Allen Shih came back to my podcast for a second conversation. We met in person in Singapore during TOKEN2049 and the F1 weekend and talked about what he'd seen this year across the US, Japan, Taiwan and Singapore, and how his approach to coaching early-stage teams has changed.
His background is wide: physics, then manufacturing, tech, ride-hailing, games at Roblox, Web3, even concert investment. Now he advises on product and going global, and the teams he coaches get into YC almost every batch.
Singapore is a base, not a market
His sharpest point: don't treat Singapore as a market. Treat it as a US company's base in Asia.
If your question is how to turn Singapore into a good market, the business stays small. If you position it as a US team's foothold for Japan, Taiwan and Southeast Asia, it's a different story.
He sees US teams starting to take Asia seriously, including the big model labs building forward-deployed engineering capacity here. They don't know the region and need local partners, and not only for sales and business development but for systems integration and engineering on the ground.
For founders going the other way, he thinks the right path is to build a solid business in Asia first and then take on the US. Teams that set out from Asia to "explore the US market" mostly don't survive.
Culture fit is a wall, unless you pick the right category
What travels most easily? Content generation: video, audio, images. Users can tell immediately whether the output is good, without needing to understand local habits.
His counterexample is from ten years ago. When Didi considered Taiwan, he told them neither they nor Uber would do well there. Taipei's metro is too convenient and scooters are everywhere. Ten minutes on a scooter gets you anywhere. Demand for rides is naturally weak, and no product fixes that.
Then he turned to workplace collaboration agents, one of the hottest categories, and one of the hardest to take abroad.
"What does your team use every day? Feishu and Tencent Meeting. Overseas it's Slack and Teams. You don't use them yourself. How do you understand those users? How do you build a teamwork product for them?"
A teamwork product has to be good, invite the whole team in, and replace the Slack or Teams they already have. The switching cost is far higher than for a video generator.
A ten-billion-dollar niche you only find in New York
One team he coached wanted to build AI workflows for insurance and wasn't focused enough. He told them to go to New York.
They went, met investors, banks and insurers, and found something very specific. US companies can make errors in tax filings, and the consequences are severe enough that there is dedicated insurance for it. That market is worth more than ten billion dollars.
Without the trip, he said, you'd never know it existed. Teams arrive in the US with a bag of solutions, and the customer says: we don't need all this, you haven't tried to understand what we actually do here.
What you think is supply is often demand
His main lesson this year: look fast for mismatches between supply and demand.
His example is domestic helpers in Singapore. A team wanted to use AI to raise helpers' skills and understanding. He thought it wouldn't work. A good helper is good mostly because the family knows what it wants: cook well, do school runs, or talk with the kids. Every household is different and nothing is standard. It looks like a supply problem, but it's a demand problem: can the family describe the task?
He applied the same thinking to AI finance for restaurants. Don't keep drilling into receipts, invoice management and receipt categories, getting narrower and narrower. Use the receipt as an entry point to the industry's real problems: inventory, payments, logistics. Restaurants fail at high rates around the world, and the cause is usually cash flow and inventory, not the food.
Ten times the growth without hiring
The biggest problem he sees in coaching has changed. It used to be finding a direction. Now revenue grows faster than the team's capacity. Five or ten customers arrive and most teams can't absorb them, so they hire, and then management problems eat the gains.
So his question now is how to grow operations tenfold without hiring. That, he says, is what AI agents should really solve: not replacing people, but doubling what the current team can do.
YC wants to see how you use ChatGPT
The point that surprised me most. YC used to ask for your LinkedIn and social profiles. Now, he says, they want to know how you use ChatGPT, whether you commit to GitHub consistently, whether your coding agents keep producing, and whether anyone uses what you ship.
"This era is getting more and more real. Whether you actually build with AI isn't something you can fake."
He also described a shift in how accelerators and investors think. They ask what will be needed in six months to a year, not whether to chase today's trend. When AI coding took off, they were already looking at the bottleneck after it: everyone is writing code, and the pile-up is in review and release.
So if you apply to YC today saying you're building an AI video tool because AI video is hot, you're already late. The better question is what this amplifier will need from you in six months.
"Don't copy someone's answer. That's gambling. Learn their way of thinking and the pattern."
*This article draws from Allen Shih's second conversation on my Chinese-language podcast on AI founders going global.