Appier (TSE: 4180) published SMITH, a reinforcement-learning framework that trains a single model to both build and use tools, with the paper accepted at NeurIPS; a 4B-parameter model trained with SMITH outperformed an untrained 30B-class model. Minister Lin Yi-Ching of the Ministry of Digital Affairs reiterated at a recent forum and at DevDays Asia 2026's public-sector track that government shouldn't lead AI industry development through budget spending, pivoting instead to guiding private capital into compute-center buildout (targeting over 10,000 GPUs by next March) and the AIEC model-evaluation center. Meanwhile, a Group-IB report ranked Taiwan third in the Asia-Pacific for August ransomware incidents (20, behind only India and Australia) — the first time this series covers Taiwan.
Generative AI: Text and Image Synthesis Principles and Practice is an introductory course taught by Yen-Lung Tsai (蔡炎龍) of NCCU's Department of Mathematical Sciences and opened to other schools as a TAICA satellite course. The most complete course page online actually belongs to the Chang Gung University satellite section, where Chih-Yuan Yang is the co-teacher. This series follows Spring 2025 (semester 1132): 14 recordings, 14 slide decks, and 12 homework specs with rubrics are public, and the demo notebooks are on GitHub, so the access grade is A3. The gaps: the notebooks keep changing, submission and grading run through each school's LMS, and final projects were never published.
A guide to the BERT and its Family unit of Prof. Hung-Yu Kao's NLP course at NTHU (Fall 2025). It starts from "I record the record": Word2Vec and GloVe give both records the same vector. ELMo fixes this with a bidirectional LSTM language model. With Transformers, pretraining splits into three roads: encoders (BERT: MLM plus NSP, strong at understanding, weak at generation), encoder-decoders (T5's span corruption, BART's five noise types), and decoders (GPT: pure next-token prediction). The last part covers GPT-3's in-context learning and scaling laws, and why decoders became today's dominant backbone.
A guide to the decoding and evaluation unit in NTHU Prof. Hung-Yu Kao's NLP course (Fall 2025). The first half covers how to pick a word once the model outputs a probability distribution: greedy decoding can't take back a mistake, beam search keeps several candidates but favors short outputs, and top-k / top-p trade determinism for diversity (missing from the slides; the professor covers it verbally in class). The second half covers scoring generated text: BLEU's modified precision and brevity penalty, ROUGE-N and ROUGE-L, perplexity, and what GLUE, SQuAD 2.0, MTEB, and MMLU each measure.
A guide to the GPT-2 / T5 tutorial in NTHU Prof. Hung-Yu Kao's NLP course (Fall 2025). One task, LCSTS Chinese summarization, is solved twice. Decoder-only GPT-2 is written in native PyTorch: you join article and summary into one sequence, switch to left padding, and set padding labels to −100. Encoder-decoder mT5 uses Seq2SeqTrainer: no left padding needed, and DataCollatorForSeq2Seq handles the −100 for you. Both segment with jieba and score word-level ROUGE.
A guide to the Hugging Face BERT tutorial and HW3 in NTHU Prof. Hung-Yu Kao's NLP course (Fall 2025). The tutorial walks through binary IMDb sentiment classification: AutoTokenizer, the input_ids / token_type_ids / attention_mask fields, AutoModelForSequenceClassification, and Trainer. HW3 applies the same tools to SemEval 2014 Task 1: one bert-base-uncased with two heads, one regressing a 1–5 relatedness score and one classifying entailment into three classes. You add the two losses and write the training loop yourself, because Trainer is not allowed.
HW1 tests word vectors on the 19,544 Google Analogy questions (8,869 semantic, 10,675 syntactic). You first answer them with pretrained glove-wiki-gigaword-100 loaded through Gensim, then train your own Word2Vec on a 20% sample of a pre-cleaned Wikipedia dump, and plot t-SNE for the family subcategory both times. Seven TODOs are worth 55%, the report 45%. Fall 2026 keeps the same TODOs but asks for an .ipynb with outputs.
Hung-Yu Kao's Natural Language Processing at National Tsing Hua University is a graduate-level flagship course in the TAICA alliance. The syllabus caps it at 1,200 students, it is taught in Mandarin, and it runs from TF-IDF and word vectors to RLHF, PEFT, and RAG. For Fall 2025, the slides, 32 class recordings, and 4 assignments with starter notebooks are all on GitHub, which rates A3. Solutions, grading, and the term-project spec are not public. Fall 2026 is in progress and only goes up to W3, so it rates A2. Grading changed to 75% assignments plus a 25% in-person midterm, and a Reasoning/Agent unit was added.
This 34-slide TA session answers a practical question. Pasting data into the ChatGPT web page one row at a time is slow and hits hourly limits, so research and homework should use the API. The notebook runs one SemEval 2014 entailment example through Gemini, Claude, and OpenAI in turn: prompts live in prompts.yaml, output is forced into JSON, then few-shot and token counting. The material is from 2024. The slide cover says 2024/11/21, and the notebook uses gemini-1.5-pro, gpt-4o, and claude-3-5-sonnet-20241022. That Claude model was retired on 2025-10-28, and Google's old Gemini SDK reached end of support on 2025-11-30.
HW2 treats expressions like "14*(43+20)=882" as character sequences and asks a two-layer LSTM to generate the answer one character at a time after it sees "=". The training set has 2,369,250 rows and the eval set 263,250, with every number in 0–49. Six TODOs run from building a vocabulary and batching with loss only after "=" to a generator, teacher-forced training, and exact-match evaluation. The W4 PyTorch TA session is the toolbox for it.
Each of the two RAG TA sessions builds one version. The first installs Ollama on Colab to run llama3.2:1b and wires up a minimal RAG with LangChain's Chroma, MMR, and retrieval chain. The second uses LangChain only for data prep and writes the rest by hand: chunking, text and vector stores, hybrid BM25 + cosine retrieval merged with RRF, then generation with Llama-3.2-1B-Instruct. HW4 applies the first session's skeleton to 150 cat facts and 150 GPT-5-generated QA pairs. The generator must be Llama3.2-1b and the embedding model jina-embeddings-v2-base-en, and you report recall@1, recall@5, and exact match. Code is 45% of the grade and the report 55%; the report analyzes how prompts, data format, document order, and counterfactual information change the results.
The second half of W11_RAG.pdf starts at the "From Retrievers to QA" slide and turns a retriever plus a reader into a full QA system. It begins with ORQA and REALM (2019–2020), where BERT is the reader, then covers the first paper named RAG and REPLUG, which keeps the LLM frozen. A single table then sorts seven recent fixes into three groups: rewrite the query (Query Rewriting, HyDE), make the generator robust to noise (RetRobust, RAFT, RAAT), and decide when to retrieve (FLARE, Self-RAG). It ends with noise types, four abilities an LLM needs inside RAG, and generative retrieval (GR) with reliable response generation (RRG). On the recording side, the W11 Thursday lecture stops at the RAG paper; I could not find a recording that covers the later slides.
"Look over there" is 3 tokens, the Chinese version 4, the Japanese version 12. When output length does not track input length, the classifier trick of adding an FFN at the end fails. This 33-slide deck starts from encoder-decoder models, works through vanishing gradients in RNNs and the three LSTM gates, and ends with attention fixing both long-range memory and parallelism, including why scores are divided by √d.
A guide to the Sub-word Tokenization unit of Prof. Hung-Yu Kao's NLP course at NTHU (Fall 2025). Splitting on white space only works for Western languages and cannot handle unseen words or German-style compounds. BPE starts from characters and repeatedly merges the most frequent adjacent pair into the vocabulary, so each merge adds one entry; its weakness is that the greedy split is not always the best one. The Unigram LM picks splits by probability and can even sample different ones. The slides give vocabulary sizes of 30522 for BERT, 50257 for GPT-2/GPT-3 and 32,128 for T5.
Fall 2025 was graded 70% assignments + 30% term project. Projects were done in groups of 3–4 and split into Proposal 6%, Progress 6%, Poster 6% and Report 12%, with no GPUs provided. The repo has no project spec, only the syllabus structure, an end-of-term reminder and the W15–W16 recordings. Fall 2026 switches to 75% assignments (4 of them) + a 25% in-person midterm in W14. The schedule drops the presentation weeks, adds a Reasoning/Agent unit, and brings in an AI-TA for grading support and TAICA compute credits. The official materials don't say why.
A guide to the Transformer unit of Prof. Hung-Yu Kao's NLP course at NTHU (Fall 2025). The slides start from two RNN problems: words interact only across O(N) steps, and time steps cannot run in parallel. Self-attention lets every word look at every other word and computes it all in one matrix multiplication, fixing both at once. The cost is that the model can no longer tell word order, so sinusoidal positional encoding is added. The lecture then assembles multi-head attention, Add & Norm, feed forward, cross-attention, masked attention and teacher forcing, and closes with GPT-2, ViT and four variants to show how far the architecture went.
Most public AI courses in Taiwan outside NTU come from TAICA, an alliance set up by the Ministry of Education. Each semester's course list says where every flagship course streams, and courses that stream on YouTube are usually watchable by anyone. Two courses are complete enough to self-study: Hung-Yu Kao's Natural Language Processing at NTHU (Fall 2025) and Yen-Lung Tsai's Generative AI at NCCU (Spring 2025), both A3. Wei-Ta Chu's Introduction to AI at NCKU, Ping-Hsuan Han's Human-AI Interaction at NTUT, and Min-Chun Hu's Robotic Navigation and Exploration at NTHU have full recordings but keep assignments on NTU COOL, so they rate A2. NYCU's TAICA courses are taught in English; the Deep Learning recordings are not publicly listed and Physical AI has just started, so neither made the main table.
cnYES publicly offers more than ad inventory: video, events, sponsored features, editorial production, and historically, B2B news licensing. Free content attracts readers, while the platform balances advertiser outcomes, production costs, and editorial trust.
BigGo Finance places public AI podcast summaries, market data, and news at the free entrance, then offers Pro upgrades around model capability, alerts, and experience. That creates a plausible product path, but public evidence does not show that summary readers convert into paying subscribers.
CMoney describes its product design as a three-step path from method to tool to community. Free content and discussion are designed to surface needs, shared data and APIs can turn investing methods into many apps, and public products show monetization through subscriptions, courses, and institutional systems—but the company has not published retention or revenue proof for the loop.
Public materials do not show Fugle charging directly for free articles or taking a commission on every trade. It attracts investors with research tools, then monetizes personal API subscriptions, information services, and B2B technology while the brokerage still holds the account and executes the trade.
Medium and Vocus can both deliver discovery, but the exchange is different. Medium controls distribution to strangers and withholds new subscribers' email addresses; Vocus handles Taiwanese payments, invoices, and member operations, but its public documentation proves only that order data—not complete email identities or payment relationships—is exportable.
Creator-platform selection is not a ranking. A Taiwan-based creator should pass four gates—payments, discovery, control, and operations—before choosing a platform, managed SaaS, or self-hosted site. No CSV represents a complete reader and billing relationship.
Taiwan's industrial density can become global enterprise intelligence, but a new company must pass five gates: unique signals, recurring demand, structured fields, lawful verification, and overseas buyers. TrendForce shows the model exists; it does not show that every vertical can copy it.
Taiwan's Cybersecurity Act 2.0 passed and FSC's three-tier system is in place, but neither mandates specific certifications — real demand comes from the job market. Taiwan has 7+ training providers: Uuu has the widest catalog (only one offering both SecAI+ and COASP classroom courses), WUSON is the CISSP legend (monthly cohorts sold out through mid-2027), and DEVCORE exclusively brings OffSec factory instructors. Test centers in Taipei and Kaohsiung; most certs also support online proctoring from home.
GPUtw.ai is a Taiwan-based short-rental GPU cloud. Its main value is not maximum scale, but Taiwan data centers, prepaid credits, Jupyter/ComfyUI/Ollama/vLLM templates, Vault storage, and team billing. Public information is enough for a service introduction, not enough for procurement or production endorsement.
On 109 posts that genuinely contain Mainland vocabulary, zhtw-mcp scored 29.4% precision at 85.2% recall; twlint scored 21.2% / 82.5% — but 195 of twlint's error-level findings are legitimate Traditional characters misread as Simplified (干→幹 60 times, 只→隻 15), so --fix corrupts the text. The most useful result was not a winner: zhtw-mcp found five Mainland terms my own list had missed, and correctly declined to flag one it wrongly included (審計). These tools calibrate your wordlist; they do not replace it.
Three Legislative Yuan budget evaluations log 23 drone-caused airport closures in Taiwan between FY2019 and August 2023, and every row's response column repeats one sentence: on notification, go to the scene with the Aviation Police and investigate. Article 99-13(6) of the Civil Aviation Act gives an airport the verb 'enforce against', not 'stop or remove' — yet the FY2024 report says every airport has already procured handheld jammers.
Paragraphs 3 and 4 of Article 99-13 of the Civil Aviation Act say 'government agencies, schools or legal persons', but paragraph 7's proviso — the only sentence letting a site operator 'take appropriate measures to stop or remove' a drone — says only 'government agencies', and Taipower, CPC and the fabs are legal persons. Their remaining path is to have a municipality announce a restricted zone and 'enforce against' violators, which drops the penalty from an NT$300,000 minimum to an NT$300,000 maximum, while importing a jammer requires 'critical infrastructure provider' status that you learn you hold by receiving a letter.
Eight municipal police departments now run drone units, and the authority to fly is clear — Article 99-16(2) of the Civil Aviation Act exempts agencies performing statutory duties. But that is a flight-safety exemption: a Taipei City legal opinion says it 'may be unable to serve as the legal basis for conducting administrative investigation by drone', and the special-compulsory-measures chapter passed in July 2024 authorises GPS, IMSI-catchers and private-space imaging with no aerial-investigation provision at all.
'Drone' appears in 19 of Taiwan's central statutes and regulations; 'counter-drone' appears in exactly one, and that one is a procurement act. Six settings are authorised to 'stop or remove' an intruding drone and none of them says by what means — Article 99-13 of the Civil Aviation Act says only 'take appropriate measures', while the 31-item police instrument list includes mortars but no jammer, and Article 67(1) of the Telecommunications Management Act has no public-duty exception.
Thermal camera prices jump in steps because ECCN 6A003.b.4.b splits control on 111,000 focal-plane elements: 384×288 clears it by 408 pixels, 640×512 is three times over, nothing common sits between, and FLIR prices the Boson 640 at 2.04–2.31× the 320. The Note 3.b carve-out then limits not resolution but focal length — you may have a thermal camera, just not a telephoto one.
The public evidence of a production ramp isn't in the factory but in the government e-procurement system: the fire service's 88 thermal drone sets were priced centrally, six fire departments' first tenders drew no qualifying bidder, all seven awards I pulled equal the budget to the dollar, and the bidder pool is about six firms. A control group added afterwards — aerial ladder trucks, at 0.54 failed-to-awarded versus the drones' 0.33 — refutes this post's original first conclusion: failed tenders are the norm in Taiwanese fire-service procurement.
Grepping all 9,199 lines of Skybrush's open-source light-show firmware for neighbour, collision and swarm turns up nothing: no aircraft knows another exists, and the entire content two hundred of them synchronise is one GPS time-of-week integer plus a millisecond offset, with collision avoidance settled on the ground in the choreography software. Chapter 7 of Taiwan's drone cybersecurity spec — the swarm chapter — leaves the 'drone' column a dash down all twelve items and tests switches and routers instead.
Reading the whole specification turns up three counter-intuitive things: of the seven mandatory items only three sit on the aircraft and the rest on the ground control station; an unencrypted link still passes, as long as the manual or the box states the reason or the risk — the rule asks for disclosure, not encryption; and the five items that actually test resilience (spoofing, jamming, link loss, known firmware vulnerabilities, app certification) are all in Chapter 8, which is optional.
The same 5 kg aircraft on the same 2 kg pack needs 512 W to hover for 41 minutes as a multirotor, or 128 W to cruise for 164 minutes and 197 km as a fixed wing — and that range is independent of cruise speed. A VTOL's cost is not the energy burned in transition (2.4%) but that it carries both a rotor system and a wing, worth about a quarter of the range; Taiwan builds something at every rung from 40-minute electric multirotors to a 6-hour piston helicopter.
Analogue FPV video in Taiwan has two separate problems: LP0002's 5.8 GHz window is only 125 MHz wide while the market's standard channel table spans 300 MHz, leaving eighteen of the forty classic channels inside the window once transmit width is counted. The harder one is that §4.10.1.1 recognises only frequency-hopping transmitters at 5.8 GHz, and analogue video is fixed-frequency — it doesn't violate a clause so much as fail to match any clause in the document.
I had filed "what's the real margin on military-grade-commercial tenders, and how long is the cash cycle" under questions requiring interviews. The published filings answer both, more precisely: gross margin runs 35–39%, normal for hardware; but operating expenses consume it, and operating income has been negative for three straight quarters while reported net income came from non-operating items. The real constraint is inventory — roughly 385 days of it, producing a cash conversion cycle near 377 days. The money in this business isn't stuck in margin, it's stuck in inventory.
In 2022 Taiwan's Ministry of Transportation issued a press release titled 'Drone Licensing Overhaul! Obscure Questions Removed, Question Banks Fully Published,' putting all 1,420 questions online — 388 general, 588 professional, 324 renewal, 120 simplified renewal. The stated reason was candid: candidates said it was too hard. And the published bank became a policy document in its own right — the meteorology subject imports the manned-aviation syllabus wholesale, and flight principles lean heavily on fixed-wing aerodynamics, while most candidates fly multirotors.
Taiwan's Civil Aviation Act drone chapter regulates flight safety, not the people being flown over — and that isn't my commentary, it's the Legislative Yuan's own 2020 research report: the chapter 'contains no specific provision on the privacy management that matters most to the public.' So privacy falls back on Article 315-1 of the Criminal Code and the Personal Data Protection Act, and case law shows that catches part of it: pointing a drone at a hot spring room drew an indictment, raising a phone to a window drew four months. What it doesn't catch is evidence — the report's own words: 'by the time the victim notices it, it may already have vanished.'
Agricultural spraying is the one drone application whose unit economics are fully transparent: the billing unit is the fen (about 970 m²), rates run NT$150–300, spraying one fen takes about two minutes, equipment costs NT$300–500k, and even the substitute's price is public (manual spraying, roughly NT$200 per fen). And precisely because anyone can run the arithmetic, everyone did — operators who entered in 2018 sprayed over a thousand hectares a year and cleared over a million; later entrants broke even after two or three years and quit. Licensed operators charge NT$300 per fen; unlicensed ones undercut to NT$150. This is the industry cycle in miniature, compressed into about six years.
I assumed inspection was blocked by beyond-visual-line-of-sight rules the way logistics is. It isn't. Bridge, transmission tower, and high-speed rail viaduct inspection are all running with hard numbers: one bridge went from 8 inspectors, 4 vehicles and 2 days to 5 people, 1 vehicle and half a day with no traffic control at all, at 60% of conventional cost; high-speed rail crews covered at most 700 metres a day on foot and now save 3–5x; a private power plant cut headcount by three quarters and cost by half with no outage required. The reason is that all of these are segmented, fixed-point tasks completable within visual line of sight. What BVLOS actually blocks is continuous long-range routes, not 'inspection' as a category.
I assumed Taiwan had no real drone logistics because it has no BVLOS framework. Wrong: the CAA has approved 10 cases across 24 flight corridors, and the Institute of Transportation has run a six-year PoC → PoS → PoB path since 2020, entering commercial validation in 2025. The way it routes around regulation is the mirror image of inspection — inspection cuts work down to within visual line of sight, logistics gets corridors approved one at a time. And its value isn't being cheaper than a boat: during the typhoon sailing suspensions at Liuqiu, a drone made the crossing in a bit over ten minutes, which is what you have when the boats don't run.
Agricultural spraying runs NT$150–300 per fen, computable to the decimal. Search and rescue has no such number, because a life recovered has no price. That difference determines two things: first, the specification is written by terrain rather than performance — the defining feature of Taiwan's fire agency drones is that they do NOT depend on GPS, because mountain signal drops and they must thread through trees; second, when the legislature moved to cut the budget, the ministry could only point to one man pulled from a flooded river in June. Defending a budget with an anecdote is fragile, and this application has no better weapon.
Electronic warfare has one structural limitation: it needs a signal to attack. Fiber-optic control emits no radio, satellite links bypass ground jammers, and AI terminal guidance needs no link at all for the final leg — three methods systematically defeating jamming, which is why defense is shifting toward interception. Taiwan carries an extra layer: a Control Yuan investigation documents the Ministry of National Defense revising its drone response SOP three times in two months, from "may shoot down" to "flare warning only, first shot requires the Minister's authorization" and back to "shoot down with 7.62mm or smaller." That is not a technology problem.
Only four drone occurrences have entered Taiwan's official aviation safety statistics — because the threshold is 'substantial damage to a drone over 25 kg.' A hobbyist crash never enters the count. The two published investigation reports are the same model, the same manufacturer, and the same agency, and both probable causes were hardware failures: main rotor servo electrical failure in one, a fractured tail rotor pitch link in the other. In both, the flight control computer and the operator were explicitly cleared.
The three most important lines on a drone spec sheet are exactly the three lines spec sheets don't print. Taiwan's Drone Cybersecurity Testing Specification defines a product 'series' as units whose flight control, communications, and satellite positioning chip modules are all identical — regulation decides whether two drones are the same drone by those three modules, not by looks or endurance. And the weight field isn't a marketing number either: 250 g, 1 kg, 2 kg, 15 kg, and 25 kg are five separate legal thresholds.
Put drone jobs back into the five-layer supply chain and the picture clarifies: Layer 2 belongs to mechanical and electrical engineers, while Layer 3 — flight control firmware, sensor fusion, RF, edge AI — is the real entry point for software people, and also exactly the layer Taiwan is short on. But check the scale first: 267 companies and NT$12.9B of 2025 output means far fewer openings than the news volume suggests.
TEDIBOA has grown from 50 founding members to over 260, but since 1 July 2025 you must first be a full member of the Defense Industry Development Association before you can even apply. R&D grants under the Industrial Innovation Platform cap out at 50% of total project cost, three concurrent cases per company, three years maximum — and explicitly ban red-supply-chain components while requiring joint cybersecurity lab testing. This piece covers only what public documents can actually establish.
Taiwan has no 'drone department.' The system's core is National Formosa University's Department of Aeronautical Engineering — a NT$90M Ministry of Education training base sited inside the Chiayi drone cluster, the only one of 18 regional bases located in an industrial park, plus a NT$50M lab with a large low-speed wind tunnel. And the 2026 Presidential Cup's three challenge topics (edge AI recognition, high-precision positioning, GNSS-denied autonomous navigation) are almost a mirror of the industry's Layer 3 gap.
Taiwan's public drone money splits three ways: an approved NT$44.2B coordination program (R&D grants), a proposed NT$210B defense procurement special statute (stuck in cross-party negotiation), and annual agency budgets (NT$7.2B+ for 2026). The largest was written to run from 1 August 2026 — that date has passed with the bill still unresolved. And the Executive Yuan version buys only three specific items.
Measuring the drone sector against supply chain chokepoint / structural demand / high switching cost / long-term institutional holding: Taiwan sits at the most substitutable layer, 80% of demand comes from public budgets rather than end-user behavior, and only the certification-driven switching cost genuinely holds. The Army's NT$988M counter-drone contract — failed three times, terminated in full, NT$98.78M performance bond forfeited — is the most expensive lesson in why winning a bid is not revenue.
The general licence is written-test only and costs NT$450 total (NT$200 test + NT$250 certificate). The professional basic tier adds a practical test and medical exam for NT$1,900. Professional tiers must be climbed in order — reaching the top bracket realistically takes 18 months to two years. Each advanced group (G1/G2/G3) costs another NT$1,200. Fee figures come from Annex 17 of the regulations; the widely circulated 'NT$500 practical test' is wrong.
Register anything 250g or heavier; the registration number expires after 2 years. Individuals only need a licence at 2kg–15kg with navigation equipment. The licence term is now 3 years, not 2, and the student licence age dropped to 14, not 16. This piece covers only the currently effective text of the Remotely Piloted Drone Management Regulations, and flags the three most common outdated claims circulating online.
The EU has had a workable path since the end of 2020 — the Specific category grants an operational authorisation based on risk assessment, with U-space Regulation (EU) 2021/664 rolling out on top. The US Part 108 rule was still at OIRA and unpublished as of July 2026. Taiwan doesn't have the framework at all: its regulations offer 'extended visual line of sight' (900m, 400ft, observer required), while true BVLOS runs on per-activity permits valid for three months.
The global drone market is roughly US$69B in 2026 (IDTechEx). China holds about 80% of it (CSIS) and DJI over 70% of multi-rotor. The FCC put every foreign-made drone on its Covered List in December 2025; Taiwan's drone output jumped from NT$5.0B to NT$12.9B in one year, and Q1 2026 exports already beat all of 2025. This piece breaks down the five-layer supply chain, the four demand blocks, and the two ceilings holding back scale.
Of the 267 companies the Ministry of Economic Affairs counted, 164 are in northern Taiwan. But geography is not division of labor — Thunder Tiger's published bill of materials shows motors, batteries, frames, and propellers sourced locally, while flight control, comms/GPS, and camera modules go to US, European, and Japanese partners. Exports were only 23% of 2025's NT$12.9B output, and 88.1% of export value sits in the 2–7 kg weight band per Ministry of Finance statistics.
A breakdown of the three-layer digital ecosystem structure: LINE's super-app, Shopify App Store flywheel, and Taiwan MarTech integration strategies. The core mechanism is using APIs and data flows to create mutual dependency among participants, collectively reinforcing the moat.