Nvidia can design a faster processor. SK hynix can stack more memory. Neither one becomes an AI chip until somebody joins the two together, and for many of the industry's top chips that somebody is TSMC.

The joining step is called CoWoS. It rarely makes headlines, but it decides how many finished AI chips exist. And the numbers people quote about it are shakier than they look.

What is CoWoS packaging?

CoWoS stands for Chip on Wafer on Substrate. It's TSMC's method for placing an AI processor and its memory side by side on one thin slab of silicon.

That slab is called an interposer. It works like a street grid between the processor and its memory. It carries thousands of tiny wires so data can move between them quickly. The finished stack then sits on a base board, the substrate, that connects to the rest of the server.

TSMC describes CoWoS as 2.5D packaging that integrates processors and high-bandwidth memory stacks. It isn't new. TSMC says the platform has been in volume production since 2012. What changed is demand. Each Nvidia Rubin processor carries eight memory stacks. Every one of them connects through the interposer. Our coverage of the HBM4 memory squeeze explains why that memory is scarce too.

What are CoWoS-S, CoWoS-R and CoWoS-L?

They are three versions of the same idea. The difference is what the interposer is made of and how large it can grow.

TSMC's CoWoS family, in TSMC's own description
Version Interposer What TSMC says about it
CoWoS-S Solid silicon Up to 3.3 times the size of a standard chip exposure, about 2,700 square millimeters
CoWoS-R Polymer and copper wiring layers In volume production since 2023
CoWoS-L Wiring layers with small silicon bridges embedded First 3.5-times version in volume since 2024; TSMC keeps making it larger

The size limit matters. A single chip can only be printed so large in one exposure. AI chips now pack more processor area and more memory than one exposure allows, so the interposer has to grow. CoWoS-L is TSMC's route to bigger packages. Its page says the version keeps scaling the interposer "to integrate more silicon and memory."

Why is CoWoS a bottleneck for AI chips?

Because many top AI chips need it, and TSMC can package only so many wafers a month. A processor without packaging is a part, not a product.

TrendForce relayed Taiwan press reports in December 2025. They said both CoWoS-L and CoWoS-S were fully booked, with AI orders from Nvidia, Google, Amazon and MediaTek. That's the shape of the problem. Several of the biggest chip buyers are waiting on the same packaging lines.

The interposer itself is now a supply question too. On October 8, GlobalFoundries announced a $2 billion, multi-year deal to make silicon interposers for TSMC's CoWoS at its plant in Malta, New York. DatacenterDynamics reported that volume production is expected to start ramping in the first half of 2028. We covered what that timing means for 2027 buyers in our piece on the GlobalFoundries deal.