Volume production of the 2-nanometre node began this quarter, eighteen months ahead of the schedule most analysts carried in 2024. The slip was upward. That is unusual, and it tells you something about how much capital and attention has been concentrated on one place.
That place is a cluster of science parks roughly sixty square kilometres in area, on the western coast of Taiwan, centred on the city of Hsinchu.
The geography of a process node
A leading-edge fab does not stand alone. It sits at the centre of a supply web that has to deliver, on a cycle measured in hours rather than days: ultra-pure chemicals, specialty gases, photomasks, precision tooling, and the engineering talent to keep yield climbing through the first six months of ramp.
Every one of those inputs has a supplier within ninety minutes of Hsinchu. Several have multiple suppliers within fifteen. That density is the actual moat — not the lithography tools, which are purchasable by anyone with the money and the export licence.
What the ramp looks like
Yield at the start of a leading-edge ramp is a closely held secret and a source of endless speculation. What can be said:
- The 2nm node uses gate-all-around transistor architecture, a structural change from the finFET design used for the previous decade
- Backside power delivery is in production for the first time at this node, separating the power network from signal routing
- Initial capacity is heavily committed — mobile SoC and AI accelerator customers hold the majority of first-year wafers
- Capital intensity per wafer start has risen roughly 40 percent against the 3nm node
That last figure is the one that matters for everyone downstream. The cost of leading-edge logic is rising faster than the transistor count, which is pushing chip architects toward advanced packaging and chiplet designs rather than pure node shrinks.
The most consequential square kilometres in the global economy are not in a financial district. They are in a science park where the water is cleaner than the air outside.
The concentration risk everyone has priced and nobody has solved
Roughly 90 percent of the world’s leading-edge logic capacity sits in Taiwan. Every government that has looked at that number has concluded it is too high. Every attempt to change it has run into the same wall: the ecosystem does not transplant.
Fabs have been announced in Arizona, Kumamoto, Dresden, and Dresden again. All of them are trailing-edge or specialty nodes relative to Hsinchu. All of them depend on Taiwanese engineering support, Taiwanese packaging partners, and ultimately Taiwanese capacity for the most advanced work.
The diversification is real at the margins and illusory at the core.
Power and water: the actual constraints
The binding limits on Hsinchu’s expansion are not geopolitical. They are physical.
A single leading-edge fab draws 50 to 80 megawatts continuously and consumes millions of litres of ultra-pure water daily. The science parks have been in a queue for grid capacity for three years. Water recycling has been pushed to levels that were considered uneconomic a decade ago.
New capacity is being sited further south, in Tainan and Kaohsiung, partly for land and partly because the power picture is marginally better. That is slowly redrawing the map — but the ecosystem gravity of Hsinchu means the new sites are still tethered to it.
What to watch
The 2nm ramp will be judged on two numbers over the next four quarters: yield at volume, and how much of the capacity premium customers are willing to absorb. If both hold, the node cements a two-year lead. If yield disappoints, the chiplet and advanced-packaging route becomes the primary path for everyone — which shifts value toward packaging capacity, and toward a different set of sites.
Either outcome runs through Hsinchu.