Tokens are the data fee of a new era: is Taiwan's compute abundant or scarce?

"A token is a bit like the data fee of this new era — a telecom-scale business model." That is how Su Yen-che (蘇彥哲), CEO of Horizon AI, summed up the token economy that Jensen Huang keeps returning to, speaking on episode 286 of the Black Swan Academy (黑天鵝學院) podcast (full interview). This piece collects the figures and realities he sees from the front line of compute supply — on cost, scale, and where Taiwan stands.

Tokens: from jargon to a basic fee

Most people cannot feel "compute" today because they do not yet draw on it. But look a few years ahead: when half the apps on your phone run AI features in the background, each carrying a subscription fee, the substance of that fee is tokens — much like the monthly bill you pay a telecom carrier, behind which sit cell towers and bandwidth.

The difference is that users will no longer need to know what a token is, just as no one today counts the megabytes their packets consume. When the unit of pricing for a technology shifts from industry jargon to a basic fee, whoever controls the supply side holds a telecom-scale business model. That is why capital worldwide is pouring into AI infrastructure.

The truth about Taiwan's "abundant" compute

A counterintuitive fact: Taiwan's compute is, for now, abundant. Not because supply is excessive, but because demand has not yet woken up.

Overseas, a new management culture has emerged: companies now require engineers to consume a set amount of compute each month, treating it as a hard metric that you are using AI to improve your workflow. Some set the figure at NT$50,000 (about US$1,600) per person per month. In Taiwan, by contrast, most people's engagement with AI still stops at liking and sharing short videos on social media; even within tech circles, those genuinely using AI agents in daily work remain a minority.

The gap is cultural. Intense competition in the Chinese market pushes companies to charge in at full speed wherever there is a chance to pull ahead — ByteDance, Baidu, Alibaba, and Tencent have all launched their own large models and agent services. Taiwan, by comparison, is comfortable, and the cost of waiting looks low. But once the demand culture catches up — and it is catching up — Taiwan's supply conditions mean abundance will quickly turn to strain.

What an AI data center actually costs

Outsiders often picture an AI data center in a single word: expensive. The actual cost structure looks like this. The figures below reflect the Taiwan market.

• The economic unit is 32 servers. Building 8 or 16 units barely changes the base engineering cost — cooling, piping, floor loading, and reserved space — so a certain scale is needed before the economics work. The industry generally treats 32 GPU servers as one unit.
• The servers are the biggest line item. A new-generation B300 machine runs about NT$15–18 million (US$480,000–570,000) each; the previous generation, the H200 and B200, about NT$10 million (US$320,000). Built out to 32 units, a full site starts at around NT$600 million (US$19 million) and fits within roughly 1MW of power capacity.
• Facility engineering multiplies that by about 1.5. At high-density configurations, a complete 1MW-class build can reach NT$1.5–1.8 billion (US$48–57 million); a 20MW site can carry a total investment of US$300 million to US$1.1 billion, depending on configuration.

The scale gap is more striking still. A new site in northern Taiwan tops out at 5MW under policy, which the industry already regards as the ceiling; US peers are talking in terms of 1GW and 2GW — a gap of hundreds of times. During last year's GTC in the US, the team set out to find a compute cluster to rent locally, only to find that every relevant data center company was fully booked, with not a single server to spare.

Power, investment, and the sovereign AI deadlock

Taiwan's sovereign AI narrative contains a structural contradiction. Sovereign AI requires data to stay in Taiwan; keeping data in Taiwan requires enough local data centers — but:

1. Power limits keep any single site small.
2. A site that stays small shows investors no economies of scale, so they hold back.
3. With too little capacity, large orders go unfilled — the market has already seen a request to rent 650 servers on a five-year contract that no Taiwan site could take on.
4. So enterprises move their computing overseas, and the data goes with it.

The knot loops back to where it began. It also explains why almost every compute operator in Taiwan has chosen to go abroad: Japan offers tax and power subsidies, Malaysia's low electricity costs suit large sites, and demand in South Korea and Vietnam is strong. What Taiwan does well is to root the roots of compute — engineering capability, supply-chain relationships, and operational know-how — at home, while placing scale around the world.

What this means for enterprises: don't wait for infrastructure, build demand capability first

For most enterprises, the meaning of these figures is not "should we build our own data center" (the vast majority should not) but a question of timing: the price of compute and the barrier to obtaining it are in a window that precedes the surge in demand.

Companies that start putting AI into their workflows now enjoy the pricing and service flexibility of abundant supply. Once demand in Taiwan fully wakes up, you will be the one in the queue. Of two companies in the same business, the one that moves first optimizes its processes at a lower cost — and in this era, a small company overtaking a large one in a short time is something that genuinely happens, and already has.

---

Horizon AI is the enterprise AI adoption partner within the KONST group — turning the group's strengths in compute and infrastructure into AI that enterprises can genuinely use: assessment, custom development, system integration, and operations and governance, end to end. Book an AI readiness assessment →