An American field study on China's AI has recently been circulated widely within Chinese technology and venture capital circles. The author, José Maria Macedo, the founding partner of Delphi Ventures and co-founder of Delphi Labs, originally published his observations on X. The piece was subsequently translated and featured by numerous tech and mainstream media outlets, rapidly evolving into a highly influential observational text. Macedo spent two weeks in China conducting an intensive survey of the domestic AI ecosystem, meeting with an extensive roster of founders, venture capitalists, and CEOs of publicly traded companies. He admits that he entered the country, expecting a bullish outlook on the Chinese AI landscape, where there would be world-class talent operating at valuations significantly lower than those in the United States. However, his perspective shifted by the conclusion of his visit. On one hand, his assessment became more nuanced. He found the hardware sector to be more robust than anticipated, while software capabilities were notably weaker. On the other hand, certain observations regarding Chinese founders caught him off guard, a point that strikes at the very heart of the Chinese tech industry's most profound vulnerabilities.
To more objectively evaluate his observations regarding the Chinese AI ecosystem, it is essential to clarify the nature of Delphi Ventures as a venture capital institution. According to Delphi's official disclosures, Delphi Ventures is the entity responsible for providing investment advisory services, whereas Delphi Labs functions more as an incubation and acceleration platform within the Delphi framework. Delphi Ventures defines itself as a "thesis-driven, high-conviction" firm; its portfolio currently reflects support for nearly 200 distinct teams.
This determines its perspective on the Chinese AI landscape, which diverges from that of traditional USD-denominated funds. Delphi Labs maintains that its acceleration model functions as an "extension" of the founding team; they look beyond the mere fundraising narrative to scrutinize product architecture, mechanism design, strategic roadmaps, execution, and organizational synergy. They emphasize their expertise rooted in the investment activities of Delphi Ventures, the rigorous analysis of Delphi Digital, and the hands-on operational experience of the Labs team. In other words, Macedo's assessment of Chinese AI does not prioritize who is best positioned for an IPO or who tells the most compelling story. Instead, he evaluates who is genuinely creating new prototype narratives, new paradigms, and new organizational capabilities. Such a lens naturally favors non-consensus founders, systemic moats, and long-term compounding, while remaining acutely sensitive to, and perhaps skeptical of, templated resumes, short-term valuation bubbles, and derivative innovation.
Within this framework, the primary conclusion drawn from Macedo's visit is that the relative certainty of Chinese AI lies in hardware rather than software. In his account, the true sense of impact came not from a model launch event or a consumer-facing AI application, but rather from the "hardware underground" of Shenzhen and the Greater Bay Area, defined by mature supply chains, high-density manufacturing networks, the teardown and reverse engineering of advanced Western products, and the resulting ultra-short iteration cycles. Reports from ABMedia further indicate that one of the most significant recalibrations of his judgment during this trip was a distinct upward revision of his assessment of China's hardware capabilities.
Why hardware? Because the essence of hardware lies not merely in inventing new ideas, but in distilling product design, manufacturing, procurement, delivery, debugging, distribution, and cost compression into a singular system engineering feat. China's advantage in this domain has never been limited to individual enterprises. Rather, it is the industrial network itself. Macedo's article cites entrepreneurs claiming that over 70% of hardware components originate from the Greater Bay Area, with nearly 100% sourced from within China. He further utilizes cases such as Bambu Lab to illustrate that the expansion speed of Chinese hardware startups often defies the intuition of Western investors. Furthermore, he contends that the genuine moat of Chinese hardware is not that individual teams are inherently smarter, but that the entire industrial network has driven the cost of trial and error to an absolute minimum.
In contrast, Macedo is markedly more cautious regarding Chinese software, particularly in the realms of Large Language Models (LLMs) and application-layer startups. This should not be dismissed as a natural bias held by Western VCs against Chinese software. The deeper context is that the competitive focus of leading American AI firms is shifting. Google's latest enterprise AI strategy, for instance, explicitly positions Gemini Enterprise and agents as the core of its commercialization, emphasizing governance, compliance, and enterprise-grade platform capabilities. Simply put, frontier companies in the U.S. have transitioned from building models to building platforms, distribution, and enterprise workflows. Under this paradigm, any consumer or light application-layer company operating within the native capabilities of foundational models faces an accelerated risk of platform absorption. Macedo's critique that Chinese software teams appear to be "impressive upgrades of existing products ("V2"), not genuinely original bets" suggests, within this context, that many opportunity windows for Chinese software entrepreneurship are being drastically compressed by global foundational model platforms.
This pressure has been further amplified recently by the controversy surrounding "distillation". In February of this year, OpenAI submitted a memorandum to the U.S. Congress alleging that DeepSeek utilized circumvention techniques and programmatic methods to harvest outputs from American models for distillation purposes. Shortly thereafter, Anthropic publicly accused DeepSeek, Moonshot, and MiniMax of engaging in over 16 million interactions with Claude via approximately 24,000 fraudulent accounts, explicitly elevating such activities to the level of national security and export control concerns. Simultaneously, the Frontier Model Forum released a briefing titled Adversarial Distillation in February, which established a clear distinction between "authorized distillation" and "adversarial distillation", signaling a push toward a unified industry standard. Consequently, Macedo's assertion that the catch-up space for Chinese closed-source models is being constrained stems not only from disparities in GPU availability and capital expenditure but also from the fact that leading U.S. labs are transforming "anti-distillation" from a series of isolated defensive measures into a matter of broad industry coordination.
While the strength of hardware and the comparative weakness of software are the most visible conclusions of this piece, a far more significant insight lies in the assessment of China's internal AI innovation mechanisms. Macedo argues that the core issue facing Chinese AI innovation is not necessarily a deficiency in talent or capital, but rather a screening process that systematically rewards high-level standardized answers over perilous originality. He repeatedly emphasizes that the vast majority of entrepreneurs he encountered in China are exceptionally diligent, maintaining an intense, around-the-clock work ethic. Their resumes are impeccable, featuring elite academic credentials, backgrounds at major firms like ByteDance or DJI, and an impressive array of papers and patents. However, the specific archetype he sought, i.e., the "independent thinker" characterized by rebellion, obsession, and the ability to formulate entirely new questions, remains relatively scarce. He attributes this phenomenon to a supply of entrepreneurs shaped by the combined forces of education and venture capital. While the educational system excels at training high-functioning executors, Chinese venture capital firms further reinforce this by prioritizing pedigrees, endorsements, and "Big Tech" labels. Consequently, the market produces a high volume of founders who are undeniably excellent, yet fundamentally similar.
This observation is one of the aspects in the article that deserves introspection within the Chinese investment community. When an educational system privileges standardized excellence, and a capital ecosystem utilizes "pedigree" as its primary filter for entrepreneurs, the result is a mechanism that systematically reinforces "non-innovation". It is not an absence of innovation entirely, but rather a proficiency in optimizing known problems, accelerating validated paths, and replicating successful paradigms. Conversely, this framework struggles to accommodate founders who initially appear unconventional or who do not fit the established profile of a successful candidate. The hardware sector continues to yield results because it relies heavily on supply chain density, engineering iteration, and organizational execution, areas where standardized high-achievers naturally excel. However, foundational models, native software, and new-paradigm products depend far more on non-consensus intuition, the endurance of long-term isolation, and the pursuit of directions that are initially misunderstood. In other words, the true deficit in the current Chinese AI ecosystem may not be a lack of intelligent individuals, but rather the absence of a mechanism that allows "outliers" to persist and thrive over the long term.
Consequently, the issue of valuation extends beyond a simple calculation of costs. Instead, it is a question of what capital is incentivizing. In his analysis, Macedo expresses a palpable wariness regarding the valuation bubbles within the Chinese AI software and humanoid robotics sectors. He notes that many early-stage AI projects command premium valuations well before establishing product-market fit or generating revenue, while the late-stage market drives a handful of scarce targets to levels that defy fundamental economic logic. Even if one does not accept every valuation figure in the text as an absolute metric, they indicate a clear systemic tendency: as capital increasingly gravitates toward "replicable success narratives", "prestige labels", and the "speculative space of imminent IPOs", funding is more readily funneled into projects that fit a template rather than teams that possess the potential to shift paradigms but appear uncertain in their early stages. This explains why he identifies "genuine alpha" as residing with those founders who fail to conform to the aesthetic preferences of local venture capitalists.
Naturally, Macedo's article should not be viewed as an indisputable truth regarding Chinese AI. Rather, it functions as a field report marked by distinct investment biases. Some of the data remains unverified, certain judgments are heavily contingent upon the specific samples he encountered, and several conclusions clearly reflect the personal aesthetics of a crypto-native, frontier-tech investor. However, it is precisely for these reasons that the piece reveals a more significant set of questions. Through the lens of a frontier American investor, China's most formidable asset is not its performance on model leaderboards, but its systemic hardware capabilities. Conversely, the primary cause for concern is not the quality of its talent, but its internal mechanisms for selecting innovation. This distinction is perhaps more deserving of attention than the reductionist debate over whether China is "strong" or "weak" in the global AI race.
Final analysis conclusion:
The true value of Macedo's observations regarding the Chinese AI ecosystem lies not in the superficial judgment of being strong on hardware and weak on software, but in the revelation of deep-seated structural issues within China's innovation framework. As Chinese AI enters its next competitive phase, a critical question remains: is capital paying for the comfort of certainty, or is it carving out space for genuine originality? If the Chinese AI ecosystem is to transition from "incremental repair" to "strategic innovation", the priority should perhaps not be the replication of another trending sector. Instead, it must center on providing those who lack standardized credentials yet possess an authentic impulse for original creation, along with more sustained and tolerant institutional and financial support.
______________
Xia Ri is an Industry Researcher at ANBOUND, an independent think tank.
