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.
Every AI certification an engineer can register for in 2026, listed one by one: AWS AIP-C01 / MLA-C02 / AIF-C01, Google PMLE, Microsoft AI-103 and AI-500, all twelve NVIDIA exams, Databricks, Snowflake, Oracle's Agentic AI track, IBM watsonx, Salesforce Agentforce, GitHub GH-300, Anthropic's four Claude exams, Taiwan's iPAS AI Application Planner, plus the governance and audit line (IAPP AIGP, ISACA AAISM / AAIA, CertNexus CAIP) — prices, validity, and registration gates all checked against vendor pages. Two things that hit your wallet: Google's PMLE exam guide has renamed Vertex AI to Gemini Enterprise Agent Platform throughout, making pre-mid-2026 study material worthless, and the iPAS intermediate certificate is valid 5 years, not permanently.
Firecrawl's open-source Rust conversion library turns 14 office formats (including legacy .doc / .ppt / .xls) into GFM at a 4.7ms median — 109× faster than Docling under the same timing basis. The trade-off: it does no OCR at all.
Scans and complex layouts leave you no choice but to infer structure with a model. But the technical gap between MinerU, Marker, and Docling is far smaller than the licensing gap — MinerU needs a separate license past $20M monthly revenue, Marker's model weights need payment past a funding threshold, and only Docling is cleanly MIT. Read the LICENSE before the benchmark.
The most common mistake in feeding documents to an LLM isn't picking the wrong tool — it's picking the wrong layer. Structure already in the file goes to the conversion layer (milliseconds); text without structure goes to extraction; only inferred structure needs parsing. anydoc's 4.7ms against Docling's 513.6ms is a 109× gap, and most people jump straight to the most expensive layer.
Digital-native PDFs already contain readable text — what's missing is structure, and heuristics can recover it. PyMuPDF, pdfplumber, pypdf, and Tika do this with zero GPU and zero inference cost. The biggest selection trap isn't accuracy; it's PyMuPDF's AGPL-3.0 license.
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.
PX4's own architecture docs say the companion computer runs Linux because 'Linux is a much better platform for general software development than NuttX; there are many more Linux developers.' That sentence is the entry point. The three transition paths differ sharply in friction — CV and MAVLink work on the companion computer transfers almost directly, flight controller firmware means learning RTOS and work-queue constraints, and estimation and control means quaternions and Kalman filters.
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.
Once studying went fully digital, paper still holds three places with real evidence: reading (paper over screens at g ≈ −0.21 across 171,055 participants, widening to 0.35–0.48 when scrolling is required), writing while you answer (on screen, harder questions draw less scratch work, not more), and drawing (45% recall against 20% for writing). The one claim most people lean on, that handwritten notes stick better, spans −0.008 to +0.248 across four meta-analyses with no consensus.
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.
Hardware runs 35–55% gross margin under permanent DJI price pressure; autonomy software and DaaS subscriptions run 60–80% and recur. Skydio's software subscriptions were already ~30% of revenue in 2023 at a 38% blended margin; India's Garuda had DaaS at 62% of FY24 revenue with a 351-day cash conversion cycle against a defense-heavy peer's 597. Taiwan is almost entirely concentrated in the lowest-margin, most substitutable cell.
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 2016 consumer drone bubble left specific wreckage: 3D Robotics stopped making hardware, GoPro recalled all 2,500 Karma units six weeks after launch and cut 15% of staff, Parrot cut 35% of its drone workforce, and Lily Robotics collapsed after taking $34M in pre-orders. In 2023 even Skydio — $570M raised — exited consumer, and three years later it is valued at $4.4 billion. This wave runs on a completely different engine, but three things are exactly the same.
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.