Count who gets paid: the least glamorous AI money in Southeast Asia

Somebody is going to stand on a stage in Singapore this month and tell you two things. Southeast Asia now has more than 60 profitable tech startups. And the region is in the middle of a $4 billion AI boom. Both numbers are true. Both are also lies. Not made up. Lies the way your uncle lies at dinner, where every word checks out and the whole thing is still nonsense. So I am going to tell you up front where this lands. The most profitable AI business in Southeast Asia is not building the model. It is not even the thing people are calling the AI boom. It is teaching people how to actually use it. That is a services business. It gets paid on day one, or net 30. And it is the one lane where being a follower is not a weakness. It is the strategy. I am writing this from KL with the haze back, the Twin Towers invisible from three streets away and a throat that sounds like I gargled a satay grill. We cannot see through the sky this week. We can still see through the press releases. One. The 40 people who never logged back in This did not start with a report. It started with watching the same movie play out across my portfolio and my network. A company I am close to spent good money last year buying every AI tool you can name. Seats for the whole team. I did the same thing myself. I had so many subscriptions I lost count, and I justified every one as my R&D budget. Fast forward to today and ask how it is going. Past the slides, the honest answer is this. A handful of people went fully red-pilled and got the lift. The rest logged in once, maybe twice, felt a bit confused or a bit silly, and quietly went back to doing the job the way they already knew how. Nothing was wrong with the tools. The tools are incredible. What was missing was the boring part. Nobody sat down with 40 people and taught them, in plain language, for their actual job, how to make the thing useful before nine o’clock on a Monday. Once you see it, you see it everywhere. Startups, SMEs, corporates. The gap is not the model. The gap is the person in the chair who was never taught, and was too embarrassed to say so. Multiply that person across 680 million people and you do not have a technology problem. You have the biggest teaching job this region has ever faced. And where there is a job that big that nobody has done yet, there is money. Two. Take a wrench to the headline numbers First, the 60-plus profitable startups. It comes from a single list, quoted everywhere, and nobody says who is on it or how profitable is defined. I sanity-checked it against my own portfolio and I would apparently be a meaningful chunk of the region’s total. I am a good solo GP. I am not that good. Here is what I will concede. Directionally, it is right. I sit across from early-stage founders every week, and genuinely profitable venture-backed companies are less rare than they used to be. The capital drought forced it. Treat the number as a direction, not a fact. The second one is worse. Native AI startups in the region raised about $4.1 billion so far this year, roughly double all of last year. Pop the champagne. Now pull the thread. Kling AI, the video generation unit of Beijing’s Kuaishou, raised a $2.8 billion Series D. That one round is about 68 percent of the entire “boom.” Strip it out and you are looking at a fraction of the headline. The number of rounds actually fell, from 41 last year to 23 this year. That is not a boom. That is one cheque and a thinner deal count. Of what is left, almost all of it lands in Singapore. With the usual caveat I will keep repeating until it stops being true: half the region is incorporated in a Singapore holdco, so the geography in these tables is partly an accounting artefact. So put both numbers down. Neither tells a founder in Manila or Bangkok where to make money in AI this year. Here is the rule I want you to carry for the rest of this piece: ***Do not count logos on a slide. Do not count dollars in one giant round. Count who sells a thing and collects cash.*** Three. AI projects do not fail. They stall on people. We have all seen the MIT-style reports on enterprise AI projects that never deliver a return. I do not think most of them fail, exactly. They stall. And not because the model is not smart enough, or the tool is badly designed. They stall because nobody knows how to use it, nobody trusts it, and the pilot dies quietly in a slide deck. If you have ever rolled out a CRM at a growing company, you know this movie. The software was never the hard part. Changing the daily habit was. Research on the roughly $400 billion corporate learning market lands on the same point: workforce readiness, not technology, is what blocks most enterprise AI. One survey has three in four companies saying they cannot keep up with their own demand for new skills. They have the tools. They have the budget. They do not have the drivers. That flips the story. If the model is not the moat and the model is not the blocker, the bottleneck in the entire AI economy is human. It is a teaching problem. A change management problem. A get-accounts-payable-to-trust-the-tool problem. That is not a San Francisco problem. That is an on-the-ground, local, human problem. And being close to the ground is one of the few things we are genuinely good at. Four. Sell shovels, not a prayer The AI training and enablement slice of corporate learning is somewhere in the six to seven billion dollar range today, heading toward something like 18 by the end of the decade. Estimates vary and I am not going to pretend to precision. Every serious forecast points the same way: up and to the right, fast. Here is why I care. Teaching people is a services business, and services get paid differently. No eight-year burn. No praying you hit an inflection before the money runs out. A company signs a purchase order, you deliver, you invoice, you get paid. Assuming a good payer. We have all met the other kind. This is the picks and shovels trade of the AI era. You can run at the hills and maybe strike gold. Or you can sell shovels to everybody running, and get paid whether they find gold or not. You do not need a frontier lab or a $200 million round. You need expertise, a curriculum, a customer and a purchase order. ***That is a business. Not a bet.*** Five. Three kinds of buyer, all moving at once Demand for this is not spread evenly around the world. It is concentrated here. Surveys have roughly 96 percent of Southeast Asian employers prioritising upskilling, against around 85 percent globally. Young workforce, fast adoption, hundreds of millions of people who came online mobile-first and now have AI in their pocket from day one. The instinct that let this region skip the desktop and skip the credit card is now pointed at AI. And this is not a prediction. Two of these moves landed this month while most people were not looking. Individuals. On 16 September, Singapore rolled out more than 200 AI courses under SkillsFuture, and citizens who take one get six months of a premium AI tool for free, so they can actually practise. A government putting shovels directly in people’s hands. Educators. ASEAN education officials endorsed AI Ready ASEAN, backed by Google.org and the ADB, moving from awareness into applied learning with a target of more than 1.7 million students and teachers by the end of 2028. That is not a webinar. That is a regional workforce being built on purpose. Governments. Here at home, Malaysia’s national AI skilling push with Microsoft, Microsoft Elevate, has already reached 80,000 learners from its pilot. The part I like is who it includes: educators, small business owners, civil servants, even retired servicemen and women. That is a whole-of-country bet that the bottleneck is people and the fix is teaching. A word to my friends in the Philippines, because this one is yours to win. The youngest population, the strongest English, and a workforce that already trains and supports the rest of the world for a living. You have done the BPO. You have done the data annotation. If any country here is built to become the place that teaches the region to use AI, it is you. Do not let anyone tell you that is a small ambition. Demand is not the question. Demand is the one thing we are not short of. The only question is who supplies it, and whether that supplier is one of us. Six. What it looks like as a company Take Gnowbe out of Singapore. AI-powered microlearning for enterprise. The client list is not hopeful startups. It is General Mills, Shangri-La, MetLife. Blue chips paying real money. The clever bit: they are SkillsFuture-accredited, so corporate clients can get up to 90 percent of the training cost co-funded by the government. Malaysia has its own version in HRD Corp. Employers pay into the levy whether they like it or not, and a lot of them genuinely struggle to spend it. There is cash sitting there. Smart operators build straight on top of it. What I respect is that this is a capital-light business. Nobody raised nine figures and set it on fire. It is a book of blue-chip customers, government co-funding pulling demand through the door, quietly compounding. That is what a healthy company looks like when the hype leaves the room. And it is not one company. A business in our portfolio has started running AI training in India and has landed some very large contracts. Someone in my network doing the same in the Middle East tells me they have more demand than they can possibly supply. Corporate bootcamps are running week-long intensives for executive teams. None of it shows up in that $4 billion number, because nobody is counting it. Which is exactly why there is still room. Seven. The plumber is now worth more than the engine builder There is a job title that barely existed three years ago and is now the hottest role in AI: the forward-deployed engineer. The idea is simple. Instead of selling a company a model and wishing them luck, you send an engineer to sit inside the customer and make the thing work in their actual mess. It is the Palantir playbook, now going mainstream, and at the top labs the people who make it work have become the people who close the deal. Why? Because the project dies on the unglamorous stuff. The AI cannot talk to the customer’s ancient database. It cannot get through their login system. It hits a data rule that says this information cannot leave the country. None of that is intelligence. All of it is plumbing. And the person who does the plumbing is now worth more than the person who built the engine. A one-week course for a bank’s back office and an engineer embedded for six months in a factory are the same business: get the thing to work for real people in a real building in this region. Teaching at one end, deployment at the other, and every dollar in between is a dollar for making AI useful, not inventing it. That entire line, from the classroom to the server room, is ours to own, because it is won on the ground, in local context and local language. Eight. Now take a wrench to the word “trained” I opened by taking apart “profitable.” Fair is fair, so let me do the same to “trained.” 1.7 million learners. 80,000 reached. 200 courses. Big, impressive numbers. And a lot of what sits underneath them is a two-hour awareness webinar and a certificate of attendance. Sitting through a webinar does not make you a skilled worker any more than watching a cooking video makes you a chef. When you hear a giant training number, ask the boring question. Trained to do what, exactly? And can they still do it the following Monday? Second, a lot of these programmes are funded by the American giants: Google.org, Microsoft, AWS. That is genuinely generous. It is also not charity. Free curriculum is a demand funnel. Teach a million people to build on your cloud and a chunk of them pay you cloud bills for the rest of their careers. The training is real, and it is also a customer acquisition cost. Hold both. Which is actually the whole opportunity. The free, generic, awareness-level stuff is covered. The giants have it. What they do not have, and probably do not care about, is the deep end. Sector-specific, applied training, taught in Bahasa, Tagalog, Thai or Vietnamese, for the way business actually gets done here. Teaching a bank’s compliance team, a hospital’s admin staff or a factory floor to use AI in their actual workflow, with their actual rules. ***Do not just hand out the shovels somebody else made. Own the part that teaches people to dig.*** Nine. If you wanted to build this on Monday Let me stop being the analyst and be the investor who would write the cheque. Here is how I would tell you to do it. 1\. Pick a lane you already know. Do not sell generic AI literacy, it is crowded and much of it is free. Sell AI for insurance underwriters, for property agents, for hospital admins. Narrow is where you can promise a real outcome and deliver it. 2\. Sell the outcome, not the attendance. The moment you can say “we cut this team’s reporting time in half, and here is the measurement,” you are not a trainer anymore. You are a consultant, and consultants charge ten times more. 3\. Stand on the government’s money. Get accredited under SkillsFuture, HRD Corp or the local equivalent, so the state picks up a big chunk of the bill. You are not competing with free. You are riding the subsidy that makes your deal an easy yes. 4\. Win one logo, then never let it go. Land one respected enterprise in your lane, do the work so well they put their name to it, and use that name to walk into the next ten. In services, one great reference customer beats any funding round. 5\. Stay capital-light on purpose. Customer cash funds the growth. That is the whole point of the model. Do not turn into a burn machine to impress somebody on a stage. It will probably not get you on the cover of anything. It will get you paid every month for solving a problem the entire region has at the same time. The bow The conference circuit is going to sell you 60 profitable startups and a $4 billion AI boom. One is a list with the definition deleted. The other is mostly one company. Neither is where the money is. Remember the 40 people who never learned to use the tools the boss bought them. That is not one company’s problem. That is the whole region, at once. Enterprise AI does not fail on the model. It fails on the people. So every worker, every boss and every government is trying to fix the same human problem at the same time, and they are all holding a cheque. We are probably not going to build the smartest model. We are not going to out-Anthropic Anthropic. I have made peace with that, and so should you. If one of the crazy ones out there can actually do it, I would love to find and back them. But we are going to teach 680 million people how to use it. We are going to send an invoice for that teaching, and we are going to get paid. Followers who teach can beat innovators who burn. So here is my question for you. Look at your own team. How many of the people you bought AI seats for are actually using them? And if the answer embarrasses you, who is going to fix it: you, a free webinar, or a local founder who just found their business? This piece accompanies this week’s episode of SEA of Startups. Real. Raw. Relatable. Listen on Spotify, Apple Podcasts, or YouTube, and subscribe to the newsletter at seaofstartups.substack.com. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit seaofstartups.substack.com