutergort for sharing his excellent Chinese network, @PonderingDurian for organizing this trip, and Claude for patiently editing my ramblings on the plane.