Taiwan MOEA · 2026-10-20
On track
>$1.1TTaiwan export orders · by 2026
Taiwan export orders rise 62–67% y/y in October 2026; full-year orders exceed $1.1 trillion.
- Why
- AI servers, chips and memory; September orders a record $114.2B.
- What it means
- Confirms Q4 strength for Taiwan's supply chain (TSMC, Foxconn, Quanta, Delta).
- Status
- October data out ~Nov 20.
Short to 2028Litho capacity vs demand · by 2028
Planned 30% capacity increases in 2027 and again in 2028 will not be enough to meet customer demand for lithography.
- Why
- Memory and leading-edge logic customers all expanding at once.
- What it means
- Litho stays the gate on new fab capacity — supports ASML pricing and Zeiss-chain suppliers; caps how fast memory supply can grow.
- Status
- 2026 sales guided €44–45B.
Tight to 2028Memory supply/demand · by 2028
Memory supply stays tight through 2028; most of calendar-2027 HBM is already sold.
- Why
- HBM uses roughly three times the wafer area per bit of standard DRAM, so every HBM generation pulls capacity from conventional memory.
- What it means
- Higher-for-longer memory prices — good for MU, SK hynix, Samsung; a cost headwind for phone and PC makers.
- Status
- Korea chip exports +222% y/y in Oct 1–20.
TrendForce · 2026-09-30
Raised
+10–15%Contract price change · by Q4 2026
Q4 2026 DRAM contract prices rise 10–15% q/q and NAND 15–20%.
- Why
- AI-server orders keep supply tight; consumer demand softer.
- What it means
- Still rising, but slower than Q2/Q3 — watch whether Q1 2027 increases shrink further.
- Status
- Oct 20 update pointed to ~+20% for DRAM.
$230BAVGO AI revenue · by FY2028
Broadcom AI semiconductor revenue roughly doubles to ~$115B in FY2027 and again to ~$230B in FY2028.
- Why
- Custom-ASIC programs (Google TPU and others) plus AI networking; FY2026 guided to $58B.
- What it means
- Custom chips taking share of AI compute — positive for TSMC, HBM makers and optics regardless of GPU vs ASIC mix.
- Status
- Q4 FY26 AI guide $21.7B (+236% y/y).
NVIDIA · 2026-08-26
Watch
$40B/GW$ revenue per GW · by 2027
NVIDIA's revenue per gigawatt of data center rises from ~$18B (Hopper) to ~$25B (Grace Blackwell) to ~$40B (Vera Rubin).
- Why
- More silicon, memory and networking per watt with each generation.
- What it means
- Multiply by GW added per year (see SemiAnalysis: ~51 GW in 2027) for a top-down ceiling on AI chip spend.
- Status
- Vera Rubin ramps through 2027.
+18% '27WFE growth · by 2027
Bernstein: wafer-fab equipment grows 21.4% in 2026 and 18.2% in 2027, driven by memory. SEMI's Dec 2025 view was far lower (+9% / +7%).
- Why
- DRAM and NAND capex surging with memory prices.
- What it means
- If Bernstein is right, tool makers' 2027 numbers are too low — especially memory-heavy LRCX, TEL and AMAT.
- Status
- SEAJ raised Japan tool forecast to +26% in July.
~30%HBM share of DRAM wafers · by 2027
HBM rises from ~20% of DRAM wafer capacity in 2026 to near 30% in 2027; HBM market +58% in 2026 to ~$54.6B.
- Why
- HBM4 for Vera Rubin and custom ASICs.
- What it means
- Less wafer capacity left for PCs and phones — supports conventional DRAM prices too.
- Status
- Trade-press figure; verify against TrendForce.
Stock figures next to each card are 1-month price changes. Forecasts are summarised from the linked public sources; where only secondary coverage was available we say so. Status is our editorial judgement. Not investment advice.