The U.S.-China AI Startup Ecosystem Divide—A Conversation with Google Cloud’s Head of Venture Ecosystem for North Asia
Summary
Google’s support for and investment in startups are two separate tracks, while investment itself runs on two ledgers: strategic synergies and financial returns. Strategic investments are built around creating synergies across education, retail, security, and use-case deployment—With is one example. Google Ventures, on the financial side, brings in internal experts for due diligence and helps portfolio companies dig deeper after the investment. The main focus over the past 1-3 years has been Agentic AI, while healthcare remains a long-term direction.
The structural gap between U.S. and Chinese AI startups remains one of problem definition versus problem solving, but the application layer is closing the gap fast. 庄明浩 asked whether a new generation of founders with overseas education or resources is changing the picture. Warren Li’s view: “China still can’t match the U.S. at the foundational layer,” but AI is so new that applications do not need to look as far ahead—and Chinese teams are stronger at exploring use cases.
Asia is not one market; startup conditions differ chiefly in paying power, local relationships, and market and listing mechanisms. China is “clearly far ahead.” India’s B2B market largely follows the U.S., while B2C is held back by insufficient income and consumption. Japan’s strong enterprise payment habits and friendly early-stage listing system support a sizable middle tier. Korea has foundations in semiconductors, robotics, and healthcare, but a small domestic market. Southeast Asia’s population has not translated into paying customers. Warren’s pointed comparison: “10 million users in the U.S. are worth all of Indonesia.”
OPC has lowered the barrier to starting a company, but early-stage outcomes still hinge on product judgment and logic. Founders first need to see “where the opportunity is and what product can meet it” and think the idea through. In the past, repeated handoffs between PMs and engineering teams created communication loss; now, founders can act more like directors—generating, inspecting, and correcting in real time: “you immediately know what’s missing here.”
AI will first monetize productivity gains in industries and markets where labor is most expensive. U.S. finance, law firms, healthcare, and technology coding are early adopters because the economics of replacing expensive labor work. In Southeast Asia, cheap labor weakens the return on automation. Warren calls the idea that traditional industries will adopt AI faster than cloud “a hypothesis”: cloud computing has been discussed for nearly 20 years without full adoption, yet companies are “very proactive” about AI.
Warren does not think Google will absorb every vertical application; room for focused startups comes from the last mile, first-mover advantage, and big-tech priorities. Gamma’s PPT market may be worth only $1B-$2B, not enough to make it a Google priority. Once a startup has established itself, it would not make economic sense for a large company to spend $200M-$400M fighting for a $1B market. A $500M market is already enough to support a startup: “Build a company with $50M AR, and I think that’s good—comfortable.” Foundational bets, meanwhile, must be sufficiently non-consensus, with founders bringing resources or a clear comparative advantage.
The deepest moat in the U.S. startup ecosystem is not any single technology, but failure-tolerant capital, legal protection, M&A exits, and the motivation to “make an impact.” Independent research projects can receive funding simply to prove a proposition wrong; companies with traction are more likely to be acquired by incumbents than copied outright. Warren uses tech employees selling stock and reinvesting the proceeds into new companies to show how tax incentives recycle wealth into early-stage startups. He summarizes Demis’s standard for purpose this way: “If you can still walk after running a marathon, you didn’t try your best.”
Deep dive
1. Google Keeps Strategic Synergy and Financial Investment on Separate Ledgers
Asked how he evaluates startups, Warren first clarified: “Support and investment are actually two different tracks.” Strategic investment must generate synergies across education, retail, security, and deployment in specific use cases. Warren cited the partnership with With; 庄明浩 also mentioned cases such as Produce AI, arguing that this strategic-investment thread is clear.
Financial investing is not fundamentally different from ordinary VC. The difference is Google’s network of internal experts. When investing in data-related companies, in-house practitioners can participate in due diligence; after the investment, experienced colleagues can help the company identify and pursue further opportunities.
Over the past 1-3 years, the main focus has been Agentic AI, while healthcare has remained under observation. 庄明浩 asked whether that reflected LP preferences and fund structure. Warren agreed: U.S. capital can underwrite research-stage bets, while RMB capital typically “doesn’t want to look that far ahead.”
2. China Is Closing the Application-Layer Gap, but Zero-to-One Innovation Remains a Weakness
Warren used education to explain the U.S.-China gap: Chinese training places more emphasis on practice, test-taking, and clear answers, producing strong execution once the problem is defined. After the iPhone defined the smartphone category, Chinese teams proved highly capable of building products within that established category. But “defining innovation and defining the problem” remains an area where the U.S. is stronger.
庄明浩 asked whether a new generation of Chinese AI founders with overseas education or resources is changing the picture. Warren acknowledged that the gap is narrowing because AI is new enough that everyone is still exploring. His caveat: “From the foundational layer, China still can’t match it.” Applications, however, do not need to look as far ahead, and Chinese teams are better at exploring use cases.
3. Asian Markets Differ Sharply in Paying Power and Localization Constraints
Warren’s regional view is blunt: “Across Asia, if you look at entrepreneurship, China is clearly far ahead.” India’s B2B market broadly follows the U.S., and investors encourage founders to move to the U.S. Earlier bets on B2C did not deliver because income growth and distribution were insufficient to support a domestic consumer market.
Japan has very few B2C products, but strong B2B payment habits and a friendly early-stage listing system can support a large middle tier of companies. The challenge is distribution: small companies struggle to sell without the backing of a large enterprise, while foreign companies without local relationships have difficulty closing deals. They must be “extremely local.”
Microsoft is a useful example of localization. Even if its AI is not necessarily the best, Japanese customers feel reassured as soon as they see Microsoft, because its interfaces have been extensively customized around Japanese usage habits.
Korea has a strong base in semiconductors, robotics, and healthcare, making it “quite interesting.” The main constraint is the size of its domestic market.
Southeast Asia’s problem is not population but willingness to pay. Singapore has only 4M-5M people and high costs; setting aside its geographic and political advantages, its commercial customer base is limited. Indonesia and the Philippines have large populations, but B2C spending still has not taken off. Land and relationship resources are concentrated among a small number of large family groups, so foreign companies must invest heavily in local relationships to enter the market. Given the same investment, Warren would rather target Japan, whose IT market is far larger than that of any single Southeast Asian country.
4. OPC Lowers Engineering Barriers, Making Product Judgment and Logic Critical
As the barriers to starting a company fall, Warren still looks at the founder first. The most important factor is the product idea: the ability to spot where an opportunity exists and judge whether a product can meet it. Founders also need sufficiently strong logic.
In the past, a PM had to relay the logic point by point to the engineering team, with every round of communication introducing loss. Warren compares the process to directing a film: if you do not understand cinematography, editing, or special effects, you can only repeatedly describe the picture in your head. Now founders can create the work themselves, inspect it, and immediately see “what is missing here.”
AI’s core value remains productivity. The more expensive the labor, the stronger the adoption incentive. In the U.S., early adoption has come from finance, law firms and other professional services, healthcare, and technology and coding. In low-cost labor markets, as with early factory automation, the economics of replacing people may not work.
Warren deliberately frames the next claim as a hypothesis: traditional industries “will adopt AI faster than some other new technologies.” Cloud computing has been discussed for nearly 20 years, yet many traditional companies still do not use the cloud, potentially because of regulation, mindset, and internal politics. Faced with AI, companies this time appear highly proactive and eager to adopt it.
5. Vertical Applications Win on the Last Mile; Foundational Bets Must Be Non-Consensus
庄明浩 raised the concern that Google covers much of the AI stack, leaving startups vulnerable to being swallowed by big tech. Warren’s answer focused on Google: it cannot build every niche application, and vertical companies have room to win through last-mile delivery.
Gamma builds slides and PPTs, as does Google, but Google will not approach the category with Gamma’s level of specialization. A $1B-$2B market may not be high enough on Google’s priority list. If a startup has already established itself, a large company may need to spend nearly $200M-$400M to compete for a $1B market—a poor trade-off.
Founders should not treat fundraising scale as the finish line: “A $1B market, or even a $500M market, is enough to support a startup.” A company with $50M AR can be a very good business; software margins are high, and the ecosystem and connections across the value chain are sufficient.
Foundational opportunities such as chips are different. They sit in areas where Google is investing heavily and cannot afford to concede, and they are highly capital-intensive. A startup should enter only with enough initial resources and a clear advantage in execution speed or some other comparative advantage. Consensus directions that also consume huge amounts of resources have very low odds of success.
Warren uses the first-mover race in GPT and conversational products to illustrate the value of non-consensus thinking. Google already had Transformer, chips, capital, and talent, yet did not produce the defining product of this phase first. He describes AlphaGo as a more scientist-led path and contrasts it with Sam Altman’s more marketer-and-PM orientation, while emphasizing that success also contains an element of chance.
6. The U.S. Ecosystem Recycles Failure, Wealth, and Mission into the Next Innovation Cycle
In the U.S. and Europe, the legal system reduces the risk that a startup will be copied outright or forced out of the market. Once a company has traction, large companies are more likely to buy it than copy it. 庄明浩 called the environment “extremely enviable,” while Warren linked it to education, tolerance for failure, and multiple exit routes.
The U.S. has large numbers of independent research labs and research programs whose funding can support researchers in “proving that this thing is wrong.” If someone is willing to pay to disprove a proposition, innovators do not need to prove that every experiment can be commercialized immediately.
Warren uses early Facebook or Google employees as an example. Assume an employee has accumulated roughly $5M in stock: selling it all could trigger a tax bill of about 40%. If the employee invests $1M in a still-growing company, then under Warren’s explanation, that tax can be offset. The resulting incentive encourages more people to start companies.
The deeper motivation is not money but the desire to “make an impact.” If Elon Musk only wanted to make money, he would not have set sending people to Mars as a goal. Demis treated solving AI as a mission from a young age, with an almost brutal standard for “try my best”: if you finish a marathon, are sent to the hospital, and are not dead, then you truly gave it everything.
Verification Notes
- The original first says “OpenAI got there first,” then asks why Google was not first and DeepMind was; it also uses “Davinci” as a reference. The body avoids attributing first-mover credit to a single entity.