Delphi Ventures co-founder José Maria Macedo recently shared his observations after making an intensive two-week trip to explore China’s AI ecosystem in depth. On this trip, he met in rapid succession with the founders in China’s AI space, venture capital firms, and the CEOs of publicly listed companies. He had originally gone in with fairly optimistic expectations, believing that China not only has world-class AI talent, but that its valuations still offer a clear discount compared with Western markets.
However, after the actual visits, his judgment has shifted toward a more complex conclusion: compared with before he left, he is even more bullish on China’s hardware capabilities, but more conservative about the software prospects, and he has also revised his understanding of the traits of China’s startup founders. While China has top-tier talent, Shenzhen’s hardware advantages, and a complete supply chain, truly original founders are scarce. China’s AI software capabilities still lag behind the West, and valuation bubbles and the humanoid robotics boom are even more worth warning investors about.
New venture investors admit it plainly: China can’t find founders who go from zero to one
Macedo points out that the top founders he has invested in in the past usually share a very distinct commonality: they think independently, have a rebellious streak, are persistent and intensely driven; they are used to not following existing rules and always keep asking “why,” and they often make decisions that outsiders can’t understand at first, but that are extremely reasonable to them.
Yet in this trip, many founders he saw in China looked more like an entirely different archetype. These people are also exceptionally capable—their résumés are almost flawless. They come from top universities, have worked at ByteDance or DJI, have Nature papers, patents, and strong first-line technical backgrounds, and they are astonishingly hardworking. They fill up meetings on weekends, across cities, and at all kinds of time slots; even founders show up for appointments on the very day their wives give birth.
But in Macedo’s view, although this group is extremely strong at execution, they are relatively less likely to have the kind of startup impulse that is truly from zero to one and strongly original. Instead of posing new questions that have never been seen before, they more often produce excellent V2 versions in existing directions.
China educates “excellent people,” but can’t produce founders
He further attributes this phenomenon to China’s education and talent development system. Macedo believes China’s system is indeed very good at producing “excellent people,” but it may not provide enough room to encourage ways of thinking that deviate from the standard answers. What ultimately gets produced is a batch of high-level executors who are extremely good at solving known problems, not the outlier founders who proactively come up with a new market question that hasn’t been recognized yet—and are determined to build it. What’s even more noteworthy is that local venture capital firms are actually reinforcing this path dependency.
Many China funds today put their investment focus on founders with “elite résumés” who come from big companies like ByteDance and DJI. They value pedigree more than the kind of personal traits in founders that are hard to quantify—and sometimes even a bit sharp-edged.
Macedo believes this instead overlooks one of the most successful groups of founders in China’s history, who often didn’t come from the standard elite pipeline. Like Jack Ma, Ren Zhengfei, Liu Qiangdong, Wang Xing, and even Liang Wenfeng, who more recently built DeepSeek, to some extent are closer to outliers than to the typical choices that mainstream VC templates highly favor. He says outright that true alpha may actually be in people who don’t fit the mainstream résumé framework, but there don’t seem to be many investors in the market who are willing to proactively look for such people.
Shenzhen’s hardware supply chain is impressive
If the originality of China’s founders made him keep some reservations, then Shenzhen’s hardware supply chain and engineering capability were the most impressive parts of the journey for him. Macedo said that what impressed him most wasn’t a particular new startup’s pitch, but the underground-like, densely existing network of hardware engineering in Shenzhen. There, engineers systematically acquire Western high-end hardware products, then disassemble them part by part, do reverse engineering—and do it with extremely meticulous care.
He admits that after going around in a full cycle, he’s even not sure whether most Western hardware founders truly understand what level of competition they’re facing. This is not a supply-chain advantage that exists only on paper; this is an industrial capability built up over decades—extremely dense—and with physical network effects.
Fellow startup founders also emphasized to him that in many hardware products, more than 70% of components can be sourced directly from the Greater Bay Area, and nearly all critical inputs can be handled within China. This enables Chinese hardware companies to iterate, test, and adjust products at a pace far faster than their Western competitors.
He observed that many Chinese hardware founders are actually replicating the path DJI once took: entering first with a certain consumer hardware niche—such as electric wheelchairs, lawn-mowing robots, next-generation fitness equipment, and so on—scaling a single product category to a meaningful size, and then expanding into adjacent categories along existing customer bases or underlying technologies.
Macedo also specifically named a company that left a deep impression on him: Bambu. This 3D printing company is still not well known in the Western world, but it’s reportedly already generating annual profits of $500 million, and is continuing to grow at roughly double-digit rates. This also gave him a clearer judgment about where China’s true advantage in AI-related investment comes from: at least in hardware and manufacturing capabilities, China’s moat is deeper and harder for many overseas investors to understand—and harder to replicate.
China’s AI products can’t beat Western teams
By contrast, Macedo is much more conservative about China’s software and model layer. He acknowledges that China’s performance in the open-source model space is indeed impressive, but if it’s closed-source models, there is still a significant gap versus Western top labs—and that gap may continue widening in the future. Reasons include the huge differences in capital expenditures, continued restrictions on obtaining GPUs, and Western labs strengthening their blocking efforts on technical paths like distillation.
More importantly, revenue scale already tells the story. Macedo cites figures showing that Anthropic’s monthly revenue is shockingly high, while China’s best model companies’ annual recurring revenue still often remains at the tens of millions of dollars range. From the perspective of software startups, he also found that a fairly common startup combination in China’s market is former ByteDance product managers and researchers, attempting to build consumer software in Western markets that is agentic or ambient in nature.
These teams of course have strong capabilities, but the problem is that many of their products themselves actually fall within the scope of what large model companies’ “next native feature update” could directly eat up—so the moat isn’t deep.
What worries him even more is that China currently seems to lack the kind of truly fast-growing, breakthrough private software companies. Macedo mentioned that in the West, besides model companies like OpenAI and Anthropic, there is also a cohort of high-growth startups that have already built nine-figure—or even ten-figure—ARR, such as Cursor, Loveable, ElevenLabs, Harvey, and Glean. But China has almost no private software companies at that level.
With the few exceptions, like HeyGen, Manus, and GenSpark, in the end they all chose to leave the China market to grow. That leads him to believe that in terms of the explosive power of AI software entrepreneurship, China still can’t be compared to the United States.
Investor warnings: China’s AI valuations have already shown a clear bubble
Although his view on software is on the pessimistic side, that doesn’t mean China’s AI market lacks hype. Quite the opposite—Macedo believes China’s AI investment market has already shown a fairly clear valuation bubble, and it’s not only present at late stages; the early-stage market is also extremely hot.
He noted that although the top talent from star companies like ByteDance, DeepSeek, and Moonshot still trades at valuations cheaper than U.S. peers, if you look at the market median level, it has already quickly caught up. In China today, it’s already quite common for consumer-style startups that haven’t yet produced a product to be valued at $100 million to $200 million, and pre-seed rounds above $30 million are no longer rare.
China’s AI is like cryptocurrency—supporting a high valuation enables even higher private deal prices
In the late-stage market, Macedo’s tone becomes even more direct. He gave examples: Minimax’s valuation in the public market is about $40 billion, yet its annual revenue is less than $100 million—roughly a 400x revenue multiple; Zhipu’s valuation is about $25 billion, with revenue around $50 million. By comparison, the valuation multiples in peak fundraising rounds for OpenAI and Anthropic are only about 66x and 61x ARR, respectively.
Today, private model companies like Moonshot are using these public-market benchmarks to push valuations upward—reaching $6 billion, $10 billion, and $18 billion within just a few months. Macedo believes this kind of playbook is actually very familiar to crypto investors: at its core, it’s using inflated public-market valuations with limited liquidity to support even higher prices in the private market.
And the reason Zhipu and Minimax can enjoy this premium, to some extent, is that when the market wants to buy “China’s AI narrative,” there aren’t enough options. Once more targets go public, this scarcity premium could be diluted; additionally, the IPO window could close at any time, so the arbitrage opportunity may not be executed smoothly.
Investors worry about the commercialization rollout timeline for China’s humanoid robots
He also projects the same concerns onto the humanoid robotics track. Macedo points out that China currently has about 200 humanoid robot companies. Of those, around 20 have raised more than $100 million, and several have valuations that have reached the scale of several billions of dollars—but most companies still have virtually no revenue, and many are planning to list in Hong Kong in 2026 or 2027.
In the long run, if the humanoid robot market truly takes shape, China, leveraging manufacturing and hardware advantages, has a strong chance to achieve a dominant position in this industry. But in the short to mid term, he doubts that the speed of commercialization rollout will keep up with the fundraising tempo of the current capital market, and he’s also not sure whether the Hong Kong market can genuinely absorb so many robot companies with valuations that easily reach tens of billions of dollars each. So at this stage, he chooses not to participate.
China’s AI teams are heading internationally, familiar with Silicon Valley dynamics, using Claude Code
However, Macedo has not lost interest in China’s AI overall. On the contrary, he particularly emphasizes that an asymmetry worth the world’s close attention is forming: almost all the China founders he has encountered are now targeting the global market first, rather than starting with China’s local market. They are familiar with Silicon Valley startup dynamics, use Claude Code, follow content from the Western tech circles like Dwarkesh, and their understanding of the San Francisco startup ecosystem may even be deeper than that of some Western investors who haven’t closely followed the market.
Macedo says plainly that Western hostility toward China is far higher than China’s hostility toward the West; and China’s founders don’t see any contradiction in combining China’s engineering execution strength and hardware depth with the West’s go-to-market and product vision. When these two capabilities truly come together in the same founding team, the future may well produce a batch of highly competitive—and even quite surprising—companies.
That’s why, going forward, Delphi Ventures is most focused on those China founders who don’t fit the mainstream local VC aesthetics, but may truly have global-level ambition and original capabilities.
This article Top talent is everywhere, but can’t build OpenAI? In-depth two-week visits by investors reveal the real problems with China’s AI. First appeared on Chains News ABMedia.
