Why Is Nvidia Always Sold Out? TSMC CoWoS Capacity Bottleneck Explained

Bottom line: The limiting step for Nvidia GPUs is not TSMC's 4nm wafers but CoWoS advanced packaging. Here are the published numbers behind the bottleneck.

What happened

• TSMC CoWoS capacity: ~70k wafers/month (2025) → ~120–140k by end-2026 (TrendForce; DIGITIMES cites ~127k).
• Nvidia reportedly holds 800–850k wafers for 2026, more than half of TSMC's CoWoS output; Broadcom ~200–240k is second, AMD third.
• Supply-demand gap: ~20% in 2026, possibly narrowing to ~10% by year-end and easing further in 2027.
• TSMC is now outsourcing the front-end CoW step to OSATs such as ASE and SPIL (TrendForce via ETNews).

Analogy: wafers are the flour; CoWoS is the machine that wraps dough and filling (HBM) into a finished bun. More flour does not help if the bun machine is the queue.

Point 1: Why CoWoS is mandatory

AI accelerators need multiple TB/s of memory bandwidth, which means HBM stacks millimetres from the GPU die, linked by tens of thousands of fine traces on a silicon interposer. Organic substrates cannot route that density, so the dies are first bonded on wafer (CoW), then mounted on substrate (WoS).

Point 2: Bigger interposers, fewer per wafer

Each GPU generation grows the package; CoWoS-L already spans several reticle sizes. Fewer interposers fit on a 300 mm wafer, so nearly doubling wafer starts does not double GPU output. That is why capacity headlines look huge while the market still reports shortages.

Point 3: HBM is contested too

Each package carries six to eight or more HBM stacks, and HBM supply is equally tight; reports say Micron's 2026 HBM output is sold out. A shortfall on either side stops the whole module.

Point 4: Relief valves

TSMC is expanding AP8 in the Southern Taiwan Science Park, converting part of AP7 in Chiayi to CoWoS, and shifting CoW work to OSATs. Non-TSMC capacity for end-2026 has reportedly been revised from ~26k to ~40k wafers/month. Longer term: larger CoWoS formats and panel-level packaging.

Open questions

• All capacity and allocation figures are supply-chain and analyst estimates; TSMC does not disclose customer allocations.
• If China-bound products like H200 ship, Nvidia orders could rise again and widen the gap.

Related videos

▶ You Think TSMC's CoWoS is Just One Technology? Think Again!

▶ The REAL Reason RAM Prices Are Exploding

Will the GPU shortage end in 2027, or will AI demand keep the gap open?

Join the discussion on our Facebook page: PLC on Facebook

The Automation Core｜theautomationcore.blogspot.com｜Videos: @TheEngineeringCore-v｜Sources: DIGITIMES, TrendForce, TrendForce/ETNews