> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # 43 Substations in 15 Years: Who Gets Paid as Taiwan's Grid Chases AI Halls - URL: https://www.sotkn.com/markets/taiwan/taiwan-taipower-ai-data-centre-grid-october-2026/ - Published: 2026-10-14T10:50:19.000Z - Updated: 2026-10-14T10:50:18.000Z - Description: On 14 October Taipower said it will build or expand 43 substations over 15 years to feed AI and chip loads. It wants big AI halls next to power plants. Substations and lines are paid first. The stuck step is getting power to the site. - Author: Sean - Tags: #market-taiwan, Markets *Taiwan's chip fabs already eat a large share of the island's power. On 14 October the state utility said the next bill is the wires. Taipower plans 43 new or expanded substations over 15 years, about three a year, and wants big AI data centres built next to power plants. Substations and lines are paid first. Halls wait until the plug arrives.* **TAIWAN PULSE** · AI DATA CENTERS · Data as at 14 Oct 2026 · Next update: any Taipower filing that names the first of the 43 substations **Affects:** [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/) · [Semiconductors](https://www.sotkn.com/stack/semiconductors/) · [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/) — Markets: [Taiwan](https://www.sotkn.com/markets/taiwan/) · [Korea](https://www.sotkn.com/markets/korea/) · [Australia](https://www.sotkn.com/markets/australia/) · [US](https://www.sotkn.com/markets/us/) **In short:** On 14 October 2026 Taipower chairman Tseng Wen-sheng said the utility will build or expand 43 substations over 15 years, add about 1,922 km of transmission lines and about 52,450 MVA of transformer capacity, and spend about NT$30 billion on 12 STATCOM units. Semiconductor electricity use is seen rising from about 52 billion kWh this year to about 110 billion kWh by 2035\. He wants large AI data centres next to power plants, and said a 100 MW AI hall can cost about NT$1.5 trillion to build. Substations and lines are paid first. The stuck step is delivering power to the site. Map of who gets paid — not a buy list. Education only, not advice to buy or sell any security. Figures are as reported from Taipower's 14 October remarks and related 13 October comments. A planned megawatt is not a running one. Our own sums are labelled. ### The numbers that matter this week Substations, 15-year plan 43 About 3 a year · 14 Oct speech Primary New transmission lines 1,922 km To move power to the load Primary New transformer capacity 52,450 MVA Framed as \~52 GW · TVBS / CNA Primary STATCOM spend NT$30B 12 units for grid stability Primary Semiconductor kWh path 52→110B 2026 to 2035 · \~30% of national use Primary Power delivery to the site Stuck Halls wait on substations and lines Stuck step Figures from Taipower chairman Tseng Wen-sheng on 14 October 2026, as reported by CNA and TVBS, and from 13 October remarks as reported by the Taipei Times and Taiwan News. Status chips are Second Order’s read. **On this page:** [Why it matters](#why-it-matters) · [What changed](#what-changed) · [The 43-substation plan](#the-plan) · [Why near the plant](#near-plant) · [Our math](#our-math) · [Who gets paid](#money-goes) · [What is stuck?](#still-stuck) · [Dates next](#dates) · [What would change this](#change-view) · [FAQ](#faq) ## Why does Taiwan's grid plan matter for AI data centres? Because Taiwan already builds the chips that fill AI halls, and the next bottleneck is not only packaging or memory. It is whether electricity can reach a dense AI site on time. For what those halls are, see [what an AI data center is](https://www.sotkn.com/explained/what-is-an-ai-data-center/). For the layer map of landlords, power and hosts, see [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/). Taiwan's Markets page has tracked exports, Foxconn racks and packaging land this month. Those stories sit upstream of the hall. This Pulse looks at the utility that has to deliver the megawatts once someone wants to run racks on the island, not only ship them abroad. The same bind shows up elsewhere this week. [Korea's Pohang financing Pulse](https://www.sotkn.com/markets/korea/korea-pohang-ai-data-centre-financing-october-2026/) showed AI halls leaving Seoul for sites where power already exists. [New South Wales](https://www.sotkn.com/markets/australia/australia-nsw-data-centre-power-condition-october-2026/) now ties new Sydney approvals to firmed clean power before a hall can run. [PJM](https://www.sotkn.com/markets/us/us-pjm-capacity-shortfall-october-2026/) is short of capacity and still rewriting who pays for new plants. Taiwan's answer on 14 October was explicit: build substations at a pace the island has not tried before, and put large AI halls next to plants. 1 · Chip and AI load rises **Who:** Fabs, memory, burn-in, AI halls **Paid:** 52B → 110B kWh semiconductor path Demand 2 · Substations and lines get built **Who:** Taipower and its contractors **Paid:** 43 sites · 1,922 km · 52,450 MVA Paid first 3 · Halls next to plants connect **Who:** Hyperscalers and local hosts **Paid:** Shorter dedicated lines Paid next 4 · Capacity squatters wait **Who:** Applicants without a real build **Paid:** 100 MW example: \~NT$1.5T to build Deferred ## What changed on 13 and 14 October 2026? - **13 October — shorter clocks.** Ahead of the Net-Zero Taiwan and Energy Taiwan shows, Tseng said electricity investment decisions that used to take about ten years now need to be made in five years, or even three. He said AI demand could rise fast and is hard to forecast if companies keep computing on-site to protect secrets. - **14 October — the 43-substation plan.** Speaking at the smart-energy week, Tseng set out a 15-year grid build: 43 substations, about 1,922 km of new lines, about 52,450 MVA of new transformer capacity, and about NT$30 billion for 12 STATCOM units. - **14 October — site next to the plant.** In follow-up answers, he said large AI data centres should sit near power plants so dedicated lines are shorter, cheaper and faster. Firms that really intend to invest, he said, accept that. - **14 October — anti-squatter filter.** On a 100 MW AI data centre application, he estimated a real build could cost about NT$1.5 trillion. Giving the power slot to someone who only holds land and a possibility, he said, can block a real investor. Taipower is still designing a fairer filter. - **7 October — Mailiao LOI (context).** Formosa Smart Energy signed a letter of intent with CXIT for a first-phase MW-scale AI data centre at the Mailiao industrial park, with a planned cloud joint venture later. No megawatt or dollar figure was disclosed. It is a direction, not a build. Read together, the week is about delivery, not a single lease. Taiwan's [September export record](https://www.sotkn.com/markets/taiwan/taiwan-september-exports-record-october-2026/) and [Foxconn AI rack Pulse](https://www.sotkn.com/markets/taiwan/taiwan-foxconn-ai-racks-october-2026/) show the hardware side is already moving. The utility is now saying the wires have to move with it. ## What is inside the 43-substation plan? ![Chart of Taipower's 15-year grid build for AI and chip loads: 43 substations, 1,922 km of lines, 52,450 MVA of transformers](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/taiwan-taipower-ai-data-centre-grid-october-2026-print.png) Wires first. Halls wait on the plug. · Education only. Tseng's point was that counting megawatts of generation is not enough. Power also has to arrive where the load sits. Taiwan's land is tight, so each science park that used to share one ultra-high-voltage substation now needs more. Southern Taiwan Science Park already has more than one, and will need a third. Taipower expects that park's load to reach about 4 GW by 2031, about 9% of national electricity use. Grid build, as stated 14 Oct Substations43 new or expanded over 15 years PaceAbout 3 a year New linesAbout 1,922 km Transformer capacityAbout 52,450 MVA (\~52 GW framed) STATCOM12 units · about NT$30B Load path Taipower cited Semiconductor use, 2026About 52 billion kWh Semiconductor use, 2035About 110 billion kWh Share of national use, 2035About 30% Southern Taiwan Science Park, 2031About 4 GW load Gas units, next 10 years25 units · about 26 GW He broke the semiconductor load into parts. Wafer fabs are the part Taipower can forecast from known process nodes, especially investments below 3 nanometres. Memory, packaging and testing, and high-end servers together use about 30% as much as wafer making today, and that share may rise. Burn-in testing of memory and servers is named as a growth source. That is the bridge from chip output to AI halls on the island. Cities are in the plan too. Alongside new industrial clusters, Taipower says it will refresh ageing underground lines in 18 districts across the six special municipalities. New industrial load and city load rise together; one cannot wait for the other. ![Chart of where Taiwan's power bind sits: semiconductor kWh growth versus the substations and lines that must catch up](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/taiwan-taipower-ai-data-centre-grid-october-2026-shift.png) Chip load rises; the substation clock is slower. · Education only. ## Why put the AI hall next to the power plant? Because a long dedicated line is slow and expensive in a dense island. Tseng said large AI data centres should be sited close to generation so the power travel distance shrinks. That cuts build time and cost for the dedicated feed. In his TVBS interview he said the hyperscalers and other firms Taipower has spoken to privately are willing, as long as they really intend to invest. He used a small physical picture. A single cabinet of GB200 or NVL72-class AI servers can draw about 120 to 140 kilowatts. About 20 such cabinets, he said, use as much electricity as Taipower's own headquarters building. AI halls run for long hours and concentrate load on a local grid. Putting that concentration next to a plant is a planning choice, not a slogan. The Mailiao letter of intent fits the same geography even though it is not a Taipower project. Formosa Smart Energy and CXIT chose an industrial park that already has power and utility systems. The announcement did not say which megawatts, which power source or which customers. It does show that builders are looking for sites where power and land already exist, the same logic Korea is using outside Seoul. For how long each stage of a hall takes once power is agreed, see [how long it takes to build an AI data center](https://www.sotkn.com/explained/how-long-does-it-take-to-build-an-ai-data-center/). For why utilities get paid before the hall earns, see [where cash sticks on the utility bill](https://www.sotkn.com/analysis/utilities-bill-data-centers-before-power-where-cash-sticks/). ## Our math: what do the headlines imply? - **Substation pace.** 43 substations over 15 years is about 2.9 a year (our arithmetic). Tseng rounded it to three a year and called that unprecedented for Taipower. - **Transformer capacity versus the semiconductor path.** About 52,450 MVA of new transformers is a stock of gear, not a yearly consumption figure. The semiconductor path from 52 to 110 billion kWh is a flow. They are not the same unit. What the plan says is that delivery kit has to scale with industrial load, not that every new MVA is reserved for AI halls (our reading). - **100 MW cost print.** If a 100 MW AI data centre costs about NT$1.5 trillion to build, that is about NT$15 billion per MW of IT load (our arithmetic). Treat it as Taipower's order-of-magnitude filter for who is a real applicant, not as a price quote for any named project. - **Cabinet math.** At 120 to 140 kW per cabinet, 20 cabinets are 2.4 to 2.8 MW (our arithmetic). A 100 MW hall is on the order of roughly 700 to 800 such cabinets if that density held (our arithmetic). Real designs differ; the point of the example is density, not a bill of materials. - **Southern Taiwan Science Park.** About 4 GW of load by 2031 at one park is about 40 times a 100 MW AI hall (our arithmetic). Chip fabs dominate that number; AI halls are an extra concentrated add-on that still needs its own feed. What this tells you: the money for substations and lines is being planned in public. The money for any one AI hall still depends on a real investor clearing Taipower's filter. For how to read cash spent long before it earns, see [cash capex, explained](https://www.sotkn.com/explained/cash-capex-explained/). ## Who gets paid first? ![Infographic: who gets paid as Taiwan's grid chases AI halls, and why delivery is the stuck step](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/taiwan-taipower-ai-data-centre-grid-october-2026-infographic.png) Who gets paid / stuck step on Taiwan's AI grid. · Education only. Follow one hyperscaler that wants a 100 MW AI hall in Taiwan next year. The money moves in a short line. - **First, the grid build.** Taipower and its contractors get paid as substations, lines and STATCOM units go in. That spend is in the 15-year plan whether or not any one AI applicant lands. Without it, no dense hall connects. - **Second, site owners next to plants.** Industrial parks and plant-adjacent land that can take a dedicated feed become more valuable because they clear the siting preference Tseng described. Mailiao-style parks are an example of the geography, not a paid lease. - **Third, hall builders and equipment sellers.** They bill once a real project has power certainty and starts construction. Until then, drawings do not pay. - **Fourth, the tenant.** Cloud and model companies pay when racks switch on. That is the last cash in the chain. - **Who waits.** Applicants who only hold a site and a power enquiry. Tseng's NT$1.5 trillion print for 100 MW is meant to separate them from builders who will actually spend. One-line rule: who gets paid first is the grid build; the stuck step is delivering power to a real hall, not finding chips or customers. ## So what is the stuck step? Delivering power to the site. Not generation headlines alone, and not demand. Three things make it stick. - **Time.** Substations and lines take years in a dense island. Tseng is compressing decision clocks to three to five years, but steel and cable still have to be built. - **Place.** Taiwan cannot spray long dedicated lines across every coastal plot. Near-plant siting is the workaround. - **Queue honesty.** If power slots go to holders without a build, real investors wait. Taipower says it is still designing that filter. So-what: Taiwan's AI data centre map will follow the substations and the plants, not only the science-park brand names. That is the same second-order move as Korea's halls leaving Seoul and Australia's halls waiting on firmed power. For the general rule, see [where the stuck step moves](https://www.sotkn.com/analysis/where-does-the-stuck-step-move/). ## Which dates test this read next? ![Calendar of dates that test whether Taipower's grid plan keeps pace with AI halls](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/taiwan-taipower-ai-data-centre-grid-october-2026-calendar.png) Dates that test the read. · Education only. - **14 Oct 2026 —** Taipower's 43-substation speech and near-plant guidance. Done. - **Date not set —** First named project among the 43 substations, with a site and a service year. - **Date not set —** Publication of Taipower's filter for large AI data centre applications. - **Date not set —** Any Mailiao follow-up that discloses megawatts, power source or a customer. - **2031 —** Southern Taiwan Science Park load toward about 4 GW on Taipower's path. - **2035 —** Semiconductor electricity use toward about 110 billion kWh on the same path. The first named substation matters most. Until a site and a year land in public, the 43 remain a programme, not a queue you can map hall by hall. ## What would change this view? - **Taipower names the first substations with dates and served loads** — the programme becomes a build schedule. - **A hyperscaler announces a Taiwan AI hall with MW, site and connection year** — near-plant siting is tested in public. - **The application filter is published and used** — capacity squatting shrinks. - **Semiconductor or AI load rises much faster than the 110 billion kWh path** — even 43 substations look tight. - **Large halls are approved far from plants without dedicated feeds** — the near-plant preference did not stick. ## Practical takeaways - Taipower plans 43 substations in 15 years, about three a year, plus 1,922 km of lines and 52,450 MVA of transformers - Semiconductor electricity use is seen rising from about 52 to 110 billion kWh by 2035 - Large AI halls are steered toward sites next to power plants - Who gets paid first: the grid build; halls and tenants come later - Stuck step: delivering power to a real site, not finding chips ## FAQ **What did Taipower announce on 14 October 2026?** At Taiwan's smart-energy week, chairman Tseng Wen-sheng said Taipower plans to build or expand 43 substations over the next 15 years, about three a year, add about 1,922 km of transmission lines, and add about 52,450 MVA of transformer capacity (framed as about 52 GW). It will also spend about NT$30 billion on 12 STATCOM units to steady the grid. **Why does this matter for AI data centres?** AI halls draw dense, round-the-clock power. Taipower said about 20 cabinets of GB200/NVL72-class servers can use as much electricity as its own headquarters building. Generation is not enough if the wires cannot deliver it to the site. The stuck step is grid delivery. **Where does Taipower want AI halls to go?** Next to power plants where possible. Shorter dedicated lines are cheaper and faster to build. Tseng said firms that truly intend to invest are willing to site that way. **Who gets paid first?** The companies and contractors that build substations, lines and grid gear get paid as the wires go in. Hall builders and tenants get paid later, once power reaches a finished site. Capacity holders without a real build plan are the ones Taipower says it does not want to queue ahead of real investors. **Is this investment advice?** No. Education only. It maps who gets paid and which step is stuck. It is not a recommendation to buy or sell any security. ## Keep reading on the map - [Taiwan Markets hub](https://www.sotkn.com/markets/taiwan/) — country page - [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/) — landlords, MW and power - [Korea Pohang AI financing](https://www.sotkn.com/markets/korea/korea-pohang-ai-data-centre-financing-october-2026/) — halls move to the power - [Sydney power condition](https://www.sotkn.com/markets/australia/australia-nsw-data-centre-power-condition-october-2026/) — approval is not the plug - [PJM capacity shortfall](https://www.sotkn.com/markets/us/us-pjm-capacity-shortfall-october-2026/) — who pays for new supply - [Singapore data centre licence](https://www.sotkn.com/markets/singapore/singapore-data-centre-licence-october-2026/) — state-rationed MW - [What is an AI data center?](https://www.sotkn.com/explained/what-is-an-ai-data-center/) · [All Markets doors](https://www.sotkn.com/markets/) **Get the Wire.** Second Order’s free daily brief on what moved in the AI chip map. [Read the Wire →](https://www.sotkn.com/the-wire/) · [Subscribe free →](https://www.sotkn.com/#/portal/signup) **Members’ Analysis (teaser):** When the stuck step is the wire, cash sticks at the substation before it reaches the hall. Who carries the wait between a chip export boom and a powered rack on the island? The free line: substations are paid before AI halls earn. [Read the Analysis →](https://www.sotkn.com/analysis/where-does-the-stuck-step-move/) **How we check this.** The grid plan comes from Taipower chairman Tseng Wen-sheng's speech and press answers on 14 October 2026, as reported by Central News Agency and TVBS. The 13 October remarks on shorter investment clocks and scattered AI load come from the Taipei Times and Taiwan News, both citing CNA. The Mailiao letter of intent comes from Formosa Smart Energy's 7 October announcement as summarised by Island Frequency from CNA. Analyst price targets and buy calls are left out on purpose. Anything we worked out ourselves is labelled “our arithmetic”. A letter of intent is not a build contract, and a planned substation is not a live one, and we say so. The references we used are listed at the end. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) Not investment advice. Do your own research. --- ### References 1. Central News Agency (CNA), Taipower plans 43 substations over 15 years for AI and semiconductor power demand (Chinese), 14 Oct 2026\. [cna.com.tw](https://www.cna.com.tw/news/afe/202610140274.aspx?ref=sotkn.com) — 43 substations, 1,922 km of lines, \~52 GW framed transformer capacity, NT$30B for 12 STATCOMs, semiconductor kWh path, Southern Taiwan Science Park \~4 GW by 2031, 100 MW cost estimate, anti-squatter stance. 2. TVBS, AIDC should sit near power plants; Taipower chair says real investors agree (Chinese), 14 Oct 2026\. [news.tvbs.com.tw](https://news.tvbs.com.tw/money/4036032?ref=sotkn.com) — 52,450 MVA of new transformer capacity, near-plant siting, consensus with firms that intend to invest. 3. Taipei Times, Taipower head wants power sector at the table in tech decisions, 13 Oct 2026\. [taipeitimes.com](https://www.taipeitimes.com/News/biz/archives/2026/10/13/2003865846?ref=sotkn.com) — investment clocks shortened to three to five years; AI demand hard to forecast; grid delivery is the bind. 4. Taiwan News, Taipower urges quicker investment planning for Taiwan AI demand, 13 Oct 2026\. [taiwannews.com.tw](https://www.taiwannews.com.tw/news/6456849?ref=sotkn.com) — on-site and enterprise AI halls may scatter load; green power and storage coordination. 5. Island Frequency, Formosa Plastics plans an AI data centre and low-carbon power at Mailiao, 7 Oct 2026\. [ccs.tw](https://ccs.tw/news/formosa-mailiao-aidc-low-carbon-power?lang=en&ref=sotkn.com) — Formosa Smart Energy LOI with CXIT for a first-phase MW-scale AI hall; no MW or dollar figure disclosed.