# He Asked, "Could It Be Mid-Next-Year?" — Notes on Statementdog on What AI Is Doing to Commercial Property Source: Realpha Blog (blog.getrealpha.com) Original article and charts: https://blog.getrealpha.com/en/blog/caibaogou-2026-11-15-578-ai-ft/ > Listening notes on Statementdog podcast episode 578: Colliers director Huang Shu-wei says the county attracting the largest property investment in Taiwan this year is Miaoli, where Micron bought a fab plus land in one US$1.8 billion deal, while NVIDIA spent NT$12.2 billion on a site in Taipei. What these buyers are paying for is time to production, and the premise behind the whole trend is that the geopolitical structure holds. Educational discussion, not investment advice; no stock recommendations or price targets. Published: 2026-11-17 Locale: en Tags: Statementdog, commercial property, industrial real estate, AI supply chain, Kaohsiung TL;DR: AI supply chain buyers take existing fabs because building from raw land means environmental review, queueing for water and power, and cleaning the soil — two or three years gone. Part of what they pay is for time to production, and run the same logic backwards: an idled fab took five or six years to sell. ![Dusk inside an industrial building, the empty long-span shop floor in the foreground sunk in shadow beneath overhead crane rails, the frame receding into depth, the loading dock doors at the far end thrown open with low sun coming through, transmission towers faintly visible beyond, a surveyor with a tablet standing small in the light with their back to the camera](/covers/caibaogou-2026-11-15-578-ai-ft-cover.png) > A mountain need not be high; one immortal makes it famous. Water need not be deep; one dragon makes it hallowed. This is a humble room, and the fragrance in it is my own virtue. > > —— Liu Yuxi, *Inscription for a Humble Room* (Tang dynasty; translated here by the author) Statementdog episode 578, released on 15 November 2026, has host Sky talking commercial property with Huang Shu-wei, a director in the occupier and landlord representation practice at Colliers. Huang says the county attracting the largest property investment in Taiwan this year is Miaoli: Micron bought a fab together with its land in a single US$1.8 billion deal, and the other headline transaction is NVIDIA spending NT$12.2 billion on a site in a Taipei technology park. His explanation is that part of what these buyers are paying for is time to production — building from raw land means environmental review, queueing for water and power, and remediating contaminated soil and groundwater, which takes two or three years. He is clear about the premise: the geopolitical structure holds, and Taiwan's core manufacturing capability stays here. ## When they used to come to Taiwan, they came to compare prices Huang tells a story from 2019. Before then, when a US semiconductor firm went site-hunting in Asia, the shortlist was Korea, Taiwan, Singapore and mainland China, and they toured all of them. Taiwan, he says, was usually along for the ride: on government support, land price, industrial ecosystem and degree of internationalisation, the only item where Taiwan led was having TSMC's supply chain here. So when a request like that came in, his side knew it was a pricing exercise, and the visitors left after the tour. The 2019 visit was different. The company opened by saying it would skip the other locations and come straight to Taiwan. Huang asked what they wanted to see; they said the science parks, and beyond Hsinchu they also wanted Taichung and Tainan. He assumed they meant fabs. They said fabs and offices — besides production, they wanted a training centre, an R&D centre and an operations centre, in the north, the centre and the south. What struck me here is how different that shape is from "find somewhere cheap to manufacture". A company willing to move R&D and operations in has changed the order of magnitude of how long it expects to stay. ## He asked, "could it be mid-next-year?" — and said renting was fine Huang then asked when they wanted to be in production. That question used to have a standard answer: buy land and build, two or three years at best. Their answer was: could it be the middle of next year? He said that was impossible. They said that was fine — renting worked too. That line is the key to this episode: what they wanted was production without a gap, and whether a building was new or old, bought or leased, ranked below time. Huang lays out how the two waves differ. The 2017–2018 wave was the opening of the US-China trade war, with Taiwanese firms pulling supply chains back from mainland China. The talk then was of five shortages, with water, power, labour and land all short, and the government package was tax relief and subsidised loans, levelling the cost so firms would come home. That wave also came with a rule: repatriated money could not go into property, only into productive equipment. The wave since 2022 has nobody subsidising it. Large firms are moving in, competing for orders and for ramp dates, and whoever reaches volume first takes the order, so the money comes from the buyer. Huang's point is that an existing building that shortens the path to production is what matches a firm's competitive position. ![Two side-by-side panels contrast opposite money flows: on the left, government subsidies flow toward companies, with a gate below blocking entry into real estate; on the right, companies fund themselves to flow into Taiwan, driven by the race to start production, reversing the direction of the two waves of money.](/figures/two-waves-money-direction-en.svg) How much time does building from land cost? He lists it: environmental review, applying for power (and power is short), applying for water, then soil and groundwater contamination — a regulated site cannot change hands until remediation is finished. The episode cites the Kaohsiung land TSMC bought, formerly an old industrial site: the soil has to be washed, then tested, and only a passing test lets construction start. Two or three years, minimum. So part of the price of an existing building is "already cleared". He calls power stability and the water, drainage and effluent systems the venous system; the visible items are ceiling height, floor loading, crane capacity, floor plate size and loading docks. Buy the wrong apartment and you buy another one — leaks, radioactive rebar and sea-sand concrete all appear in the disclosure statement, and there are public databases for comparing price. For industrial buildings all of that has to be contracted out for inspection, and he says a single deal takes them anywhere from one or two to three or five years. ![Two timeline bars side by side: the upper self-build path must first clear environmental review, power and water applications, and soil remediation, roughly thirty months before ground is even broken; the lower ready-made facility bar starts at production itself, and the gap between the two is exactly the time the price buys.](/figures/build-vs-ready-timeline-en.svg) ## Which is why the biggest number this year sits in Miaoli Huang likes asking audiences which county drew the largest property investment in Taiwan this year. Five guesses. Sky tried Taichung, Tainan, Chiayi and Pingtung, then reached for something obscure, and still missed. The answer is Miaoli — Micron's fab-plus-land purchase in Zhunan, at US$1.8 billion. He says the structure would have been unthinkable before: a foreign company buying land in Taiwan, and buying the building together with the land. When Google, Apple and Amazon set up plants, data centres or offices here, they leased; nobody was writing cheques to own. This year NVIDIA spent NT$12.2 billion on a site in a Taipei technology park. The buyer list reaches past semiconductors. He mentions ASE buying nine industrial properties in Taiwan this year, and Largan buying seven inside a single quarter. Server assembly needs something else again — floor space where people fasten screws and plug in cables, so it follows labour, landing in Taoyuan and Taichung. ## The same plant took five or six years to sell Run the logic backwards and it eats value. Huang describes another deal: a panel fab in a decent location, with reasonably new land and buildings, but a process a few generations old, which makes the equipment close to scrap. The company's finances gave out and the plant stopped. After it stopped nobody was managing it, so toxic gases may have escaped; water and power were cut for a long stretch, and equipment was stolen. Property like that ends in liquidation, and because the bank had lent heavily against it, it wanted an auction price that limited the loss, which kept the reserve high. He says it only sold recently, well below that original reserve, after five or six years: first the equipment went out in several packages handled by specialist contractors, then the interior environment was cleaned, and the property was disposed of last. The winning bidder was an environmental remediation firm with soil and groundwater capability — the only kind of buyer that can absorb something like this. My own read of that stretch is that a building is worth what its distance from "able to start production" says it is worth. For that panel fab the distance was five or six years, and the price was discounted to match. ![A slope running from upper left to lower right, with time until you can start work on the horizontal axis and price on the vertical axis: a ready-to-run plant with zero waiting sits at the high end, while a panel fab idled for five or six years lands in the deeply discounted lower-right corner.](/figures/distance-to-startup-vs-price-en.svg) ## The people have not arrived; the price has Then there is Kaohsiung. Huang's figures: TSMC's capital expenditure in Kaohsiung runs to at least NT$1.5 trillion, and the whole supply chain at least NT$4 trillion. TSMC has five fabs planned in Kaohsiung, P1 through P5, all to be in production within five years, with the first already running since last year; ASE has at least three campuses there, with another new one coming at Qiaotou. On the population gap he is blunt: Kaohsiung has been a place people leave, and the number of engineers advanced nodes require bears no relation to who is there now. He puts the sequence in order: industry arrives, then employment, then public facilities, and population last. Hsinchu is the control case — forty years of development and public investment is still short, so a child finishing junior high has only a couple of local high schools and many families leave. Ten years ago most of the companies were already there, and prices were far below today's, which means somebody still had to fire the starting gun. ![A four-step staircase rises from left to right: industry arrives, workers arrive, public infrastructure arrives, and people arrive; yet the price arrow jumps to the top at the first step, with the handover date in the middle and a gap between them.](/figures/price-arrives-before-people-en.svg) In Kaohsiung the sign to watch is schools: Kang Chiao is opening a campus, primary schools near Qiaotou are adding classes, the experimental high school is adding a campus, while other counties are cutting classes. So the demand looks real. The problem is the lag in price. TSMC announced the Kaohsiung investment at the end of 2022, developers nearby started presales that same year, and prices went through two rounds of increases before the fab was finished. Now the people are arriving and the units are completing — and the districts with the most pre-completion cancellations are the ones with the most technology plants: Tainan, Kaohsiung, Taichung. ## The story is intact, so why am I the one getting liquidated This is the question I most wanted to take apart, because it has the same shape in equities: the thesis has not been refuted, the industry is still investing, and the position blows up first. Huang's path runs like this. Presales sold best in 2023 and 2024, when the youth mortgage scheme let buyers of completed homes get financing, so developers pushed low deposits alongside — a few hundred thousand NT dollars to sign, settle it at completion. Then the selective credit controls of the third quarter of 2024 capped loan-to-value on a second home and widened the scope from the seven special municipalities to the whole country, so anyone already holding one unit, whether inherited or mortgaged, was restricted on the next. On the other side, the Equalization of Land Rights Act stopped presale contracts from being assigned, which killed the resale of reservation slips. That left two roads: fail to get the loan and default, or complete and deal with it afterwards. What crushes the timeline is the completion date. He mentions a project originally promised for 2028, which investors reckoned they could stall past the presidential election in case rules loosened — and then the developer announced completion in 2027. The bank came round to sign mortgage documents six months ahead, and the cancellations began. Every step on that path is a cash-flow timing mismatch — loan-to-value, no contract assignment, the completion date — three reasonable rules that stack into forced selling. I now ask one more question about an equity position: what is this one's completion date? Margin interest, an option expiry, mortgage pressure, or some day I know I will need the money. ![Three arrows press inward from the top, left and right toward the center, squeezing what was once a wide exit into a narrow slit, and the only paths marked at that opening are cancelling the contract or selling at a loss.](/figures/three-constraints-squeeze-en.svg) He mentions something you can check yourself: the government's actual-price registry now has a cancellation field you can open, with cancellations by district, and the stated reasons often read as relatives, related parties, or a NT$2 million decoration or appliance allowance. Prices propped up by those transactions look fake with hindsight; had the controls never landed and the units completed smoothly, the same unit would have become an asset on a balance sheet awaiting disposal. The other layer is who is selling. He says the firm the industry fears is an unlisted developer with no funding needs and branch companies in the hundreds: when it cuts prices it clears inventory fastest, and everyone else is stuck, because nobody matches its cost. Its projects are the benchmark ones — the most expensive luxury homes and the most expensive office-industrial buildings are its — so when it cuts a tier, the neighbours follow. The equity analogue is that the most liquid name gets sold first, and its price becomes the anchor for the whole group. ## I added electricity to my own site checklist The other sore point: you hear an industry story, and how do you know which condition to go and check? This episode's answer is power. Huang says data centres have become the kind of facility nobody wants nearby — people want them in Changbin, in Yunlin, as far away as possible, and nobody would back one next to Songshan Airport. He also mentions a site in Taipei where a tower has been delayed again and again because the substation beside it has been protested for years and cannot be built. They laugh about old Chinese advertisements for industrial buildings, with a row at the bottom listing water, power and road connections, which read like a rural pitch at the time; today's Taiwanese listings carry the same list. The way I noted this down: the bottleneck in this layer of AI right now is power and space you can start work in, so some site choices that look strange make sense once you fill power in — fewer offices and industrial buildings competing for the supply there, and plots large enough. Huang uses the same line to explain why some Taiwanese firms are no longer investing here: utilities are the bottleneck, while Texas has cheap power and cheap land, and when the leader goes the suppliers follow. ![Two concentric circles show a dense urban core inside, where a substation cannot be built after years of protests, and an outer ring with ample land and no competition for power; an arrow pushes the data center from the inner ring to the outer ring.](/figures/power-pushes-siting-outward-en.svg) In the US the binding condition is holding cost. He says Phoenix home prices have risen fifty to seventy per cent over a few years, while plenty of owners now cannot sell: property tax starts around one per cent of assessed value every year, mortgage rates run 7.4% to 7.5%, there are management fees, and rental yield has to cover all of it — which makes that market look more like a rental market. As for buyers who can pay all cash, he asks back: why would they not put it in equities? ## Where to look yourself - Statementdog episode 578, "Industry talk: the property market's Vin Diesel returns — AI is minting money in commercial real estate, ft. Colliers director Huang Shu-wei", 15 November 2026 - The central bank's press releases on each round of selective credit controls, for how second-home loan-to-value caps and their scope changed - *Apple in China*, the book Huang recommends at the end of the episode, on how Apple spent thirty years building a manufacturing fortress in mainland China ## The One Thing to Take Away One idea: a lot of price differences come down to how far something is from "can start". The panel fab that took five or six years to auction had reasonably new land and buildings; the discount came from the years still standing between it and anyone working inside. Of the US$1.8 billion Micron paid in Zhunan, part of it bought away two or three years of environmental review, utility applications and soil remediation. The same thing holds for people — the task you will pay someone else to do is usually the one you do not want to wait for. Something I have tried that works outside investing too: pick one thing you have been putting off — renewing a passport, the dentist, the car — and write down two numbers. First, how many days from today until it is done. Second, if you needed it tomorrow, how much you would pay to buy that waiting away. Put the two on the same line and you will see your own price for time. I have written it out a few times, and the number I would pay came out higher than I expected — while all those "already done, ready to use" options had only ever looked expensive to me. ![A horizontal waiting axis runs from today to done, labeled 90 days; at its start stands a short bar marked with what you would pay, and the two numbers side by side form your price for that stretch of waiting.](/figures/pricing-your-own-waiting-en.svg)