Update: CoWoS YE Capacity to ~140k wpm — Gap Reprints ~20% as NVIDIA Locks >50%

What changed since the Aug WWSEMS Q1 update

In our August equipment-billings update we treated SEMI’s WWSEMS Q1 2026 print as the newest official flow vintage: global billings at a record $36.55 billion (+14% YoY), Taiwan +24%, Korea +16%, and mid-2026 packaging trackers that narrowed the CoWoS supply-demand gap from ~20% to about 10% as capacity printed ahead of the July research target. That post answered the billings pace question — is the Mid-Year $165.9B path showing up in quarterly dollars, and did packaging ease?

This Q3 note answers the next tracker print the market trades on: what moved when early-August CoWoS / OSAT notes and TSMC’s mid-August advanced-packaging disclosures raised year-end capacity toward ~140k wafers per month — while demand reservations reprinted a ~20% gap and put NVIDIA above half of 2026 CoWoS?

Three information events rewrite the dashboard. Capacity rises again: institutional and tracker midpoints now cluster near 140,000 wpm by late 2026 (versus 125k in the July research target and ~130k in the August mid-print), with some paths extending toward ~220k in 2027. Demand does not sit still: NVIDIA’s reported 800,000–850,000 wafer reservations for 2026 account for more than 50% of TSMC CoWoS, and aggregate demand trackers push the gap back toward ~20% even as capacity expands. Process news matters too — TSMC’s 5.5× CoWoS solution entered volume production with yields cited above 98% across some lines, while the company expands outsourcing of a key CoWoS front-end step to OSAT partners. The equipment ledger is unchanged as the official OEM path: SEMI Mid-Year still puts total equipment at $165.9B in 2026 (+23%), WFE at $143.9B, and DRAM equipment at $38.8B (+39%). Q1’s $36.55B remains the pace check, not a revision.

The dashboard above is built as a vintage delta — CoWoS capacity versus demand across research / Aug mid-print / Q3 tracker, NVIDIA reservation shares, Mid-Year segment bars, bottleneck tightness Δ and dual-vintage scatter, plus an upstream-dollars → downstream-gate stack with the Q1 AI-install regional signature. For the full equipment-cycle map see the July research post; for the earlier HBM/CoWoS bottleneck frame see AI GPU packaging & memory.

The headline table: Aug mid-print → Q3 tracker

The non-obvious cut is not “tools stopped arriving.” Upstream equipment dollars are still booming. The Q3 news is that packaging demand re-accelerated faster than the mid-print gap story, so a higher capacity path can coexist with a wider shortage print.

Capacity up, gap wider: the CoWoS paradox

July’s research vintage treated 125,000 wpm as the 2026 year-end CoWoS target against demand near 150,000 — a roughly 20% gap. The August billings update celebrated a mid-year print around 130k capacity versus 145k demand, compressing the gap toward 10%. Early-August tracker notes (TrendForce synthesis of OSAT outsourcing and customer reservations) and mid-August institutional estimates push year-end capacity toward 140k — a clear upward revision — but they also lift implied demand, reprinting a gap near 20%.

That is the paradox the dashboard’s composed chart is built to show. Bars for capacity rise; the amber demand area rises faster; the gap line goes back up. Markets that only watch capacity headlines will read relief. Markets that watch gap and customer lock-in will read continued rationing into 2H26. The 2027 path toward ~220k wpm is the relief valve — not the next two quarters.

Yield and footprint reinforce the same story from the process side. Volume production of larger-reticle CoWoS with >98% yields on some lines raises effective output per tool dollar. Ten advanced packaging facilities are a scale statement. Neither fact removes allocation — they change how much scarcity costs.

NVIDIA’s >50% lock and the OSAT overflow valve

The reservation panel is the concentration meter the prior update lacked. Reports that NVIDIA has reserved roughly 800k–850k wafers of TSMC CoWoS for 2026 — more than half of the year’s capacity — turn packaging from a generic “AI bottleneck” into a named customer-concentration risk. Remaining GPU, ASIC, and rising server-CPU CoWoS demand share the residual. When one buyer clears a majority of slots, secondary buyers inherit longer queues even if total capacity rises.

TSMC’s reported expansion of outsourcing a key CoWoS front-end step to OSAT partners is the overflow valve. Partner CoWoS-compatible capacity in the 50–60k wpm band (tracker estimate, not identical to TSMC process-of-record) absorbs spillover — and, in our editorial scores, re-tightens the OSAT layer because the overflow is no longer optional capacity; it is load-bearing. ASE / Amkor-class additions remain helpful but secondary to TSMC’s share of high-end CoWoS.

Equipment dollars still boom — DRAM +39% is the memory signature

None of the packaging drama revises SEMI’s Mid-Year OEM forecast. Total semiconductor manufacturing equipment is still projected at $165.9 billion in 2026 (+23.2% from the record $135.1B in 2025), with wafer fab equipment at $143.9B. Inside that path, DRAM equipment jumps 39% to $38.8B — the clearest AI-memory signature in the official OEM print — while NAND equipment rises about 31% to $13.9B. By 2028 SEMI still sees the total market near $229.5B.

Q1’s $36.55B global billings (+14% YoY, +1% QoQ) remain the only hard 2026 quarterly stamp against that path — about 22% of the full-year Mid-Year total on a simple share basis. Taiwan ($8.77B, +24% YoY) and Korea ($8.93B, +16%) still carry the AI install signature; China stays the largest absolute market at $10.99B but cooled 16% quarter-over-quarter. Simple ×4 annualization of Q1 would understate a typical H2 AI tool ramp — treat it as a pace check, not a cut to $165.9B.

Toggle the dashboard’s equipment panel: 2025→2026 grouped bars show the level jump; horizontal YoY bars show DRAM leading the growth ranking. The stack-flow panel then connects those dollars to the downstream gate — funded tools do not automatically become shipped GPUs when CoWoS and HBM allocate the last mile.

Bottlenecks: CoWoS re-tightens; HBM stays allocation-gated

Editorial tightness scores (1–10 composites, not SEMI metrics) move for the layers that the August trackers actually changed:

• CoWoS: 9 → 10. Capacity up, but gap and customer lock-in dominate.
• OSAT advanced package: 6 → 7. Outsourcing makes partners load-bearing.
• HBM / advanced DRAM: 9 → 9. Korea’s tool dollars help, but HBM remains allocation-gated into 2H26.
• EUV, test, substrates, leading-edge logic: broadly unchanged; logic remains less binding than packaging.

The dual-vintage scatter plots tightness against lead weeks with bubble size for supplier concentration. CoWoS and HBM still sit in the upper-right — long leads, high tightness — while logic fab sits lower. That geometry is why an equipment boom and a GPU shipment constraint can coexist for another year.

What would rewrite this vintage

1. WWSEMS Q2 2026 billings (not yet public in this window) confirm or break the Mid-Year $165.9B path after Q1’s $36.55B stamp.
2. CoWoS demand trackers revise NVIDIA / ASIC reservations materially below the 800–850k bandcollapsing the reprinted ~20% gap.
3. HBM supply from Korea eases allocation faster than packaging, flipping which layer binds first.
4. OSAT process qualification fails to absorb the outsourced front-end step at yieldtightening TSMC-only scarcity again.
5. SEMI’s next OEM forecast (year-end cycle) cuts WFE or DRAM equipment after the Mid-Year print.

Until those resolve, the shareable framing is narrow: capacity rose to ~140k wpm, but the gap reprinted near 20% because demand — led by a >50% NVIDIA lock — rose with it.

Caveats and methodology

• Official vs tracker. Equipment dollars are SEMI Mid-Year OEM forecast and WWSEMS Q1 billings. CoWoS capacity, demand, gap, and NVIDIA reservation figures are industry-tracker / disclosure synthesis (TrendForce, institutional notes, TSMC packaging commentary) and may revise.
• Gap arithmetic. Gap percentages are tracker midpoints (demand vs capacity), not a SEMI series. The Aug mid-print ~10% and Q3 ~20% can both be “true” for different demand assumptions in the same quarter.
• OSAT ≠ CoWoS POR. Overflow capacity is CoWoS-compatible advanced packaging, not identical to TSMC’s process of record.
• Bottleneck scores are editorial composites for visualizationdirectional, not survey microdata.
• Q1 annualization understates typical H2 AI tool intensity; do not treat 4×Q1 as a forecast.
• This post is explanatory data journalism, not investment advice.

Primary sources: SEMI, Mid-Year Total Semiconductor Equipment Forecast – OEM Perspective (July 14, 2026); SEMI, Q1 2026 Equipment Billings $36.55B (June 4, 2026); TrendForce, CoWoS OSAT outsourcing / NVIDIA reservation notes (Aug 5, 2026); TSMC advanced packaging disclosures (Aug 2026) for 5.5× yield / capacity path context.

The shareable takeaway

Versus the August WWSEMS Q1 update, the Q3 packaging vintage raises CoWoS year-end capacity toward ~140k wpm but reprints a ~20% gap as demand — with NVIDIA locking 800–850k wafers (>50%) — re-accelerates. SEMI’s Mid-Year equipment path stays $165.9B with DRAM tools +39%; Taiwan and Korea still print the AI install signature in Q1. Upstream dollars fund the stack; CoWoS and HBM still gate the ship.