Building a Fab: Where the Money Goes Beyond the Equipment
Process tools account for 70–80% of the total investment in an advanced fab. The remaining 20–30% — building, cleanroom, ultrapure water, power, gases, wastewater, AMHS, land, and engineering design — is the so-called "construction cost". This page brings together the shares, unit prices, and reference projects that can be found in public sources into a parametric model that estimates cost by process node × capacity (WSPM), and states plainly which figures come from official sources, which are triangulated across multiple sources, and which cannot be found at all.
Parametric Estimator
Method: total investment = cost per 1k WSPM (process anchor) × capacity; non-equipment construction cost = total investment × facility share. All anchors are published figures from the "Node scaling" table below; the facility share uses the report's 20–30% range (the upper bound can be relaxed to Samsung Taylor's official 35% building share). This is a rough model, not a quotation.
The report gives no per-1k-WSPM cost anchor for 7nm, so no 7nm option is offered — do not fabricate data by interpolation. IBS figures are considered high by SemiWiki practitioners; for the same 2nm node there is also Samsung's published $340M figure, listed side by side for comparison.
Actual scale of the anchor projects: Kumamoto Fab 1 55k, IBS's 2nm model 50k, Yongin Y2 up to 360k at full build-out (six phases).
20–30% is the reported range for large advanced fabs (facilities $4–6B / total investment >$20B); 35% corresponds to Samsung Taylor's official building share. Mature-node fabs have a lower equipment share (in IMARC's split of an Indian mature fab, equipment is only 45%), so the facility share would be correspondingly higher — this slider's upper bound does not cover that case; judge such results separately.
Timeline basis: permitting + design → first wafer out (Blaschitz, 2025). The report gives no credible regional cost multiplier, so region does not enter the cost calculation; what is known is that building a fab in the US costs 30% (Intel basis) to 4× (TSMC basis) more than in Taiwan, and much of the premium comes from the longer schedule itself.
Model limitations (from the report): ① No parametric capacity curve (cost ∝ WSPMx) exists in public sources; this model extrapolates total cost linearly with capacity — but the facility portion actually scales sub-linearly (a UPW plant, substation, or CUB does not double with capacity), so the facility share will be underestimated for small fabs and overestimated for large ones; a phase-2 fab on the same campus shares infrastructure and costs significantly less. ② Land (~3–5%, $20M–200M, 100–500 acres; weak source quality) and engineering / permitting / commissioning (~3–8%) are not itemized separately and are contained within the uncertainty range of the facility share. ③ The ~+30% per-node rule of thumb (construction-physics) is driven mainly by more expensive tools and more mask / process steps, and lands on the equipment side.
Total Cost Breakdown
The table below is stitched together from multiple sources, and the bases are not consistent across rows (some are shares of total project cost, others of the construction budget) — cite any figure together with its note. A table on civilstructures.in lists "manufacturing equipment" at 35–50%, but its line items sum to over 100% and it describes a construction / installation budget, not total project cost; this page does not adopt it on a total-cost basis, citing it only where no better source exists for a facility line item, with a flag.
| Component | Share of total cost | Notes | Source |
|---|---|---|---|
| (a) Process tools | 70–80% | Advanced-node basis. Samsung Taylor's official announcement was 65% ($11B / $17B); IMARC's split for a mature fab (India) is only 45%. The equipment share has risen over time: mid-1980s DRAM fabs were roughly 50/50, and equipment has dominated since the late 1990s. Lithography alone is ~20% of total cost. | construction-physics (2024); Samsung official announcement (Nov 2021); imarcgroup.com |
| (b) Building + cleanroom | ~10–15% | An implied value backed out by subtracting the utilities portion from "total facilities 20–30%", not a directly published figure. Samsung Taylor's $6B (35%) is the official figure for building + site (incl. cleanroom), not fully comparable in basis. | Exyte/M+W practitioner via TechSpot (2025); Samsung official (2021) |
| (c) Utilities (UPW, power, HVAC, gases, chemicals, wastewater) | ~8–15% | Another basis: ~2/3 of total construction cost. In IMARC's split, utilities are 8% and cleanroom infrastructure (HVAC) 15% (mature-node basis). See the "Facility line items" table below for unit prices. | construction-physics (2024); imarcgroup.com |
| (d) Land | 3–5% | Roughly $20M–200M, 100–500 acres. Cases such as Kumamoto and Boise show land itself is a small share; site development is counted separately.Weak source | civilstructures.in (2026) |
| (e) Other (engineering design, permitting, commissioning) | ~3–8% | civilstructures basis: design 5–10%, test / commissioning 3%; IMARC basis: IT / security / compliance 5%. This is the weakest line item, with no strong industry source. | civilstructures.in; imarcgroup.com |
Node Scaling: Total Cost per 1k WSPM
On a consistent basis (total investment incl. equipment), node is the second-largest variable — the largest is actually logic vs. memory: advanced DRAM costs less than half per 1k WSPM ($69M) of mature 28nm logic ($156M). Rule of thumb: each node advance adds ~30% to total cost.
Units: millions of dollars per 1k WSPM. The 3nm figure is the range IBS gives ($15–20B @ 40k WSPM), shown as a range bar; the rest are single-point anchors. The early 5nm figure (TSMC Arizona 2020 announcement, $12B / 20k) includes a first-fab premium and does not represent mature 5nm levels.
| Node | $/1k WSPM | Basis | Source |
|---|---|---|---|
| Advanced DRAM | ~$69M | SK hynix Yongin Y2: $24.8B / 360k WSPM (full build-out, six phases) | fool.com / techtimes (2026) |
| 28nm-class (mature) | ~$156M | TSMC Kumamoto Fab 1 (22/28nm, 12/16nm): $8.6B / 55k WSPM, incl. $3.2B Japanese government subsidy | semiwiki / trendforce (2024–25) |
| 28–90nm (mature, general) | ~$125–275M | India target fabs: $5–11B each | fabs.pranaykotas.com India tracker (2026) |
| Mature specialty | ~$133M | GlobalFoundries Malta phase 1: $10B / ~75k WSPM (900k wafers / year) | medium / tai-semicon (2026) |
| 3nm | $375–500M | IBS: $15–20B @ 40k WSPMIBS figures considered high by practitioners | IBS via semiwiki / techspot (2014/2023) |
| 2nm (published basis) | ~$340M | Samsung Taylor: $17B / ~50k WSPM (announced in 2021 as 5nm; now 2nm GAA) | dmc-education / sammyguru |
| 2nm (IBS basis) | $560M | IBS: $28B @ 50k WSPM | techspot.com (Dec 2023) |
| 5nm (early adopter) | ~$600M | TSMC Arizona 2020 announcement: $12B / 20k WSPM | nasdaq.com (2020) |
| 3nm (small fab) | ~$1.1B | Kumamoto Fab 2: $17B / 15k WSPMNot confirmed by TSMC Small fabs have markedly higher unit facility costs — see "Capacity scaling" | datacenterdynamics citing Japanese local media (2026) |
Facility Line Items (the Core of Non-Equipment Cost)
Below are the line-item anchors found in the report. Note the enormous spread in cleanroom unit prices: base fit-out and "tool-ready" state differ by more than 4×, and EUV lithography bays double again — a unit price without a state condition is meaningless.
| Line item | Price / scale | Notes | Source |
|---|---|---|---|
| Cleanroom (ISO Class 1–3 ballroom) | Fit-out + hardware $1,200–2,500/m²; tool-ready $5,000–10,000/m²; EUV bays $20,000+/m²; another ISO Class 3 basis cites >$700/ft² (≈$7,500/m²) | Large fabs can have 500k–1M sq ft of cleanroom; a typical large fab is ~430k sq ft with 2,000 tools (Blaschitz) | civilstructures.in (2026); sst.semiconductor-digest.com; asianometry via construction-physics |
| Ultrapure water (UPW) system | Advanced node, 100k WSPM: $200–500M, capacity 10–20 MGD, ~2,000–5,000 gallons of water per wafer start, reuse systems add another $50–150M; mature node: $50–150M, 2–8 MGD | Hard anchor: Veolia's contract for Intel Chandler Fabs 52+62 was $120M, 6 MGD total, 18.2 MΩ·cm (Sep 2023) | semiconductorx.com; emergenresearch (Veolia contract) |
| Power | A large advanced fab draws on the order of 100 MW; TSMC Arizona at full build-out ~200 MW; 2nm uses 3.5× the energy of 28nm | Electrical infrastructure total $500M–2B (weak source quality); no standalone substation cost found. TSMC power intensity per wafer·mask-layer: 27.7 kW in 2022 → 40.5 kW in the 3nm ramp period (S&P basis; units as in the original source) | construction-physics (2024); sanieinstitute (2025); Deloitte; datacenterdynamics |
| HVAC / cleanroom air handling | No standalone unit price: included within "MEP services ≈ 2/3 of construction cost"; in small cleanroom projects HVAC can be 30–50% | Cleanroom air recirculates hundreds of times per hour, making HVAC one of the main drivers of facility energy use and cost | construction-physics (2024); crestmontcapital.com |
| Bulk gases (N₂/O₂/Ar) | On-site air separation unit (ASU): large cryogenic plants $50–300M; >500 t/day class $200–350M | Hard anchor: Linde's $1B Phoenix gas complex (2 new ASUs serving two other fabs, Jul 2026). A large logic fab uses ~50,000 Nm³/h of nitrogen | dataintelo (2026); openpr (Linde announcement); construction-physics |
| Chemicals + wastewater / exhaust treatment | No standalone cost found; only a combined water + wastewater basis of $200M–1B (incl. UPW) | TSMC Arizona has committed to 100% water reuse (semiconductorx) | civilstructures.in (envelope basis, weak quality) |
| AMHS (overhead transport + stockers) | 5–10% of construction cost; or 3–5% of equipment investment — i.e., ~$0.2–0.8B for a $20B fab | Both bases come from market research firms, not project disclosures | snsinsider.com; indexbox.io (2025–26) |
| Buildings (fab + subfab + CUB central utility building) | Total facilities $4–6B (for a fab with total investment >$20B, incl. cleanroom and utilities); Samsung Taylor building $6B / total $17B | Construction inputs for one large fab, order of magnitude: 30–40M labor-hours, 83,000 tons of steel, 785,000 cubic yards of concrete (Blaschitz) | Blaschitz/Exyte via TechSpot (2025); Samsung official (2021) |
Reference Projects (Published Figures)
All totals in this table are total investment including equipment — "building-only" figures are almost never disclosed, and Samsung Taylor's $6B building figure is the only clean official anchor. Projects without disclosed capacity are marked as such; nothing is inferred.
| Project | Node | Capacity (WSPM) | Published total investment | Year announced | Source |
|---|---|---|---|---|---|
| Samsung Taylor, Texas | 2nm GAA (Fab 1; announced as 5nm) | ~50k (trade reporting) | $17B | 2021 | dmc-education.com (official breakdown); actual regional spend including R&D and Austin ~$37B |
| TSMC Kumamoto Fab 1 (JASM) | 22/28nm, 12/16nm | 55k | $8.6B | 2021 | semiwiki.com; trendforce.com (incl. $3.2B Japanese subsidy) |
| TSMC Kumamoto, both fabs | 40/22/28/12/16/6/7nm; Fab 2 moving to 3nm | >100k combined; Fab 2 15k @3nm | >$20B | 2024 | ejinsight.com; datacenterdynamics.com; Fab 2's "$17B" is a local-media figureNot confirmed by TSMC |
| TSMC Arizona (all phases) | Phase 1 4nm, Phase 2 3nm, Phase 3 1.6/2nm | Phase 1 10k → planned 30k; Phases 2–3 not disclosed | $12B → $40B → $65B → $165B → $265B (successive increases) | 2020 / 2022 / 2024 / Mar 2025 / Jul 2026 | en.wikipedia.org/wiki/TSMC_Arizona; the $265B figure and "600k capacity by 2030" come from earnings-call reporting and low-quality aggregatorsTreat with caution |
| Intel Ohio One (two fabs) | Advanced logic (node not disclosed) | Not disclosed | $20B → $28B | 2022; raised Mar 2024 | sdxcentral.com (2025); ~$5–7B spent so far |
| SK hynix Yongin Y2 | Advanced DRAM / HBM | 360k (full, 6 phases × 60k) | ₩35.2T ≈ $24.8B | Aug 2026 (board approval) | fool.com; siliconangle.com; techtimes.com |
| SK hynix Cheongju M17 | NAND | Not disclosed | ₩19.1T ≈ $13.5B | 2026 | fool.com |
| Micron Boise (two DRAM fabs) | Advanced DRAM | Not disclosed | $50B (~$25B each) | 2022–2026 | intellectia.ai; industrialinfo.com |
| Micron Hiroshima expansion | DRAM / HBM | Not disclosed | $9.3B (¥1.5T) | 2026 | zerohedge.com (citing Bloomberg) |
| Micron Clay, New York (four DRAM fabs) | DRAM | Not disclosed | $100B (~20 years) | Jan 2026 (groundbreaking) | techtimes.com |
Capacity Scaling and Back-End (OSAT)
Capacity: equipment linear, facilities sub-linear
- Equipment scales roughly in proportion to wafer starts at a given node (double the capacity ≈ double the equipment), so the 70–80% equipment share broadly holds across fab sizes.
- Facilities show real but limited economies of scale: a UPW plant, ASU, substation, and CUB do not double with WSPM. Indirect evidence: Kumamoto Fab 1 at $156M per 1k WSPM, versus the unconfirmed Fab 2 at ~$1.1B/1k — small fabs carry far higher unit facility costs.
- A phase-2 fab on the same campus is significantly cheaper: shared power, water, logistics, permits, and talent pool (GF Malta case; both Kumamoto Fab 2 and Samsung Taylor Fab 2 sit on existing campuses).
- Part of the driver of giant fabs has nothing to do with cost: utilization headroom, fewer managers, faster yield learning, fewer financing rounds (chiphistory.org).
- Modeling suggestion: total cost = campus fixed items (land, substation, UPW, CUB — sub-linear) + cleanroom / facility items linear in WSPM + equipment (linear). No parametric curve (e.g., cost ∝ WSPM0.x) has ever been published in public sources — this is an explicit data gap.
OSAT / back-end packaging & test plants
- Reference anchor: Amkor's Arizona advanced packaging campus grew from a $2B plan to $7B (before the first building produced; broke ground 2025, production early 2028), covering advanced packaging + test, not traditional packaging alone (3dincites.com, Oct 2026).
- Cost drivers: high-end testers cost several million dollars each (wevolver.com); fine-pitch RDL lithography and precision bonding are the other two major capital items. 2025 bonding equipment market structure: wire bonding 38.5% ($3.54B), die bonding 31.2% (dataintelo).
- Equipment share is higher than in front-end fabs: ASE's 2025 ATM revenue mix was packaging 81%, test 18.5% (dataintelo), and test is precisely the capital-intensive bottleneck.
- Scale reference: Taiwan test house Sigurd raised 2025 capex to NT$8.8B (~$270M) for test capacity expansion, including a new plant (asiadaily.org).
- Model magnitudes: traditional packaging/test plant $0.5–2B; advanced packaging (CoWoS / fan-out) campus $2–7B. "Equipment ≈ 60–75% of OSAT capex" is an estimate backed out from cost drivers, not a directly sourced figure — cite it as such. No unit-capacity cost found (e.g., dollars per million units / month).
Timeline: Time Itself Is Cost
Typical durations for permitting + design → first wafer out (Blaschitz / Exyte·M+W, via TechSpot, 2025). Building a fab in the US costs 30% (Intel basis) to 4× (TSMC basis) more than in Taiwan, with a schedule roughly twice as long; a substantial part of the premium is "time-driven" — labor, extended site management costs, a ~12,000-person construction workforce (TSMC Arizona). Kumamoto Fab 1 compressed its schedule with 24/7 construction, but no public source quantifies the expediting premium (dollars per month saved) — not found.
Later stages
- Tool install → acceptable yield: a further 6–12 months of ramp (construction-physics, 2024).
- First wafer out → volume production: a further 6–18 months, depending on node maturity (e.g., VSMC risk production Sep 2026 → volume production Q1 2027).
- A large fab typically needs 2–5 years from groundbreaking to full capacity (e.g., SK hynix Y2 planned groundbreaking Jul 2027 → first cleanroom ready Jun 2029).
Where the regional gap actually is
SK hynix's schedule (groundbreaking → first cleanroom in just 23 months) shows cleanroom shell construction speeds are similar across regions; the US gap lies mainly in permitting, labor supply, and the tool-install / qualification ecosystem, not in pouring concrete. In other words, the first lever for lowering US fab costs is process and ecosystem, not building unit prices.
Unverified / Gaps in the Public Data
The following questions have no reliable public answer from this round of research. Any model claiming to answer them precisely should be asked for its sources.
- No authoritative breakdown table (in public channels): no SEMI / McKinsey / BCG / SIA table of "equipment / building / facilities / land / other" percentages was found; this page's breakdown is triangulated from second-hand sources of uneven quality.
- Capacity undisclosed by several makers: WSPM for Intel Ohio and Micron Boise / Hiroshima / New York is not published, so unit-capacity costs cannot be computed.
- No parametric capacity curve: no published cost ∝ WSPMx curve exists — only qualitative evidence.
- Standalone substation cost; standalone chemicals / wastewater cost: no standalone figures found for either — only combined bases.
- OSAT unit-capacity cost (e.g., dollars per million units / month): not found.
- Expediting premium (dollars per month saved): not found.
- Official equipment-vs-building splits have nearly vanished: treated as company secrets; Samsung Taylor's $6B / $11B is the only official announced split.
- Some 2026 news figures come from low-quality aggregators (e.g., TSMC's $265B total, Samsung's $37B regional spend): flagged in the tables above — treat with caution; the core anchors (Wikipedia, company announcements, practitioner interviews, IBS) are relatively stronger.
Key Sources
- construction-physics.com — How to build a $20 billion semiconductor fab: cost structure, equipment 70–80%, MEP ≈ 2/3 of construction cost, +30% per node, timelines
- TechSpot — Building fabs in the US takes twice as long and costs twice as much (Blaschitz / Exyte): facilities $4–6B / total >$20B; 20/38/34-month timelines; construction inputs
- TechSpot — IBS 2nm $28B / 50k WSPM
- SemiWiki forum — IBS 3nm $15–20B / 40k WSPM (incl. practitioner pushback that IBS runs high)
- Wikipedia — TSMC Arizona: capacity and investment by phase, CHIPS Act support
- dmc-education.com — Samsung Taylor $6B / $11B official split
- semiconductorx.com — Fab water use and UPW benchmarks
- emergenresearch — Veolia's $120M UPW contract for Intel
- SNS Insider — AMHS market report; IndexBox — AMHS share basis
- DataCenterDynamics — TSMC power intensity (S&P basis)
- Deloitte — 2nm uses 3.5× the energy of 28nm
- openPR — Linde's $1B Phoenix gas complex
- dataintelo — ASU costs $50–300M; dataintelo — OSAT market and revenue structure
- 3DInCites — Amkor advanced packaging campus $2B → $7B
- SemiWiki — TSMC Kumamoto; TrendForce — Kumamoto enters volume production; DCD — Kumamoto Fab 2 to run 3nm
- Motley Fool — SK hynix approves new memory fabs; TechTimes — Yongin 360k WSPM roadmap
- SDxCentral — Intel Ohio $28B
- IMARC — Mature-node fab cost breakdown
- civilstructures.in — Semiconductor fab construction (weak quality; cited only where no better source exists, and flagged)