What TSMC is building
On July 12 2026, TSMC said it will build three more advanced-packaging fabs in the Chiayi Science Park in southern Taiwan, on a site of roughly 90 hectares. This is Phase II of a plant complex whose first two packaging fabs entered mass production only a month earlier, in June 2026.
The full build-out targets more than NT$300 billion in annual output value, about US$9.35 billion or roughly EUR 8.6 billion. These are not wafer fabs that etch transistors. They are packaging lines that stitch finished chips and high-bandwidth memory onto a single substrate, the CoWoS-class step that turns silicon into an AI accelerator.
One detail carries most of the weight. The capital is being poured in Taiwan, at Chiayi, and not in Arizona. TSMC is expanding the part of its supply chain that has been physically short, and it is doing so at home.
Why packaging is the real bottleneck
Advanced packaging, not the leading-edge wafer, is now the hard physical constraint on AI-accelerator supply. For two years the market assumed the scarce thing was the 3-nanometre transistor. It was not. The wafers existed; what could not keep pace was the CoWoS packaging that bonds those wafers to memory stacks, and that shortage is what throttled GPU shipments through 2025 and into 2026.
Packaging capacity cannot be conjured with a firmware update. It needs cleanrooms, bonders, and substrate supply, all of which take quarters to stand up and years to scale. That is why three more fabs matters more than another wafer node. TSMC is not chasing a faster transistor here; it is widening the neck of the bottle.
What it means if you buy compute
If you buy compute, packaging-driven allocation should ease later, not sooner, so budget accordingly. Concrete poured in 2026 becomes qualified, yielding capacity on a 2027-and-beyond horizon, not in the current planning cycle. Anyone modelling GPU availability on the assumption that this announcement frees supply next quarter is reading the calendar wrong.
The signal itself is clear. TSMC committing to three more Chiayi packaging lines, on top of two already in mass production, is the strongest evidence yet that the constraint which held back accelerator shipments is being relieved structurally rather than patched. For European fabless designers and compute buyers who have spent two years on allocation lists, the direction is right even if the timing is patient.
Plan procurement around the horizon, not the headline. The bottleneck is being widened in concrete, and concrete keeps its own schedule.
Read next: The Nvidia Tax Just Found a Second Source | The Chip Behind Every AI Rack Gets a Second Source
