AI Industry Watch · Chinese Founders’ Night: The Game of Chinese Founders Going Global in the AI & Token Era — co-hosted by Google Cloud, Master Concept, and Unique Research, Singapore, during TOKEN2049
“This is the best era for AI startups. None of the AI applications have actually begun yet.”
Let’s start with a few numbers.
Nearly half of Americans now use AI — but only 4.5% actually pay for a subscription. Last year that figure was 2.1%.
Most people pay around $25 a month, and just 1% of paying users account for roughly 20% of total spend.
These numbers come from an a16z report, and Wu Wei, founder of Unique Research, threw them out on stage at an event in Singapore. The moment he finished reading, Kevin Gao, sitting beside him, added a line:
“Of that 1%, maybe 90% are Engineers.”
It sounds like a joke, but sit with the numbers for a second and they’re deeply counterintuitive: AI startups are so crowded that a hundred new products appear every day — and yet more than 95% of users have never paid a cent.
Is the market broken, or have the good times simply never started?
The timing of this conversation matters. It happened during TOKEN2049, when everyone working on AI and Web3 from around the world had gathered in Singapore — streets full of people talking tokens and compute.
The “Chinese Founders’ Night” gathered a different crowd: Chinese AI founders who are actually rooted in Singapore, knocking on clients’ doors one by one.
Their conversation had nothing in common with the main-stage grand narratives. It was all arithmetic.
The event was co-hosted by Google Cloud, Master Concept, and Unique Research, titled “Chinese Founders’ Night: The Game of Chinese Founders Going Global in the AI & Token Era.” Four people on stage:
Jimmy Zhao, head of Master Concept’s startup ecosystem, who has connected with over a hundred startups using Google services in recent years;
Yuchen, AI product lead at amber.ac, who just wrapped up a global startup competition in Singapore with 319 projects registered;
Kevin Gao, CEO of Arros AI, who spent fifteen years as an executive at Adecco before founding his own company — self-described as a “lao deng” (old hand) founder;
Tedd Li, founder of DEEPTENSOR, who moved from Toronto to Singapore this July, with Huawei Noah’s Ark Lab and Cerebras on his résumé.
The thing that struck me most listening to the whole roundtable: they all spoke blunt truths.
Products alone no longer carry value. To B is agonizing. Small clients basically lose money. Of 300-plus projects, fewer than 10% survive a year.
And yet these same people, all at once, said the same thing: AI applications are just beginning.
Why Singapore?
Tedd’s path is fairly representative. GPT had just come out when he was in the first year of his master’s degree; he was researching natural-language reasoning. His first project at Huawei Noah’s Ark Lab was knowledge distillation. Later he went to Cerebras and ServiceNow. This July he moved from Toronto to Singapore.
Is Singapore a springboard, a transit station, or a starting point for AI founders? Tedd’s answer was honest: first of all, it was a personal choice. When he was working as an independent consultant, a Singapore company had already reached out to work with him on AI solutions.
But objectively, Singapore touches both worlds: you can reach entrepreneurs from mainland China, and you can reach North American capital. “Singapore is a fairly ideal port — a Middle Port, a transit harbor.”
Kevin has been in Singapore for three or four years, and his framing is more down-to-earth. He has many Chinese enterprise clients, and their reason for choosing Singapore is simple: it’s a predominantly Chinese community, communication has no barriers, yet the legal and political environment is aligned with the West.
“Many Chinese companies going global — especially Chinese enterprises — treat Singapore as a testbed for going overseas.”
If they do well here, they ride the Chinese community and cultural ties into Southeast Asia, and then toward the US and European markets.
Jimmy added a more technical angle. Helping over a hundred startups apply for Google for Startups support over the years, he noticed that many companies only start seriously thinking about questions they never had to think about before after landing in Singapore: where to connect cloud services, where to put infrastructure, where to source models, where to put the database.
Before, it was a Chinese company with Chinese products for Chinese customers. Going global, all of that has to be re-chosen.
So Singapore is more of an adaptation point for these Chinese founders. They come in with Chinese products and technology, and they need to leave as global companies. That transformation happens in Singapore.
Kevin was the most outspoken person on stage.
He spent fifteen years in the recruiting industry, and has been doing an AI startup for two years. He opened with self-deprecation: “Investors like to fund two kinds of people: ‘old drivers’ and ‘young geniuses.’” Three young geniuses on stage, plus one old hand.
The old hand’s judgment:
“I think AI applications have just begun. It’s like the internet in the 1990s — the infrastructure isn’t even finished yet.”
You might ask: how can it still be early when things are this crowded? Kevin offered two observations.
First: “In vertical industries, there are very, very few companies running AI applications really well.”
The second is more interesting: two years ago when he went to see clients, many people reacted to AI products with “what is this thing, is it a scam?” But this year, many clients are reaching out on their own — even the ones he contacted two years ago have come back asking for a trial.
And once a client pays and actually uses the product, they basically never leave, “because it’s cost-cutting and efficiency-gaining, after all.”
This shift is the story behind that 4.5%. Penetration is still pitifully low, but the direction has flipped: from founders begging clients to try, to clients knocking on the door themselves.
The paid rate went from 2.1% to 4.5% — more than doubling in a year. At this slope, the real explosion is still ahead.
In the second half, the conversation moved to commercialization, switching from idealism to accounting mode.
Kevin laid out the underlying logic first: “Customers ultimately pay for value.”
Individuals pay for Hongguo short dramas because they’re buying a service and an experience — not because they’re paying for tokens. The B-side is the same: customers don’t care which model you use, they only care what results their money buys.
So should you go after big enterprises or SMBs? Kevin’s math was clear:
“One big B client can sustain a company of ten thousand people.”
Palantir, to some extent, was built on one big client — the US government — and grew to a valuation of tens or hundreds of billions. But big-B competition is brutal, procurement cycles are long, and your product has to be ten times better than competitors’ just to get in the door. It’s very unfriendly to startups.
Small and medium enterprises? The entry bar is low, but so is their ability to pay. “Starting from SMBs, you’re basically losing money.”
He stepped on this rake himself. Arros AI used to sell SaaS at $399 or $499, grinding through a pile of SMB clients, “with conversion cycles of three to six months. Exhausting.”
Customers had willingness to pay, but limited budgets.
Then they pivoted to big B: “One big B client paying several hundred thousand a year, and that sustains the company.”
But big B has its own problems: you can’t scale it through replication, and sales has to be done one account at a time.
“We have exactly one salesperson in the company — me,” Kevin said. “I’m the product manager, the salesperson, and the fund-raiser all at once.” Could sales be moved to Malaysia? They tried. Went through several people, spent a lot of time. “It’s actually worse than doing it yourself.”
What about letting AI do it? “AI can only help with lead generation — it gives you leads. But you still have to follow up, and explain why the product creates value.”
This is the real startup scene of the AI era: technology changes by the day, but the old problems — sales, delivery, trust — haven’t disappeared at all.
Tedd complemented commercialization from another angle: pay-per-result. He’s now embedded in an HCM company doing workflow transformation, and his observation is that if a company can see 50% or 60% efficiency gains, they’re very willing to pay, “because in the end the money they save is more than the money they spend.”
He also noticed a slower but more certain trend: AI is entering highly regulated, highly sensitive domains — finance, payroll, accounting — processes that demand extreme accuracy, where mistakes have serious consequences.
These are exactly the hardest bones for AI to chew, and exactly the direction he wants to chew. The logic is simple: the higher the cost of error, the more expensive human labor is, and the greater the value of AI replacement.
But pay-per-result has a fatal knot: how do you prove the result? You need a highly trustworthy partner to give you honest feedback — “I saved 60%” — and only then is it 60%.
He’s recently been tracking a more native solution, like a Payroll API: help companies process payroll, help individuals file taxes — one API call completes one piece of work. It’s naturally measurable, naturally chargeable per call.
Whether that business can scale remains to be seen, but the billing model is inherently smooth.
PLAUD, a company Unique Research has worked with, is a ready reference point: it makes AI voice-recording hardware, appears on the AI Consumer App ranking, and its subscription revenue and hardware revenue each make up about half.
That structure is worth pondering: hardware gets users in the door, but the real long-term business is the subscription that follows. Hardware is the entry; services are the repeat purchase.
Which is bigger, To B or To C? Yuchen’s view: both exist, but each has its own traps. C-end is closer to consumers, gets timely feedback, and can turn around fast; B-end can land large orders quickly, but too much enterprise-grade customization risks “a leaf blocking the view of the forest” — you think you’re building a product, but you’re actually running projects.
There’s no right or wrong — it all depends on your own resources and product form.
If Kevin was the one pouring cold water on To B, Yuchen was the one pouring it on products.
His point can be said in one sentence:
“Making a product is just a starting point. What matters more is building your product’s moat.”
The reasoning is solid. Anyone can vibe-code now; the cost of making a product has collapsed.
“A product that used to take a team a month or two now takes one person a week or two.” When products are that cheap, the product itself can no longer be a moat.
He has first-hand data. amber.ac just finished a three-month global startup competition at CHIJMES in Singapore: 319 projects registered, almost all AI startups, spanning about five vertical tracks including OPC and agents.
The conclusion: there was no shortage of gorgeous products — great UI, silky experience. “But the ones that made Top 10 were, without exception, projects whose market demand had been validated to varying degrees.”
Conversely, the trap that strong product teams most easily fall into is self-indulgence (”zihai”): thinking a feature is cool and stuffing it into the product. It feels great while you’re stuffing. But whether these things are useful before demand validation — nobody knows. The only certainty is that the cost of adjusting later keeps rising.
How many of those 300-plus projects will still be alive in a year? Yuchen’s answer: “From our experience in previous years, actually fewer than 10% survive.”
Jimmy added two words: “That’s conservative.”
Yuchen himself thinks even that estimate is optimistic. Three people, three lines — they turned the AI startup survival rate into the coldest number of the evening.
Why do so many die? Kevin jumped in with the explanation: “The key is that they’re constantly pivoting, because technology iterates too fast.”
The product you build on this year’s model capability becomes a built-in feature of the big model next year, and your moat disappears.
Startups can only sprint alongside the technology. Those that can’t keep up, or pivot in the wrong direction, become the 90%.
But think about it: this is the other side of the same story as that 4.5% at the opening. Demand is real, opportunity is real — but most products never live long enough to be validated.
So Yuchen’s final message to founders: validate market demand first. Figure out where your minimal core loop is. “Don’t build products in self-indulgence. Absolutely don’t self-indulge.”
Jimmy is the observer. He also had a striking set of numbers.
Google for Startups gives startups cloud credits of up to $250,000. Roughly 100 companies got approved through Master Concept. “But the number that actually manages to use it all up — maybe 10. Definitely fewer than 20.”
That number stings if you think about it.
If you can’t even spend free money, it means most companies haven’t even developed a real compute need — their business volume simply hasn’t taken off.
A $100,000-a-year credit, used properly, is a signal that the company is doing reasonably well.
Conversely, an unused cloud credit is a signal that a company hasn’t found a real need yet.
Can you resell the unused credit? No. “That’s non-compliant.”
On the back of this observation, Jimmy laid out a “fast and slow” framework.
Market validation must be fast. “If you have an idea, validate it fast — because if you don’t validate it today, in two months ten companies with similar products could be doing the same thing.”
Technology iteration must be fast. Models keep upgrading; your product has to keep up, or even get ahead. He threw out the question every AI application founder has to answer: “Why would users use your product instead of Gemini directly?”
But organizations can be slow. Products that used to need dozens or hundreds of people now need only ten or twenty — there’s no rush to expand headcount.
Security should be handled early. The event coincided with TOKEN2049, and Singapore was full of security talk — just days before, a Korean bank and a Japanese cloud vendor had been breached, with big losses.
Giving employees overly broad permissions early on to ship versions fast — those small risks stack up into a big vulnerability.
“If you can solve security problems early, it costs far less than patching after something breaks.”
Summed up in one line: be fast enough, while controlling risk. Like F1.
Kevin’s words: “This is the best era for AI startups. None of the AI applications have actually begun.”
Tedd’s words: “AI applications at the enterprise level — especially deep in production processes — have only just begun.”
Put those two statements next to the opening numbers, and the whole evening’s logic lines up: the 4.5% paid rate says the money is still ahead; the fewer-than-10% survival rate among 319 projects says the road is hard; the hundred companies that can’t spend free cloud credits says most people haven’t even found the right way to enter.
But precisely because of all this, the window is still open. If you wait until the paid rate hits 40% to get in, that’s when it’s truly too late.
Moderator: Wu Wei, founder of Unique Research
Guests: - Jimmy Zhao, head of startup ecosystem, Master Concept - Yuchen, AI product lead, amber.ac - Kevin Gao, CEO & co-founder, Arros AI - Tedd Li, founder & CEO, DEEPTENSOR
Wu Wei: Today we’re talking about AI and Chinese entrepreneurship. Please introduce yourselves, then share what you think is the biggest trend right now, or the most important view about AI’s development. Jimmy, why don’t you start?
Jimmy Zhao: Hello everyone, I’m Jimmy from Master Concept. I’m responsible for the startup ecosystem, mainly providing support to startups using Google services. I’ve connected with over a hundred startups over the years and have seen their development trajectories.
My view: in the process of globalization, some things need to be done very fast, and some things should go slowly. What needs to be fast is mainly technology iteration and market validation. Also, as I emphasized earlier, data security and compliance should be handled as early as possible.
What should be slow: talent and organizational development. Organizational growth can be slower, because products that used to take companies of dozens or hundreds of people to build can now be done by ten or twenty people. So organizational development can be as slow as possible. That’s one observation from this period.
Wu Wei: So technology needs to keep up with the iteration speed, but organizational development can be slower. Yuchen?
Yuchen: Hi everyone, I’m Yuchen from amber.ac. I focus on incubating AI startups, and on AI product innovation and landing.
For one view: in today’s AI-saturated world, anyone can vibe-code, and the cost of making a product has collapsed. But making a product is still a huge step away from commercial landing.
So my view is: making a product is just a starting point — what matters more is building the moat around your product.
Wu Wei: AI technology is now accessible to everyone and making products is easy, but making one well, with barriers, and commercializing and scaling it — that’s the challenge, right? Kevin, introduce yourself. You can share a view, a challenge, or a non-consensus opinion — pick one point.
Kevin Gao: Hi everyone, I’m Kevin, co-founder and CEO of Arros AI. I was always in the recruiting industry — a senior executive at Adecco in Singapore for fifteen years.
I’m what you’d call a “lao deng” (old hand) founder. Investors like to fund two kinds of people: “old drivers” and “young geniuses,” right? On stage we have three young geniuses plus one old hand.
My feeling after two years of founding: AI applications have just begun. That’s an intuition. Why do I say intuition? Because Mr. Wu gave me a question — can I state the question first?
Wu Wei: Briefly state your conclusion and view; we’ll expand later.
Kevin Gao: OK, let me state the conclusion first. I think AI applications have just begun — it’s like the internet in the 1990s, the infrastructure isn’t even finished.
Wu Wei: So the opportunity is huge and it’s still very early, right?
Kevin Gao: Too early. In vertical industries, there are very, very few companies running AI applications really well.
Wu Wei: We’ll come back to that. Tedd?
Tedd Li: Hi everyone, I’m Tedd. I moved from Toronto, Canada this July. I previously worked at Huawei Noah’s Ark Lab in Canada, then as an engineer at the recently hot Cerebras, and also spent some time at ServiceNow.
This year has been special: I spent time on the now-popular OPC, and did some consulting and related AI optimization work alongside my main work. I felt it was time to come to Singapore, hoping to do more enterprise workflow optimization. We’re currently embedded in an HCM company, turning their workflows into AI-driven solutions.
We’ve hit many challenges. My view: AI applications are gradually entering sensitive, high-compliance domains — finance, payroll, accounting — processes that demand extreme accuracy, where errors have serious consequences. How AI enters these domains is the direction I want to pursue.
Wu Wei: Quick question: why did you move from Toronto to Singapore? For AI development, is Singapore a springboard, a transit station, or a starting point?
Tedd Li: I think it’s more of a personal choice. When I was doing OPC, I did a lot of consulting. There was already a Singapore company working with me on AI solutions.
I’ve watched AI develop the whole way. When I was in the first year of my master’s, GPT had just come out. I was researching natural-language reasoning on Stanford’s SNLI dataset. My first research project at Huawei Noah’s Ark Lab was called Knowledge Distillation.
At the time, nobody was sure about the mainstream approach. After GPT-2 and GPT-3 came out, suddenly there were very large models, and Huawei quickly followed with the Pangu large model. People genuinely weren’t sure how capabilities would change as model scale grew. If I remember right, it hadn’t come out yet.
Wu Wei: 2021, 2022?
Tedd Li: I was working in 2020.
Wu Wei: 2020?
Tedd Li: End of 2020, yes. There were a lot of layoffs in North America, so I chose to join a research institute.
Wu Wei: So moving to Singapore was a personal choice, unrelated to the place itself?
Tedd Li: Of course, Singapore is also a good springboard. Here you can reach entrepreneurs and businesses from mainland China, and also North American capital. So for me, Singapore is a fairly ideal port — a Middle Port, a transit harbor.
Wu Wei: Kevin, how long have you been in Singapore?
Kevin Gao: Three to four years.
Wu Wei: What does Singapore mean to Chinese AI founders?
Kevin Gao: We have many Chinese enterprise clients. I think many Chinese companies choose Singapore mainly because it’s a predominantly Chinese community, and at the same time the legal and political environment aligns with the West — it uses the Western system. So many Chinese companies going global, especially Chinese enterprises, treat Singapore as a testbed for going overseas.
If you do well in Singapore — because there’s the Chinese bond, community, culture — it’s easier to connect internationally, and easier to enter unfamiliar environments, like Africa.
Wu Wei: But then would you be stuck circling around the Chinese community, unable to break out?
Kevin Gao: I’d say it’s a testbed, a transition — right? Because the rule of law and business environment here differ from China’s. At the same time it connects to Southeast Asia and pan-Asian markets, and to developed European and North American markets, because the legal and political systems here resemble those markets.
Wu Wei: So one is a Middle Port, one is a testbed. Yuchen, how long have you been in Singapore?
Yuchen: I actually spend more time in Hong Kong, but amber.ac has dual headquarters in Singapore and Hong Kong, so we come to Singapore quite frequently and run a lot of global events here.
Wu Wei: Comparatively, which of the two is more international, or better suited to your current business?
Yuchen: I think both have pros and cons; both are fairly international. Hong Kong leans more toward fintech, while Singapore currently leans a bit more toward AI.
Wu Wei: I think this is also soil suited to AI business, right? Jimmy, what’s your view?
Jimmy Zhao: The previous speakers were mostly thinking from the founder’s perspective. Over the years I’ve connected with over a hundred startups applying for Google support.
Google has a Google for Startups funding program: startups can apply for Google Cloud Credits, up to around $250,000.
After I help clients get approved, I watch these companies as an observer. I think more companies treat this place as a testbed. I strongly agree with Kevin’s point: start your globalization in Singapore, so you can connect with other countries faster, and respect and adapt to local laws and customers.
Of course, from my own work perspective, I also think they should reconsider: where to connect cloud services? Where to put infrastructure?
Before, it was a Chinese company, Chinese products, Chinese customers. Once in Singapore, they can think about where to put infrastructure, where to source models, where to put the database.
These questions deserve early thought. Singapore is a great adaptation point, and these questions can also be thought about from a technical angle.
Wu Wei: In the first half of this panel, we mainly talked about trend judgments and the feel of founding a company in Singapore. In the second half, we want to talk about commercialization of AI products.
Earlier we mentioned that helping enterprises optimize workflows hits bottlenecks. I happened to see an a16z report the day before yesterday, and a few numbers caught my attention.
First is AI’s paid rate. Guess what it is? 4.5%. Nearly half of Americans use AI, but the paid ratio is 4.5%; last year it was 2.1%.
So I strongly agree with Kevin: the opportunity is huge, because more than 95% of people have never paid for these subscriptions.
Also, most people pay around $25 — well, the starting point is $20, though I’ve always been at $200 (laughs). Within that, 1% of paying users account for roughly 20% of total spend.
Kevin Gao: Of that 1%, maybe 90% are Engineers.
Wu Wei: Engineers, right? Yes — coding agents. I think these are signals worth discussing. Earlier Jason mentioned PLAUD, which we serve. PLAUD is also on the AI Consumer App ranking. It makes hardware, but it also has subscription revenue — subscription and hardware revenue are about half and half.
The reason I set up all this context: I want to discuss an important topic — what opportunities did you discover while founding your own companies? Maybe what you’re doing now, or maybe you think what we’re doing isn’t as good as doing emotional companions or something, which might find value users are more willing to pay for.
So I want to hear: which AI products do you see having real paying opportunities? And in your own startups, what’s the bottleneck to getting clients or users to pay more? Let’s start with Tedd and work around.
Tedd Li: Since I’ve just started and I’m still in the client pilot stage, I can only share my own intuition. I think many enterprises still place great importance on replacing labor with AI. If they can see efficiency gains of, say, 50% or 60%, they’re very willing to pay. Because in the end, the money they save is more than the money they spend.
Wu Wei: In that process, would the payment model change? Like result-based services?
Tedd Li: It depends on what kind of service you provide. For consulting, per-use payment might work; but if you’re building a more general product, SaaS might be better.
From my current intuition, result-based services are still quite hard. First, you have to prove the result — you need a highly trustworthy partner to give you real feedback: “I saved 60%,” and only then is it 60%.
Wu Wei: So how do you measure it?
Tedd Li: Right, that’s the challenging question I’m currently facing.
Wu Wei: What might be a solution?
Tedd Li: I’ve been looking at a direction recently, like a Payroll API. You help a company do payroll, or help an individual file taxes — one API call completes the work.
That’s measurable, so you can charge per call. Whether that business can ultimately scale, we’ll have to see.
I think this is a more native kind of API approach — naturally chargeable by result.
Wu Wei: Understood. At least when doing enterprise AI, you can look at the number of completed tasks, right? By task, by result — various models are worth trying.
Kevin should have a lot of experience here. We were just talking about payroll, and you also do human-resource services, right?
And I see that you might have been more To-big-B before, and now you might also be doing some blue-collar work, right? How do you think about it?
Kevin Gao: We’re starting to do C-end now.
Wu Wei: Your customers and users are changing. How do you make that adjustment, or stack them?
Kevin Gao: Actually, Elon Musk has a great saying: “first principles.” Mr. Wu just said, do customers want to pay for tokens? In the end, customers pay for value. What results can my money buy? What can individual consumers buy — what service? Like Hongguo short dramas: customers can buy that service, so they’re willing to pay. But they’re not buying tokens.
Wu Wei: There’s no livestream today, so speak freely.
Kevin Gao: Indirectly, they’re buying a service. B-end clients can be picky; large enterprises have long procurement processes; your product has to be ten times better than competing products just to get into a big B. For SMBs, the entry bar is lower, and you can provide value, but SMBs have lower ability to pay — so starting from SMBs, you’re basically losing money.
Wu Wei: So is it To big B, or something else?
Kevin Gao: Look at your product and your resources — those are the two things. Because everyone wants to do To big B; one big B can sustain a company of ten thousand people, right? Palantir might rely on just one big client — the US government, right?
Wu Wei: Right.
Kevin Gao: And it’s grown into a company valued at tens or hundreds of billions. So excellent companies definitely want to do big B. But big B competition is fierce, and it’s unfriendly to startups. Startups often enter from SMBs, which is relatively easier.
But entering from SMBs — how do you cut costs and provide a valuable service? That requires the founders here to build a genuinely disruptive product. Not just an idea — build the product and actually provide a valuable product or service to society and the business world. Only then can the startup...
Wu Wei: How are you charging right now?
Kevin Gao: We’re in transition. Originally it was SaaS — maybe $399, $499 — grinding through a pile of SMB clients, with conversion cycles of three to six months. Exhausting. Customers had willingness to pay but limited budgets.
Later we moved to big B, which was better — one big B client paying several hundred thousand a year sustains the company. But big B has a scaling problem. We have exactly one salesperson — me. I’m the product manager, the salesperson, and the fund-raiser. Hiring one person in Singapore is very expensive.
Wu Wei: You could put it in Malaysia.
Kevin Gao: Malaysia is possible — we tried. But we went through several people, spent a lot of time, and it’s actually worse than doing it ourselves, right? That’s the current dilemma.
Wu Wei: Or let AI do it, right?
Kevin Gao: AI can’t do it. AI can only help with lead generation — it gives you leads. But you still have to follow up and explain why the product creates value.
Two years ago, many people rejected AI products: “what is this thing, is it a scam?” But this year, many clients come to us proactively. The ones we contacted two years ago are now coming back on their own, willing to use it, willing to trial — the conversion rate is decent. And once they pay and actually use it, clients basically never leave, because it’s cost-cutting and efficiency-gaining after all.
Wu Wei: Yuchen, you see how agonizing their To B journey is, right? Do you think To B or To C is the bigger opportunity? Tie it to your own business.
Yuchen: I think To B and To C opportunities both exist, of course. C-end’s advantage is being closer to consumers and getting timely feedback. B-end does more enterprise-level customization and can land large orders quickly, but for the long-term product ecosystem, there’s a risk of the leaf blocking the forest. Of course, there’s no right or wrong between B-end and C-end — it entirely depends on your own resources and product form.
We just wrapped up a three-month global startup competition at CHIJMES a few days ago. During this period we reviewed over 300 projects — 319 projects registered in total.
Wu Wei: What was the mix like?
Yuchen: Almost all AI startups — including OPC, agents — about five vertical tracks. Among the projects we reviewed, there was no shortage of gorgeous products — great UI, silky user experience. But the ones that made Top 10 were, without exception, projects whose market demand had been validated to varying degrees.
A product that used to take a team a month or two now takes one person a week or two. When product cost drops like this, teams with strong product skills easily fall into self-indulgence: “this feature is cool, let’s stuff it in.” But before market validation, are these things useful? Maybe. But they will undoubtedly raise the cost of later product adjustment.
Wu Wei: So how many of those 300-plus startups do you think will still be alive in a year?
Yuchen: From our experience in previous years, actually fewer than 10% survive.
Jimmy Zhao: That’s conservative.
Yuchen: (Laughs) We’re all being quite optimistic.
Wu Wei: So they need to pivot fast, right?
Kevin Gao: The key is that they’re constantly pivoting, because technology iterates too fast.
Wu Wei: Jimmy, at Master Concept you run the startup ecosystem and have seen a lot of projects. Today we talked about Singapore and Chinese founders going global. Is there still opportunity here? From your observation, what should an AI founder pay attention to? And how would you expand on your “fast and slow” framework?
Jimmy Zhao: I was just thinking about this topic too. We genuinely feel it: support has been granted to about 200 startups, and of those applying through us for Google Cloud Credits, the ones that actually manage to use it all up — maybe 10, definitely fewer than 20.
Wu Wei: Could they sell it to someone else?
Jimmy Zhao: That’s non-compliant and not allowed. Many companies genuinely can’t spend the credit, because it’s $100,000 a year — if you can use it up normally, it shows the company is doing reasonably well. That’s one observation.
Back to Mr. Wu’s question: in globalization, what should be fast and what should be slow? What I said at the start was that market validation needs to be fast. If you have an idea, validate it fast — because if you don’t validate it today, in two months ten companies with similar products could be doing the same thing. So that must be fast.
Technology iteration must also be fast — you have to keep up with model development. Why would users use your product instead of Gemini directly? So technology iteration must be fast — ideally, get ahead of technological development.
Wu Wei: So you’ve summed up your view: speed must be fast.
Jimmy Zhao: Right.
Wu Wei: Like F1.
Jimmy Zhao: Right, like F1. Tying into today’s theme and TOKEN2049, everyone places great importance on security. As mentioned, yesterday or the day before, a Korean bank and a Japanese cloud vendor were breached by hackers, with big losses.
From a founder’s perspective, solving security problems early costs far less than patching after something goes wrong.
Wu Wei: So be fast enough while controlling risk.
Jimmy Zhao: Right. For example, in the early days, to move fast and ship new versions, you give different employees broader permissions, or grant excessive internal permissions. Those small security risks stacking up can turn into a very large vulnerability.
Wu Wei: I feel you’ve already summarized what you wanted to say. Given the time, the remaining three can each wrap up in one sentence — a feeling or a view. Like Jimmy just said: be fast enough while controlling risk and ensuring security. Yuchen?
Yuchen: I still want to talk about fast and slow.
Wu Wei: That’s fine — just a final one-line summary.
Yuchen: I think everyone should work hard on products, but you must first validate market demand. Figure out where your minimal core loop is. Don’t build products in self-indulgence. Absolutely don’t self-indulge.
Wu Wei: Good — nail PMF first. Kevin, I’ll summarize for you.
Kevin Gao: One line: this is the best era for AI startups — none of the AI applications have actually begun yet.
Wu Wei: That also gives everyone great confidence.
Jimmy Zhao: In the AI era, rapid iteration matters a lot.
Tedd Li: I think that right now, AI applications at the enterprise level — especially deep in production processes — have only just begun.