Written by Al Vigier.
Image credit: 08.06 總統視導「漢光42號演習─八里守備區『戰力防護與保存』實兵演練」 by 總統府/ Flickr, license: CC BY 4.0.
On 10 and 13 August, for thirty minutes each time, Taiwan deliberately slowed its mobile networks to a crawl. The National Communications Commission had the three carriers cut 4G and 5G data to 256 kilobits per second, roughly 1 per cent of normal capacity, across fourteen cities and counties in central and northern Taiwan, while voice calls and text messages kept running. Some sixteen million people, about 70 per cent of the population, were inside the throttle. It was the most instructive element of Han Kuang 42, the ten-day military exercise held from 5 to 14 August, and of the civil resilience drills integrated with it for two full days for the first time. The rest of the programme rehearsed evacuations, underground hospital operations and drone resupply. The throttle was different in kind. It isolated the one variable on which all of the rest depends: whether people and institutions can still make sound decisions when the network fails. My argument is that this capacity, which I will call the decision layer, is the piece Taiwan is not buying. It is buying sensors, hulls and bandwidth at speed. It has legislated the principles a trustworthy decision layer should meet and has a precedent for writing such principles into contracts. It has not written them into the contracts that matter for civil defence, and this summer’s budget politics widened the gap.
The binding constraint in a crisis is the capacity to convert fragmented, multi-source information into fast, trusted and defensible decisions while communications are degraded and dozens of civil and military actors must coordinate without the cloud, meaning the distant data centres most modern software silently depends on. Sensing and bandwidth are comparatively abundant. Disciplined judgement under those conditions is scarce, and it fails faster than any sensor does. When two undersea cables to Matsu were severed in February 2023, the islands endured roughly fifty days of what observers called an invisible blockade, thirteen thousand residents falling back on a microwave link. A coercive blockade of the main island would reproduce that condition at scale and for considerably longer. That, rather than a thirty-minute drill from which the south was excluded so as not to disturb share trading during exchange hours, is the condition against which capability should be specified.
Degradation of this kind is a recurring operating condition, not a war-game abstraction. At least eleven cases of undersea cable damage were recorded around Taiwan between 2023 and early 2025. The coast guard has blacklisted ninety-six China-linked vessels, most under flags of convenience. The captain of the Togo-flagged Hong Tai 58 was charged after a cable to Penghu was severed in February 2025 and sentenced to three years, and a Chinese fishing captain was fined and deported for damaging the second Taiwan-Matsu cable in October 2025. As recently as 30 March this year, a Chinese-flagged salvage barge was suspected of damaging the third Taiwan-Matsu cable. Prosecutions deter. They do not restore connectivity, and they do not tell a district emergency operations centre what to do in the hours after a link drops.
Taiwan’s answer so far has been substantial and overwhelmingly material. The 2026 defence budget was set at a record NT$949.5 billion, above 3 per cent of GDP for the first time since 2009 on the NATO-style measure Taipei now uses. On 20 August, the Cabinet approved a 2027 proposal of NT$1.1225 trillion, the first above NT$1 trillion, against a stated target of 5 per cent of GDP by 2030. The coast guard is completing a 141-vessel shipbuilding programme. The Ministry of Digital Affairs is building 773 non-geostationary satellite sites, ground terminals linked to low-orbit satellites rather than to the undersea cables, as an emergency backup. Each hull, sensor and terminal generates data. Data is not a decision. A duty officer during Han Kuang holds more information than any predecessor and less time in which to act on it. Additional feeds do not resolve that asymmetry; they intensify it, because a human still has to reconcile an incomplete picture and remain accountable for what follows.
The budget politics of the summer turned that gap from an argument into an accounting fact. President Lai Ching-te’s proposed NT$1.25 trillion eight-year special defence budget was cut on 8 May to NT$780 billion and tethered strictly to United States Foreign Military Sales, the channel through which Taiwan buys certified American hardware. About a third of the original package had been earmarked for unmanned systems, more than 200,000 drones and 1,320 crewless surface vessels. Governing-party legislators noted that the settlement also leaves out the AI-assisted decision-making modules planned for the T-Dome air-defence system, the closest thing to a decision layer Taiwan had ever put in a procurement document. The Cabinet came back in June with a separate NT$210 billion special act for domestic unmanned systems, and on 27 August the legislature rejected it and passed its own version, which promises NT$240 billion over six years but only through annual budgets capped at NT$40 billion a year. The defence ministry, which had planned NT$42.1 billion for 2027 alone, says it cannot build real drone combat power on those terms. The domestic, autonomy-adjacent portion of the build-up therefore exists on paper in a form the ministry rejects, and the decision-support element is funded nowhere. Taiwan is buying more feeds while the capacity to use them waits on a budget fight.
That capacity deserves to be treated as national-resilience infrastructure and specified accordingly. Four properties are load-bearing. First, it must run on-premises and, while disconnected, on computers Taiwan controls, and be able to keep working with no link to the outside world, because a blockade removes the cloud and a throttled network removes the assumption behind it. Second, it must keep a human in control, what engineers call human-in-the-loop: an operator accepts, modifies or rejects each significant recommendation rather than having the decision automated away. Third, it must be explainable and auditable: every recommendation carries a stated level of confidence, its named sources and a record of who decided what and why, so an official can defend the decision afterwards before the legislature or a court. Fourth, the data must remain in Taiwan, on infrastructure and a supply chain whose provenance, meaning where each component and line of code came from, can be verified. Together, they separate decision support a government can lean on under stress from a system that fails silently when needed.
Taiwan has already legislated the principles. The Artificial Intelligence Basic Act, passed on 23 December 2025 and in force since 14 January 2026, commits the state to seven principles, among them human autonomy, transparency and explainability, and accountability. It requires the Ministry of Digital Affairs to build a risk classification framework interoperable with international standards, and the ministry published that framework on 7 July. The precedent for converting a principle into a purchasing condition exists too. In its satellite resilience programme, the digital ministry required participating operators to be free of Chinese ownership. The same move is available here: do for the civil-defence decision path what Taiwan has already done for satellite backhaul and for semiconductors, and convert a stated principle into a contract clause with a test attached. A statutory principle that never becomes a procurement requirement is an aspiration, and an aspiration cannot be audited after a failure.
What did the August drill show? At the centre, the system held. The government reported that police took 243 calls on the 110 emergency line during the central Taiwan throttle, in line with last year, and that local governments’ satellite and radio backups worked. At the edge, it did not. In Taichung, a person sheltering in a basement tried repeatedly to send a photo message reading “HELP” and could not. Defence Minister Wellington Koo had framed the exercise as a test of command-and-control backup capabilities whose results would decide whether to extend it to other regions. The drill proved that a thin channel survives. It did not test whether anyone on that channel could reach a validated decision, because the tool for doing so has not been bought. Bandwidth was the variable measured. Judgement was the variable assumed.
The window for that purchase is open. Allied technology firms have been partnering with Taiwanese hardware makers since Computex, producing sensors and platforms quickly. Whether that coalition also produces trusted decision-making under degraded conditions depends on a procurement choice not yet made. The 2027 budget is now before a legislature that has just rewritten the drone programme once. The NT$480 billion tranche of the special budget still waits on a United States letter of offer and acceptance, the formal sales offer from Washington, and on a further legislative review. So the argument closes where it opened. Han Kuang 42 tested the right variable. The task now is to fund the capability that lets Taiwan pass that test on the worst day, and to write explainability, human control, audit-ready records and a verified supply chain into the contract as hard requirements, beside the ships and the drones.
Al Vigier is the CEO of Caseway, an AI company that co-creates products with large enterprises and shares the revenue with them. He spent seven years in the Canadian Army, where he was shot. The views expressed are the author’s own.