TSMC is expected to expand its 2nm and 3nm production capacity further through mid-2027. Taiwan's Economic Daily News reported on September 14, citing supply chain sources, that capacity will grow 22% for 2nm and by more than 16% for 3nm from the end of 2026. A forecast has also emerged that CoWoS, TSMC's advanced packaging technology, will double by the end of 2028. However, front-end manufacturing and back-end packaging differ in both the expansion timeframe and what they process. To gauge how far AI chip supply can grow, it is necessary to look not only at monthly fab capacity but also at the size of the products made there.
Expansion timing differs between front-end processes and CoWoS
According to Economic Daily News' September 14 report, TSMC's monthly 2nm capacity is expected to rise from 90,000 wafers at the end of 2026 to 110,000 by mid-2027. 3nm is said to expand from more than 180,000 wafers to 210,000 over the same period. Both figures concern future capacity, not actual production as of September.
TSMC has not formally confirmed the reported growth rates. The paper says TSMC declined to comment on market rumors on September 13 and replied that explanations about capacity should follow the company's own announcements. The specific monthly figures are therefore best read as plans relayed through the supply chain.
Separately, an analysis by market participants cited in Economic Daily News on September 11 forecasts that monthly CoWoS capacity will grow from 130,000 wafers at the end of 2026 to 260,000 at the end of 2028. The expansion is said to center on AP7 in Taiwan and a site in Arizona in the US.
The 2nm and 3nm expansion outlook covers roughly six months, from the end of 2026 to mid-2027, while the CoWoS doubling outlook covers two years, from the end of 2026 to the end of 2028. Lining up the reported starting and ending points by process gives the following picture.
| Process | Projected monthly capacity, end of 2026 | Target date and projected monthly capacity | Nature of the information |
|---|---|---|---|
| 2nm front-end | 90,000 wafers | 110,000 by mid-2027 | Supply chain information in the Sept. 14 report |
| 3nm front-end | 180,000+ wafers | 210,000 by mid-2027 | Supply chain information in the Sept. 14 report |
| CoWoS | 130,000 wafers | 260,000 by end-2028 | Forecast by market participants in the Sept. 11 report |
Sources are the two Economic Daily News articles above. "kwpm" in the original CoWoS report was converted as thousand wafers per month. The same article also gives 3nm figures as "180,000" and "more than 210,000," so the values in the table are rounded figures from the reports, not exact endpoints. "More than 16%" is treated as the growth rate the paper presented, and I avoid recalculating finer rates from the rounded monthly figures.
The table shows that front-end processes are expected to see additions over a relatively short span, while CoWoS is expected to be built out over a longer period. Because everything starts from the end of 2026, the figures also do not represent growth from today. The size of a growth rate alone does not tell us which process's supply will become ample first.
3nm grows too, through equipment conversion and new fabs
Economic Daily News reports that TSMC is adding 3nm fabs in Taiwan, Arizona and Japan, and converting 5nm equipment in Taiwan to 3nm production. Investment in new plants and the redirecting of existing equipment to a different generation are proceeding in parallel. However, the monthly capacity outlook cannot be taken to mean that fabs in all three regions will be operating by mid-2027.
Even as 2nm ramps up, 3nm expansion continues. The paper explains that strong market demand remains for 3nm even though it is no longer the most advanced generation. The report suggests that customers' products will not all move to the new process at once, and that equipment must be allocated across demand for several generations.
This difference matters when tracking AI chip supply. More 2nm capacity does not translate directly into supply for customers who want 3nm products. Along with overall fab scale, one needs to check capacity for the process generations customers actually require.
Larger CoWoS in 2028 and the supply of finished products
In its technology symposium announcement in April, TSMC laid out plans to make CoWoS products themselves larger. The company said it is producing CoWoS at 5.5 times the reticle size and plans to introduce products at 14 times the reticle size in 2028. It said a single package could integrate about 10 large compute dies and 20 stacks of HBM (high-bandwidth memory).
A reticle is the master plate used to transfer circuit patterns during exposure in chip manufacturing. Here, "14 times" is a way of describing area scale and does not mean production capacity will rise 14-fold. CoWoS is a technology for densely connecting compute chips with memory and other components, and TSMC's plan points toward fitting more compute and memory into one package.
In other words, the number of wafers a plant can process and the number of finished products that can be supplied from them must be considered separately. If product size or configuration changes, the conversion to finished products changes even at the same monthly wafer count. There is no basis for directly linking the September forecast of doubled CoWoS capacity to a doubling of AI GPU shipments.
That said, the April plan for larger packages shows a new product option. It is not a statement that all CoWoS products will move to 14-reticle size in 2028. To measure how much larger packages might offset the effect of added capacity, we would need information on the product mix and actual yields of good units. The official announcement confirms the direction toward larger products; the forecast of 260,000 wafers per month should be viewed as less certain.
Economic Daily News' September 11 report said that, amid surging demand, customers are also qualifying supply chains other than TSMC's. Even as expansion proceeds, there are moves to secure the needed packaging capacity from multiple suppliers. Once equipment at each site is actually running and can reliably supply good units in the required process generation and package type, added monthly capacity will lead to more AI chip supply that customers can use.