CoWoS capacity, ABF substrates, and silicon wafers are the three bottlenecks that determine whether $650B in chip demand becomes revenue or backlog — and the market is mispricing all three.
TSMC's chairman said the quiet part out loud at the annual shareholders' meeting last week: capacity will not meet demand "for a long time." That was not a throwaway line. Morgan Stanley followed with a report framing advanced packaging — not transistor scaling — as the decisive factor in AI chip economics, projecting the AI semiconductor market at $753B by 2030. Meanwhile, Korea's semiconductor exports hit $25.2B in April, up 158% year-over-year, and SK Hynix announced plans to double DRAM wafer capacity to one million wafers per month by 2031. The numbers say the same thing from every angle: demand is not the constraint. Supply is.
What makes this cycle unusual is where the bottlenecks are forming. They are not at the transistor level — TSMC's N3 and N2 nodes are ramping on schedule. The chokepoints sit upstream and downstream: in the materials and packaging layers that the market has historically treated as commodity infrastructure. Three specific bottlenecks are now reshaping the profit map.
The first is advanced packaging. TSMC is racing to scale CoWoS from roughly 70,000 wafers per month in early 2025 to 130–140,000 by year-end 2026 — nearly a doubling. Yet C.C. Wei's "not enough" comment suggests even that pace trails demand. Foxconn (2317) announced a French joint venture targeting 50 million SiP components by 2033, signaling that packaging capacity is becoming a strategic asset worth building offshore. Hiwin (2049) is embedding Qualcomm edge-AI chips into panel-level packaging equipment — the tools themselves are getting smarter because the throughput pressure is that intense.
The second chokepoint is ABF substrates. Foreign brokers issued a rare consensus call this week: ABF supply is entering a fresh crunch, with lead times stretching to three quarters as AI GPUs and inference CPUs demand larger, 18–20-layer substrates. Glass substrate (TGV) is emerging as the successor technology for post-Rubin Ultra packaging beyond 9x reticle size, but 2026 is still a validation year — mass production could be three years away. The substrate trio — Unimicron, Nan Ya PCB, Kinsus — has order visibility through 2027. That kind of backlog does not happen in a balanced market.
The third is silicon wafers themselves. GlobalWafers (6488) reported that 12-inch wafers are running at full load, with even small-diameter wafers ramping back up, and the company is negotiating price hikes across all product lines. Shin-Etsu Chemical (4063), the world's largest silicon wafer maker, saw a US broker raise its target to ¥9,520, though a Japanese broker simultaneously downgraded to Neutral — a split that typically signals near-term pricing power is already in the stock but long-term capacity additions may cap margins. Winbond is planning a second clean room expansion because memory shortages have spread from DRAM to NAND and NOR Flash.
What Cuts Against
What to watch
- TSMC Q2 earnings (July 2026) — CoWoS wafer starts guidance and capex revision for advanced packaging.
- GlobalWafers Q2 revenue (July 2026) — whether 12-inch wafer price hikes flow through to margins.
- ABF substrate lead times (Q3 2026) — if 3-quarter backlogs persist or extend to 4 quarters on Rubin demand.
Sources
- [1]TSMC 2026 AGM — C.C. Wei: capacity won't meet demand for a long time— CoWoS capacity insufficient despite doubling
- [2]Morgan Stanley AI packaging report (June 2 2026)— AI semi market projected at 753B by 2030
- [3]Korea Customs HS 8542 April 2026— 25.2B exports, +158% YoY
- [4]SK Hynix capacity doubling plan (June 4)— DRAM wafer input to 1M wpm by 2031
- [5]ABF substrate broker consensus (June 7)— Lead times stretch to 3 quarters
- [6]GlobalWafers AGM (May 30)— 12-inch full load, price hike negotiations
- [7]Shin-Etsu Chemical split analyst ratings (June 5-8)— US broker PT raise vs JP broker downgrade
- [8]Formosa Chemicals COMPUTEX (June 3)— Semi materials target 30% of revenue by 2030
- [9]Foxconn France SiP JV (June 1)— 50M SiP components by 2033
- [10]Winbond clean room expansion (May 30)— Memory shortage spreads to NAND and NOR
If this analysis was helpful · ☕ Support Us · ✈️ Telegram