TSMC Is Reportedly Weighing a 20% Price Hike on CoWoS — the Step That Turns a Bare Chip Into an AI Accelerator. Fabricating the Chip Was Never the Bottleneck.
For two years the AI chip shortage got blamed on wafer fabs maxed out making dies. The part actually running out of room is the one after that — gluing a GPU die to its memory — and TSMC may now charge up to 20% more just to ration it.
The packaging backlog: why more chips don't mean more AI accelerators
Drag "packaging bays" to try different capacity scenarios, then hit Run to watch the backlog play out over a simulated 28 weeks. Illustrative model, not TSMC's real figures — see the note below for the assumptions.
The plain version
An AI chip like Nvidia's Blackwell isn't really one chip. It's a tiny, tightly packed neighborhood: a giant "brain" die surrounded by several stacks of ultra-fast memory (HBM) that it constantly reaches into, all glued onto a silicon base with microscopic wiring, which then gets mounted onto a circuit board. That gluing-and-wiring step is called packaging, and TSMC's version of it is called CoWoS — Chip-on-Wafer-on-Substrate.
For a while, the bottleneck in AI chips was simple: not enough raw silicon. TSMC and its rivals could only make so many wafers. But that's no longer the real limit. Today, TSMC can produce more brain dies and more memory stacks than it can actually stitch together into finished, shippable chips. Packaging — the fiddly, precision step that has to happen last — is now the ceiling.
Taiwanese media reported in late October that TSMC is weighing raising CoWoS prices by as much as 20%, and is even skipping its usual haggling with the companies that sell it packaging equipment, because that equipment is backordered for months. Think of a bakery that can mix unlimited batter and has unlimited ovens, but only a couple of people who can pipe the frosting — and decorating is the one step nobody can skip before the cake ships.
Why it matters for tech + supply chain: "GPU shortage" headlines for the next couple of years are really packaging-capacity headlines, and a 20% packaging surcharge flows straight into what every AI cloud and data-center builder pays per chip.
The expert version
CoWoS (Chip-on-Wafer-on-Substrate) is TSMC's 2.5D advanced-packaging platform: a logic die and several HBM stacks are placed onto a silicon interposer carrying through-silicon vias (TSVs) and redistribution layers, which is itself diced from its own wafer. The dies are reflow-bonded to the interposer at sub-10-micron alignment tolerances, underfilled, and the resulting module is mounted onto an organic substrate. Because that alignment step can't be parallelized the way lithography can, packaging throughput scales on a completely different curve than wafer-start throughput — which is why raw fab capacity stopped being the limiting variable for finished AI accelerators.
Analyst estimates of 2026 CoWoS capacity vary by source: Goldman Sachs has put it near 106,000 wafers a month, while Commercial Times-sourced analysts cite figures above 130,000, with industry plans targeting roughly 200,000 a month by 2027 as TSMC works to roughly double its prior run rate. Nvidia is estimated by various analysts to reserve somewhere between about 50% and 65% of available CoWoS capacity across Blackwell and the coming Rubin generation. Blackwell Ultra's interposer reportedly uses a reticle size around 3.3 times a standard reticle, adding further strain on packaging tools, and final test time is reported at roughly 4 times Hopper's for Blackwell packages, with Rubin potentially needing around 7 times as long.
Against that backdrop, Commercial Times and Economic Daily reported on October 27, 2026 that TSMC and its OSAT partners are weighing CoWoS price increases of up to 20%, while reportedly skipping customary price renegotiations with packaging-equipment suppliers; those suppliers' lead times have reportedly stretched from roughly 3–6 months to as long as 8. TSMC has not officially confirmed either the increase or a start date, and multiple outlets note the figure likely wouldn't apply uniformly across customers. Even accounting for planned expansion, at least one analyst estimate puts 2026 CoWoS demand outstripping supply by roughly 20–27%, and TSMC's Arizona packaging lines aren't expected to contribute meaningful volume before 2027 at the earliest.
Why it matters for tech + supply chain: A packaging-capacity ceiling, not a wafer-capacity ceiling, now sets the real production rate for AI accelerators — which means CapEx plans, cloud GPU pricing, and chip-collateral loan models all need to track CoWoS allocation, not just TSMC's advanced-node wafer starts.