Singapore wants to be the place startups go global from. Enterprise Singapore pitches Startup SG as "your launchpad to the region and beyond," and Budget 2026 set aside S$1 billion to top up Startup SG Equity for Singapore-based deep tech startups, now including at the growth stage. I share the objective. I think most founders reach for it in the wrong order. If you can start in the US, start there, and then build your Asia-Pacific headquarters in Singapore. That gets Singapore what it wants more often, because the power law does not work well for companies that begin here.

Venture returns come from a handful of companies that get very large, and how large a company can get is shaped early by where it starts. Your first customers decide what you learn and how fast. Your first prices set what every later customer expects to pay. A company that starts in a city-state spends its formative years learning from a small set of buyers and pricing for them, then expands into Southeast Asia, which is several markets with their own rules, languages and price points. By the time it reaches the US, if it ever does, its product and its price have been shaped by the wrong customers.

The capital stack points the same way. In 2025, about $274 billion of startup funding went to US-based companies, 64 percent of the global total by Crunchbase's count. All of Southeast Asia raised $5.4 billion across 461 deals by DealStreetAsia's tally, and Singapore took 91 percent of it. The big growth checks and the acquirers who pay up sit on the American side. A Singapore company with the same numbers as a US peer gets priced as a regional story, and that discount compounds every round.

AI widens the gap. The labs, the design partners willing to run something half-built, and the engineers who have shipped on frontier models are concentrated in the Bay Area, and an AI company lives on iteration speed. Being an ocean away from your best customers is a tax you pay every week. Fortune reported in July on a founder building an AI company in Singapore who moved to the Bay Area after noticing his early customers were already based in, or selling to, the US. The power law made the decision for him, late.

Now look at what happens when a company that has already won abroad comes to Asia. Carta was primarily US-focused before it chose Singapore for its first international expansion and regional headquarters, and that team has grown from about fifteen people to over sixty. Ambiq, a US-headquartered chipmaker, runs its regional R&D and commercial hub here and opened a new Singapore office in February, after its July 2025 IPO. Senior jobs, regional revenue, a global brand, with Singapore running Asia. I would argue it achieves "go global from Singapore" better than ten Singapore-headquartered companies that stall in the region.

The state's own capital already behaves this way. EDBI, which now sits under the state's SG Growth Capital platform, says it invests in "high-growth global technology companies and industry leaders looking to scale in Singapore and across Asia." Its write-up on Ambiq, which it backed in a 2018 Series E, describes "its mandate to anchor global innovators." EDBI is not betting that Singapore must be the birthplace, only the anchor. The startup end of policy should take the hint.

There are real exceptions. If your market is Southeast Asia, start here. Grab, which began in Kuala Lumpur, and Sea built for the region, and local knowledge was the moat. Runway is cheaper here, US visas are a hard constraint for many founders, and Singapore's grants and co-investment are real money. This is not an argument that Singapore is second-rate. It is about sequence, for founders who can get to the US and whose customer is global.

If that is you, incorporate and sell where your first great customer lives, and plan the Singapore headquarters as a milestone you reach with revenue, not a fallback you start from. If you are a GP, underwrite where the first customer and the next round come from, not the founder's passport. If you set policy, count the global companies running Asia from Singapore as wins, and let support follow the founder, so a Singaporean who starts in San Francisco has every reason to bring the regional headquarters home. The region is not a discount to SF. It is not a substitute for it either. Singapore wins at the second office.

The most sophisticated investors on both sides of the Pacific told us this week what they value most right now, and it is not the next model. It is the ability to change their minds. They are not retreating from AI. They are paying to keep the door open.

On October 7, at the Milken Asia Summit in Singapore, Temasek's chief investment officer Rohit Sipahimalani said about half of Temasek's AI exposure sits in public markets, and "we want to increase that number ideally to 70% to 75% so it gives us more flexibility to pivot with changes in the environment, as against being stuck with a private asset which there's little you can do when things are changing." He called an unwinding of the AI trade "the biggest risk," said Temasek does not see it as imminent, and allowed that there could be bumps in 2027. This is not a retreat. In July, Temasek said it aims to more than double AI from 6 percent of its portfolio to up to 15 percent by 2031. What is changing is the mix: more AI, held in a form it can sell.

The same week, the PitchBook-NVCA Venture Monitor showed American LPs making the same trade with different instruments. Established firms, those on fund four or later, took 88.2 percent of US venture capital raised this year, the highest share in PitchBook's data, and PitchBook says that preference "largely reflects liquidity constraints." The number of emerging firms closing a fund fell from 927 in 2022 to 211 through the third quarter. First-time funds raised $4.9 billion across 81 vehicles. The top line looks like a boom: a record $515.8 billion of deal value with AI at 82.7 percent of it. PitchBook's Nizar Tarhuni was blunt about what sits underneath: "for most of the market, the liquidity won't show up." On Forge, shares of companies that last raised in 2021 trade at a median 59 percent discount.

Put these together and the story is not that LPs have soured on AI. They are buying optionality wherever they can find it. Temasek finds it in listed stocks. US LPs find it by recommitting to managers who have already sent cash back, on the theory that a firm that distributed once can do it again. The pattern runs beyond venture, too. In private equity, the most patient institutions are selling to get it: Yale was reported in September to be selling about $1 billion of fund stakes, and HarbourVest was reported on October 1 to be buying roughly $1 billion of fund interests from GIC.

I think this is rational, not panic. When the technology moves this fast, the cost of being locked in rises. A private position you cannot exit is a bet that the world you underwrote will still exist when the position matures, and in AI that world can change inside a single fund cycle. So the premium for holding illiquid private assets has to be higher than it used to be. Most managers cannot show that it is. That is the uncomfortable conclusion inside the 88 percent. Capital is not being allocated by manager quality. It is being allocated by liquidity, and those are not the same ranking.

That should change how everyone on the chain behaves. Founders should drop the assumption that their investors are happy to hold for ten years. Build a cap table that is ready for secondaries, and have a tender-offer plan before someone asks for one, because early holders who cannot get out will eventually act like it. GPs should treat a credible path to returning cash early, through continuation funds or structured secondaries, as part of the product, and the LPs underwriting them should pass on funds whose only exit is an IPO window that, in Tarhuni's telling, keeps slipping further out.

In the US, LPs should be honest that an 88 percent allocation to established managers is a liquidity bet, not alpha selection. Emerging-manager terms are at their best precisely when that cohort is starved, so a disciplined small allocation now is worth protecting. In Southeast Asia the lesson is sharper. Temasek is the allocator this region watches most closely, and if it wants the freedom to pivot, the funds that grew up in its ecosystem should build the exit door into the fund design: secondary programs, explicit DPI targets, continuation options. Family offices committing to regional funds should ask to see that plan before they sign.

Everyone is paying for the exit door. The managers who build one will be the ones who get paid.

Who kept the equity is the market story. How they were trained is the founder story.

Forbes' real-time billionaires list this week, late September 2026: Elon Musk, Jeff Bezos, Larry Page, Michael Dell, Sergey Brin, Mark Zuckerberg, Jensen Huang, Larry Ellison, Steve Ballmer, Warren Buffett.

Nine of ten studied a STEM field at some point, finished or not. Buffett is the holdout. Business and economics, around tenth. Amancio Ortega, no university degree, sits just outside at eleven.

A decade ago the annual list was Gates, Ortega, Buffett, Carlos Slim, Bezos, Zuckerberg, Ellison, Michael Bloomberg, Charles Koch, David Koch. Count honestly and STEM was already about eight of ten. The move from eight to nine is not a curriculum revolution. The people who own the compounding machines still came up speaking engineering.

What changed is which balance sheets compound, and who still owns them.

Software, cloud, chips, and AI created equity value a department store, a phone network, or a conglomerate could not match. The names that pushed Gates, Slim, Bloomberg, and the Kochs out of the top ten are mostly people who kept large stakes in those machines. Page, Brin, and Huang were not in the 2016 top ten. Musk went to the front. Ballmer never founded Microsoft, but he got paid in stock and never sold enough of it. Buffett did not get poorer. He got outpaced, then slipped back in when the tape moved.

That is the market half of the story. Leftover ownership in a business that scales without opening another store.

The founder half is less fashionable and more useful. The people who built those machines usually started in physics, CS, electrical engineering, applied math, biology. Not because a diploma is IP, but because STEM training teaches you to model systems, break them, and rebuild them under constraint. That is the same muscle you need to ship a product, read a cost curve, and know when the demo is lying.

I have watched founders treat the major as a costume for the deck, and I have watched other founders use it as the way they think. Only the second group keeps compounding when the narrative turns. Get more STEM in your education, not less. Learn to model systems before you learn to model a pitch. Keep ownership in the thing that compounds. The degree is not the moat. The training is how you build one.

Forbes moves with the tape. A bad week for Tesla or Nvidia reshuffles the order. Ortega, Bernard Arnault, and Buffett trade the tenth slot depending on the day. Underneath, the pattern holds: the top of the list is leftover equity in software and semiconductors, held by people who mostly learned to think like engineers first.

If you are underwriting those founders, look for the ones who can reason about the product in engineering terms and still own enough of the upside. Pass on STEM as a brand on slide two with no systems fluency underneath.

The list did not get smarter because of business school. It got richer because engineers kept the equity.

Most investors are still asking the wrong question about defensibility. They ask whether a competitor could rebuild the product. In 2026, the more important question is whether the customer could.

The test changed while we were not looking.

Retool surveyed 817 builders this year and found that 35% of enterprises have already replaced at least one SaaS product with something they built internally. Seventy-eight percent plan to build more. They are replacing workflow automation, admin tools, BI, CRM, project management and support software. These are not bad categories. These are categories venture capital has funded for fifteen years. McKinsey independently puts the build-instead-of-buy number at 32%.

The buyer did not suddenly become a software company. The buyer got a coding agent.

So when a founder tells me the moat is feature depth, integrations or UX, I now apply a simple customer test. If your largest customer decided to clone the product next quarter, would their version be materially worse? If the honest answer is no, you do not have a moat. You have a head start.

What can a customer not build for itself?

I keep coming back to four things.

The first is accountability. When an AI agent acts autonomously inside a customer's systems, someone has to be responsible when it gets something wrong. An internal tool gives you nobody outside the company to hold accountable. In high-consequence workflows, the startup is not just selling software. It is selling the transfer of risk.

The second is data the customer cannot create alone. Not its own data sitting inside your product. I mean data that only exists because you sit across many customers, markets or counterparties. Benchmarks. Fraud signals. Pricing intelligence. A customer can reproduce what it teaches you. It cannot reproduce what everyone else teaches you.

The third is owning the record, not just the interface. AI is making interfaces incredibly cheap to build. But systems holding state and history remain painful to remove because removing them means reconciling records and figuring out what breaks downstream.

The fourth is multi-party position. If you sit between parties that do not want to integrate directly, a customer cannot rebuild your product because it cannot rebuild the other side of your network.

Put these together and the objective has changed.

Stop trying to be hard to copy. Start trying to be hard to remove.

Removal is an organizational problem, not a technical one. I care less about whether rebuilding your software takes six months or six days. I care about how many people must approve removing it, what breaks when you leave and who gets blamed when the migration goes wrong.

This changes how I do diligence.

I want to see the MSA. Who is responsible? Is there an indemnity? What is the liability cap? A contract with real teeth tells me the customer is buying more than functionality. It is transferring responsibility.

I ask whether customers have discussed rebuilding the product internally. I ask why churned customers left and how many rebuilt themselves. I ask how many departments use the product. And I ask the simplest question: what happens when your software is wrong?

There is an uncomfortable implication here.

If defensibility increasingly comes from accountability, trust and relationships, then part of the AI moat is human. That sits awkwardly with our current obsession with tiny AI-native companies.

A company may be able to build its product with almost no employees. It cannot necessarily hold the customer relationship, absorb liability and earn the right to act autonomously on a customer's behalf with almost no employees.

There is an accountability floor below which companies cannot shrink.

We are about to find out where it is.

So stop asking whether a competitor can copy the startup.

Ask whether the customer can.

Then ask who is on the hook when the software gets something important wrong.

Increasingly, I think that person is the moat.