Electron·Economics
Primary research · electroneconomics.substack.com ↗
Electron Economics · Source-Based Market Intelligence  ·  Last updated: Aug 2026

The AI Data Center Cost Stack —
three markets, three margin dynamics.

AI data center capex is not one cost curve. It is three stacked markets: GPU-heavy IT spend (~53% of all-in), margin-rich powered infrastructure (~15% of capex but 20–30% OEM margins), and low-margin civil delivery risk (~4%). The remaining ~29% is engineering, permitting, and commissioning — fragmented across specialists. The investment opportunity is not where the dollars are largest. It is where scarcity, margin, and contractual leverage overlap.

Layer 1 · IT Equipment
~$17.50/W
GPU servers, networking, storage. All-in basis. ~53% of $32.75/W midpoint (46–64% across source range). Tenant-owned in colo.
~$1,650M per 100 MW
Layer 2 · Powered Shell
~$4–5/W
UPS, switchgear, cooling, generators. Facility-only basis. OEM margin pool.
~$400–500M per 100 MW
Layer 3 · Physical Shell
~$1–1.5/W
Civil structure, envelope, fit-out. Smallest % share vs non-AI equivalent.
~$100–150M per 100 MW
Gap · Design / Contingency
~$9.50/W
Engineering, permitting, contingency, commissioning, logistics. Not supplier revenue.
~$950M per 100 MW · 29% of all-in
All-in $/W — IT equip + facility + gap. Layer 1 uses this basis. Facility-only $/W — shell + MEP only, excl. IT equipment. Layers 2–3 use this basis. Shell-only $/W — civil + powered shell; what colo operators pay (no IT fit-out) Tenant fit-out — IT equipment only; JLL: up to $25M/MW on top of shell in colo AI $M / 100 MW = $/W × 100 — always
View mode:
Executive = 8-row summary · Component Detail = full breakdown with sources
Layer / Component AI all-in $/W $M / 100 MW % of all-in OEM margin Lead time Signal
Layer 1 — IT Equipment  · All-in $/W basis · tenant-owned in colo
GPU / AI servers + networking + storage
Accelerators, HBM, NVLink, interconnect, NVMe
~$17.50/W ~$1,650M ~53%
midpoint; 46–64% range
Nvidia captures GPU
pass-through for operator
~4–16 wk Dominates stack
Layer 2 — Powered Shell  · Facility-only $/W · OEM margin pool
Electrical back-end
UPS, switchgear, transformers, generators, substation
~$2.1/W ~$210M ~6.4% 22–30%
Eaton 30% · Vertiv 22.3%
40–80 wk
switchgear; 18+ mo genset
Lead time risk
Cooling
Chillers, CDU/liquid cooling, CRAH (air-cooled only), towers, pumps
~$2.3/W ~$230M ~7.0% 22–25%
Trane ~25% · Vertiv 22%
18–30 wk
chillers; CDU varies
AI-mandatory CDU
White space fit-out
Server racks, PDUs, busbars, cabling
~$0.6/W ~$60M ~1.8% 20–30%
Legrand 20.7% · Eaton 30%
6–12 wk Stable supply
Powered shell subtotal ~$4–5/W ~$400–500M ~15% 22–30%
Layer 3 — Physical Shell  · Facility-only $/W · EPC margin 4–8%
Civil structure + envelope + fit-out + site works
Steel frame, cladding, earthworks, GC prelims
~$1–1.5/W ~$100–150M ~4% 4–8%
EPC/GC margin
Parallel to MEP Smallest share in AI
Gap / Overhead  · ~$9.50/W · 29% of all-in · not supplier revenue
Design, engineering, contingency, commissioning, logistics
Fragmented across specialists: engineering firms, commissioning agents, owner's reps, permitting consultants — low-scalability margin pool
~$9.50/W ~$950M ~29% — — Fragmented margin pool
Total — all-in AI basis ~$27.50–38/W ~$2,750–3,800M 100% —
▸ Full source registry & methodology notes

Primary sources

Epoch AI (May 2026) — 1 GW AI facility $38B upfront capex; servers 60% of TCO annualised. epoch.ai ↗

Lean Research (Mar 2026) — Infrastructure layer split: gray space 29%, white space 10%, cooling 32% of infra; UPS 16%, switchgear 11%, transformers ~2%, chillers 9%, CRAHs 12%; AI vendor revenue $3.1–3.5M/MW vs $1.2M/MW non-AI; shell 7% / $1.90/W (400 MW AI).

Turner & Townsend DCCI 2025 — Cost category split: GC/prelims 10%, shell & core 14%→9%, mechanical 22%→33%, electrical 54%→48% (air-cooled → liquid-cooled AI). Liquid cooling premium +7–10%. Lead time data. turnerandtownsend.com ↗

JLL Global Data Center Outlook 2026 — Non-AI shell+core $11.3M/MW global avg; 7% CAGR 2020–2025; avg lead time 33 wks (+50% from 2020).

DLR 10-K 2024 — Colo pipeline ~$9.5M/MW; premium markets ~$12M/MW.

McKinsey (2026) — Switchgear: 80 wk hyperscale lead. Transformer: 50 wk lead.

Dell'Oro Group (Jan 2026) — Liquid cooling market: "nearly doubled in 2025, approaching $3B"; ~$7B by 2029.

Earnings filings used as supplier proxies

Eaton 8-K (Feb 26 2026) — Electrical Americas Q4 2025: $3.5B revenue (+21% YoY, +15% organic); 29.8% op margin (Q1 2025 record: 30.0%); DC orders +200% Q4 2025; $15.3B backlog.

Vertiv 8-K (Oct 22 2025) — Q3 2025: $2.676B revenue (+29% YoY); $9.5B backlog; book-to-bill 1.4x; 22.3% adj op margin (+220 bps); orders +60% YoY.

Cummins 8-K (Feb 5 2026) — Power Systems FY2025: $7.5B revenue (+16%); 22.7% EBITDA (+430 bps); DC revenue $3.5B/yr run rate; order book to 2028; $150M MN expansion.

Trane Technologies 8-K (Jan 29 2026) — Americas CHVAC applied bookings +120% Q4 2025 (2nd consecutive >100%); $7.8B backlog; Stellar Energy acquired.

ABB Q4 2025 press release (Jan 29 2026) — Electrification Q1 2025: $3.825B revenue, 23.2% op EBITA; $8.69B record backlog; DC "very strong double-digit" orders Q4; 800VDC partnership with Nvidia.

Legrand FY2025 press release (Feb 12 2026) — Total €9.48B (+9.6%); DC revenue ~€2.4B (26% of group); +~40% organic DC growth; 20.7% adj op margin; DC CAGR +19% since 2019.

Arista Networks Q4 FY2025 (Feb 12 2026) — Implied FY2025 ~$9B revenue (Q4 $2.488B +28.9% YoY); GAAP gross margin 63–65%.

Schneider Electric FY2025 (Feb 26 2026) — Energy Management €33.1B (+10.3% organic); DC demand "triple-digit YoY" in Q4 2025. Secure Power (UPS+PDU) not separately disclosed.

EE channel checks — Eaton AlphaSense expert sessions (EE, May 2026): enterprise architect (5–8→30–50 kW rack density; Eaton Brightlayer BMS); EPC VP (40–60 wk enterprise switchgear; modular construction −20–30% schedule). EE Centerview BTM session (May 2026): Permian 105–115°F; LSTK vs GMP contract structure. Channel check data marked EE↑ in source buttons.

Confidence system

⬤ Primary Single primary document directly states the figure.

◎ EE est. Cross-triangulated from multiple sources; not single-primary-sourced. Treat as directional.

○ EE↑ Previously EE est.; upgraded using named EE research session evidence.

Third-party market research firm TAM estimates (MarketsandMarkets, Mordor Intelligence, GMInsights, etc.) excluded — methodology not disclosed, figures not reproducible. Not investment advice.

Electron Economics · Source-Based Market Intelligence  ·  Last updated: Aug 2026

Build Cost Vintage Trend

Global average shell & core construction cost for non-AI air-cooled hyperscale (30–50 MW IT load). 7% CAGR 2020–2025. AI liquid-cooled premium: add +7–10%. Source: JLL Global Data Center Outlook (annual); Turner & Townsend DCCI 2025.

2020
$7.7M/MW
baseline
2021
~$8.3M/MW
+~8%
2022
~$8.9M/MW
+15% (T&T)
2023
~$9.5M/MW
+~7%
2024
~$10.1M/MW
+9% (T&T)
2025
$10.7M/MW
+5.5% (T&T)
2026E
$11.3M/MW
+6% (JLL)
7% CAGR 2020–2025 (JLL). From $7.7M to $10.7M/MW. Driven by construction labour inflation (+15% peak in 2022), equipment escalation (transformers, switchgear), and increasing MEP density requirements. 2025 showed deceleration (+5.5% vs +9% in 2024) — T&T notes sector-wide construction inflation of 4.2% in 2025. AI premium (not in JLL/T&T non-AI index): liquid-cooled AI facilities carry +7–10% construction premium vs equivalent air-cooled (T&T 2025). All-in with IT equipment: $27.50–38/W (Lean Research / Epoch AI).
Lead time trajectory (JLL 2026): Average equipment lead time now 33 weeks globally — up 50% from pre-2020 levels. More than half of 2025 projects experienced 3+ month construction delays. Cost inflation is decelerating; schedule risk is not improving. Switchgear: 80 wk hyperscale (McKinsey). Transformers: 50 wk (McKinsey). Generators >2 MW: 18+ months, order book to 2028 (Cummins Q4 2025 8-K). The binding constraint is electrical procurement, not civil works.
Component / metric 20202022 20242025 2026E YoY trend Source
Shell & core (non-AI global avg)
$7.7M/MW~$8.9M/MW ~$10.1M/MW$10.7M/MW $11.3M/MW +7% CAGR JLL 2026; T&T DCCI 2025. 2022: T&T +15% peak. 2024: T&T +9%. 2025: T&T +5.5% (decelerating). 2021/2023 interpolated at 7% CAGR — EE est.
Avg equipment lead time
All categories, global
~22 wks~26 wks ~30 wks33 wks 33–36 wks +50% from 2020 JLL 2026: "average equipment lead times reaching 33 weeks, a 50% increase from 2020." Prior years: EE est.
Switchgear (hyperscale)
~40 wks~55 wks ~75 wks80 wks 80+ wks Worsening McKinsey (2026): 80 wk hyperscale. EE↑ Eaton AlphaSense EPC VP: 40–60 wk enterprise (EE, May 2026). Prior years: EE est.
Transformer (MV, dry-type)
~30 wks~38 wks ~48 wks50 wks 50–55 wks Worsening McKinsey (2026). GOES steel = binding physical constraint. Korean challengers (HD Hyundai, Hyosung) gaining share. Prior years: EE est.
Generator >2 MW
~9 mo~12 mo ~16 mo18+ mo 18–24 mo Worsening Cummins Q4 2025 8-K: order book to 2028; $150M MN expansion. Prior years: EE est.
Liquid cooling CDU market
Global revenue
~$0.4B~$0.8B ~$1.8B~$3B ~$4B Accelerating Dell'Oro Group (Jan 2026): "nearly doubled in 2025, approaching $3B"; forecast ~$7B by 2029. 2020–2024: EE est. interpolated.
T&T construction cost inflation YoY
Non-AI air-cooled DC
—+15% +9%+5.5% +5–6%E Decelerating T&T DCCI 2025: "+5.5% increase in cost per watt, markedly lower than 9.0% in 2024." Broader construction: +4.2% (T&T GCMI 2025). 60% of respondents expect +5–15% in 2026.
Sources: JLL Global Data Centre Outlook (2026) · Turner & Townsend DCCI 2025 · McKinsey (2026) · Cummins Q4 2025 8-K · Dell'Oro Group (Jan 2026). Years 2021, 2023 interpolated at 7% CAGR from JLL anchors — EE est. Not investment advice. electroneconomics.substack.com
Electron Economics · Source-Based Market Intelligence  ·  Last updated: Aug 2026 · T&T DCCI 2025 · v16 QA pass

Geography — $/W by Market

Shell & core construction cost per watt, air-cooled hyperscale 30–50 MW IT load. Add +7–10% for liquid-cooled AI. Source: Turner & Townsend Data Centre Construction Cost Index 2025 (primary; FX avg Oct 2024–Oct 2025). CBRE demand signals for context only. Click column headers to sort.

How to use for underwriting: Take the $/W below as your shell & core baseline. Add +7–10% for liquid-cooled AI (T&T 2025). The AI all-in estimate column (shell + $17.50/W IT + $9.50/W gap) is a screening figure only — the gap/overhead does not stay fixed across Tokyo, Singapore, Phoenix, and NoVA. In-market engineering cost, permitting complexity, labour rates, and logistics premiums vary substantially. Use the AI all-in column for relative ranking, not for project underwriting. For site-specific underwriting, adjust using local labour indices and apply T&T DCCI market differentials. Colo operators pay shell cost only — JLL notes AI tenant fit-out adds up to $25M/MW on top of shell.
View:
Market Shell $/W
T&T 2025 baseline
vs NoVA
% delta
AI all-in est.
shell+IT+contingency
Demand
signal
Market view
underwriting signal
Key constraint
Key underwriting read-through: The Cost Build tab's NoVA/DFW baseline understates costs in Tokyo (+34%), Singapore (+28%), Zurich (+26%), and Silicon Valley (+18%). It overstates in Charlotte (−16%), Columbus (−13%), Phoenix (−13%). For cross-market deal comparison, apply the delta from this table before benchmarking $/W assumptions.
Sources: Turner & Townsend Data Centre Construction Cost Index 2025 (52 markets; shell & core only; air-cooled 30–50 MW IT; FX avg Oct 2024–Oct 2025). CBRE Global Data Center Trends 2025 (demand / pricing signals). JLL 2026 Global Data Center Outlook. EE Centerview BTM session (May 2026) — Permian temperature data. AI all-in estimate = shell $/W + $17.50/W IT equipment + $9.50/W contingency/engineering — see Cost Build tab for full decomposition. Not investment advice.
Electron Economics · Source-Based Market Intelligence  ·  Last updated: Aug 2026

The powered shell is ~15% of capex —
but where 20–30% OEM margins live.

Equipment OEMs serving data center infrastructure earn manufacturer-grade margins (20–30%) on products sold into projects where EPC contractors earn 4–8%. The powered shell (~$4–5/W) is the investable layer — small in $/W but high in margin density. All figures from public earnings filings.

⬤ OEM Operating Margins — from Public Earnings Filings (2025–2026)

Compare to EPC / GC margin: 4–8% (EE Centerview BTM session, May 2026; CoVolt/EPC analysis). The gap between OEM margin and EPC margin is structural — OEMs hold no delivery risk after purchase order; EPCs absorb schedule delay; developers warehouse financing cost on deferred commissioning.

Eaton · Electrical Americas
27.5%
Electrical Americas Q2 2026 op. margin (+190 bps sequentially from Q1's 25.6%; substantially above 22.6–23.0% guided)
EA Q2: record $4.0B revenue (+18% organic) · DC orders +85% · DC revenue +65% vs Q2 2025 · Electrical backlog +103% YoY
Boyd Thermal: $1.8B FY2026 raised; $2.5B Electrical Global Q2 (+44% total, +25% Boyd contribution) — first full quarter post-acq.
Mobility: Reverse Morris Trust with Dana; $1.1B cash distribution to Eaton pre-close; Q1 2027 target — accretive to growth & margins
FY2026 raised: adj. EPS $13.40–$13.60; organic growth 11–13%; segment margins 24.1–24.5%
Eaton Q2 2026 8-K, Jul 30 2026
Vertiv · Total Company
22.6%
Adj. operating · Q2 2026 (+410 bps YoY) — above guidance of 20.7–21.7%; Q1 was 20.8%
Q2 revenue $3.27B (+24%); Americas +29%; APAC +29%; EMEA returned to growth +2%
Backlog $15B+ · adj. EPS $1.52 (+60% YoY) · FCF $925M (+234% YoY)
FY2026 raised: net sales ~$14.0B · organic growth 31% · adj. op. margin 23.8% · adj. EPS ~$6.70
ThermoKey acq. closed Jun 12 2026 · 2030 targets: 20–22% revenue CAGR, ~27% adj. op. margin
Vertiv Q2 2026 8-K, Jul 29 2026
Cummins · Power Systems
29.5%
Power Systems EBITDA record · Q1 29.5% · Q2 segment margin expanding YoY (only segment to do so)
Power Systems Q2: record $2.3B (+19% YoY) · total company Q2 revenue record $9.46B (+9%)
FY2026 total company raised: revenue +10–13% (was 8–11%); EBITDA 18.0–18.5%
Power generation order book extends into late 2028; capacity additions underway
$50B 2030 revenue target · low-pressure fuel cell sold to Alstom Q1 2026
Cummins Q2 2026, Aug 4 2026
Trane · Americas CHVAC
~25%+
Adj. operating · applied bookings +130% Q2 2026 (4th consecutive >100%; 2-yr stack 4×)
Record $12.1B backlog (+70% YoY) · $6B committed 2027+ · book-to-bill 123% across all segments
Americas CHVAC bookings all-time high +50%; enterprise organic bookings +37%
Stellar Energy integrating well; management focused on capacity for liquid cooling & modular chiller
FY2026 raised: ~9% organic growth · adj. EPS $15.20–$15.30
Trane Q2 2026 earnings call, Jul 30 2026
ABB · Electrification
20.2%
Op. EBITA · Q2 2026 (+90 bps YoY) — Q1 was 23.5% (step-down mix/investment-driven, not demand)
Q2 record orders $12.0B (+30%) · revenues $9.5B (+14%) · record backlog $30B (+28% YoY)
Americas orders +53% (US +62%) · book-to-bill 1.27 · FCF $881M · ROCE 28.4%
Rotork acquisition: ~$5.5B — expands automation into valve actuators; margin-accretive yr 2
FY2026 guidance: low double-digit to low-teens comparable revenue growth; margin improving YoY
ABB Q2 2026 results, Jul 16 2026
Legrand · Data Center
20.7%
Adj. op. margin · H1 2026 (20.8%) · H1 sales €5.4B (+13.1% YoY; +17% excl. currency)
DC revenue >30% organic growth in H1 · DC now 32% of total revenue (first time disclosed)
US revenue +26.7% H1 (+24.5% in Q2) — DC driving North America outperformance
7 acquisitions in 2026: Keydak + TES (Q1) · SRS Power Engineering Malaysia (€90M) · Girtz Industries US + others
FY2026 guidance raised: +16–19% sales growth (was +10–15%)
Legrand H1 2026 results, Jul 28 2026
Schneider · Energy Mgmt
~19%
Adj. EBITA guidance FY2026: 19.1–19.4% (group level)
Q1 2026: €9.77B revenue (record); EM €8B (+12.8% organic)
DC double-digit gains Q1; Secure Power (UPS+PDU) not separately disclosed
Schneider Electric Q1 2026 results, Apr 30 2026
⬤ EPC / GC margin for comparison: 4–8% — EE Centerview BTM diligence session (May 2026); CoVolt/EPC analysis. EPCs absorb delivery schedule risk; OEMs collect margin and move to next order. Note: Eaton EA Q2 2026 operating margin 27.5% — substantially above 22.6–23.0% guided; capacity ramp converting to margin. Mobility separation: Reverse Morris Trust with Dana; $1.1B cash to Eaton pre-close; Q1 2027. Vertiv Q2 2026: 22.6% adj. operating margin — above high end of 20.7–21.7% guidance; 2030 target ~27%. Cummins Power Systems: 29.5% Q1 EBITDA record; Q2 PS revenue record $2.3B. ABB Q2 2026: 20.2% EBITA (down from Q1 23.5% — mix/investment, not demand); Rotork ~$5.5B acquisition announced. Schneider Electric: FY2026 group-level adj. EBITA guidance 19.1–19.4% (Secure Power not separately disclosed). Arista Networks: GAAP gross margin ~63–65% FY2025.
The architecture transitions to watch:
800VDC: Schneider (800VDC sidecar, Apr 2025) and ABB (Nvidia partnership, Jan 2026) are both moving toward 800VDC architecture. Only 15–25% of facilities on 800VDC by 2030 (Lean Research). Direction is set — UPS content risk is real but slow-moving.
CRAH → CDU (accelerating): CRAH is 12% of infra in non-AI builds; physically obsolete above 41 kW/rack. Vertiv, Schneider, and now Eaton (acquired Boyd Thermal, Mar 2026 — liquid cooling; Boyd Q1 revenue 2×+ YoY, backlog doubled in 6 months) are all competing in CDU. Trane's Stellar Energy ($1B backlog; CEO: "$1B business in 2–3 years"). CDU market ~$3B 2025 (Dell'Oro). The CDU competitive set is now wider than ever — Vertiv, Schneider, Eaton, Trane, JCI, plus specialists (Submer, LiquidStack, GRC).
Cummins fuel cell: Cummins sold its low-pressure fuel cell business to Alstom (Q1 2026). Bloom Energy remains the BTM fuel cell wildcard — Cummins is now a pure generator play.
Bottom line: Backlogs are at record levels across every OEM — Vertiv $15B+, Trane $10.7B, Eaton EA $14.5B. Cummins Power Systems hit a record 29.5% EBITDA in Q1 2026. Eaton's Q1 margin dip (25.6% vs 30% prior record) is a ramp-cost story, not a demand story. The question is whether these margins are structural or scarcity rents — GOES steel and generator capacity constraints still point to 2027+ before normalisation begins. Watch Boyd Thermal integration progress and Korean transformer challenger share gains as the leading indicators.
Component Market leaders Challengers / disruption Market proxy (filing-based) OEM margin AI architecture signal
Layer 1 — IT Equipment
GPU / AI servers
Nvidia (~80%+ AI GPU), Supermicro (GB200), Dell, HPE AMD, Intel; ODM trend (Meta, Google) Nvidia DC: ~$44B annualised (Q1 FY2026 $13.5B ×4) — Pass-through; facility operator does not capture margin
Networking
Nvidia/Mellanox (IB), Arista, Broadcom Cisco, Juniper; Ethernet AI fabric challenge to IB Arista FY2025 ~$9B (Q4 $2.488B +28.9% YoY) 63–65%
Arista GAAP GM
IB dominance at 5-yr risk; Broadcom Ethernet gaining
Layer 2a — Electrical (Gray Space)
UPS systems
Schneider, Vertiv, Eaton ABB, Huawei, Legrand (Keor) Vertiv Critical Power ~$5.5B; Schneider EM €33.1B (Secure Power unsplit) 22–30%
Eaton 30% / Vertiv 22.3%
⚠ 800VDC content risk — Schneider + ABB both partnering with Nvidia on 800VDC architecture
Switchgear
Schneider, Eaton, ABB Siemens, Powell Industries ABB Electrification Q1 2025: $3.825B rev; Eaton EA $13.4B FY2025 (all electrical) 23–30%
ABB 23.2% / Eaton 30%
80 wk lead times; GOES steel bottleneck on transformers; Korean challengers gaining share
Backup generators
Caterpillar, Cummins mtu, Generac; Bloom Energy (fuel cell BTM) Cummins Power Systems FY2025: $7.5B revenue; DC $3.5B/yr run rate 22.7%
Cummins EBITDA
Order book to 2028; 18+ mo lead; Bloom Energy fuel cell emerging for BTM prime power
Layer 2b — White Space
PDUs / busways
Legrand (Starline 6,000A), Eaton, Vertiv nVent (Enlogic), Schneider Legrand DC ~€2.4B FY2025; Q1 2026: +18% total sales, 20.7% margin; Keydak + TES acquisitions 20.7%
Legrand adj. op.
Busway displacing fixed copper (GPU refresh reconfigurability); Vertiv CoolChip Max hybrid CDU-PDU (Feb 2026)
Layer 2c — Cooling
Chillers
Carrier, Trane Technologies, JCI Daikin, Airedale, Stulz Trane Q4 2025: $7.8B backlog; applied bookings +120% Q4 2025 ~25%+
Trane Americas CHVAC
Trane DSX to 145°F ambient. JCI offloaded residential HVAC to Bosch ($8.1B) to focus on DC cooling
CRAH / CRAC units
Vertiv, Schneider, Stulz Airedale, Rittal Vertiv thermal ~$3.5B est. (proxy: Vertiv $10.2B × ~35%) ~22%
at risk
⚠ Most structurally challenged category. Physics ceiling at 41 kW/rack. Vertiv + Schneider pivoting to CDU to protect revenue
Liquid cooling — CDUs
Vertiv, Schneider, Eaton (Boyd Thermal, acq. Mar 2026), JCI (Silent-Aire) AAON/BASX ($2.13B backlog, +107% YoY; liquid cooling +187% TTM), Trane (Stellar Energy), Submer, LiquidStack, GRC, Asetek ~$3B (2025, Dell'Oro); ~$7B by 2029 ~22–25%
incumbent proxy
Fastest-growing category. AAON/BASX (specialist): $2.13B backlog (+107% YoY), BASX liquid cooling +186.8% TTM, book-to-bill >2× for 4 consecutive quarters, FY2026 revenue +40–45% — fastest organic growth rate in segment (AAON Q1 2026 8-K, May 7 2026). Eaton acquired Boyd Thermal (Mar 2026) — liquid cooling; Boyd Q1 2026 revenue 2×+ YoY, backlog doubled in 6 months, adds $1.7B+ to Eaton 2026 revenue. Trane Stellar Energy: $1B backlog; CEO: "$1B business in 2–3 years." JCI Silent-Aire CDU (Sep 2025). Vertiv CoolChip Max hybrid CDU-PDU (Feb 2026). Competitive set now 5+ major incumbents plus specialists — OCP standardisation lowers specialist barrier.
Layer 3 — Civil / EPC
EPC / GC
Turner, Kiewit, AECOM Skanska, Jacobs, Matrix Service ENR Top 400 DC category (no precise TAM) 4–8%
GC/EPC margin
Absorbs schedule delay from long-lead electrical equipment. LSTK on power plant side; GMP on data hall
Margin data from public earnings filings only: Eaton 8-K (Feb 2026) · Vertiv 8-K (Oct 2025) · Cummins 8-K (Feb 2026) · Trane 8-K (Jan 2026) · ABB Q4 2025 press release (Jan 2026) · Legrand FY2025 (Feb 2026) · Arista Q4 FY2025 (Feb 2026) · Schneider FY2025 (Feb 2026). Market sizing from public company revenue disclosures and Dell'Oro Group (Jan 2026). Third-party market research firm TAM estimates excluded. Not investment advice. electroneconomics.substack.com
Electron Economics · Source-Based Market Intelligence  ·  Last updated: Aug 2026

AI Infrastructure — 12 Charts

All charts derived from primary sources in the Cost Build, OEM Margins, Vintage Trend, Geography, and Demand tabs. Hover any element for data detail.

1 — Capex Flow & Margin Pool
$32.75/W all-in AI capex decomposed. Bar width = $/W; bar below = OEM margin. Powered shell = 15% of capex, 22–30% OEM margins.
The powered shell (~$4.50/W) carries 22–30% OEM margins on 15% of all-in capex. IT equipment is the dominant spend but a pass-through — facility operators do not capture GPU margin.
2 — T&T Cost Category Shift (Air → Liquid AI)
Mechanical share nearly doubles (22%→33%). Electrical shrinks (54%→48%). Shell falls (14%→9%). Source: T&T DCCI 2025.
Mechanical share nearly doubles (22%→33%) when moving from air-cooled to liquid-cooled AI. Shell & core shrinks (14%→9%). This is the structural shift driving CDU demand and CRAH displacement.
3 — OEM Margin vs EPC Gap
Equipment OEMs earn 20–30% on products sold into projects where EPCs earn 4–8%. Q1 2026 most recent. All from public filings.
Equipment OEMs earn 20–30% operating margins on products sold into projects where EPCs earn 4–8%. The gap is structural: OEMs hold no delivery risk post-PO; EPCs absorb schedule delay.
4 — OEM Order Backlog (Q1 2026)
Record backlogs across all six OEMs — demand is contracted, not forecast. Vertiv $15B+, Trane $10.7B, Eaton EA $14.5B.
Backlogs are contracted, not forecast. Vertiv $15B+ and Eaton EA $14.5B represent 12–18 months of forward revenue at current run rates. Demand is visible; execution is the risk.
5 — Build Cost Trend 2020–2026E
Non-AI shell & core global average. 7% CAGR 2020–2025. 2025 deceleration to +5.5%. 2026E: $11.3M/MW (+6%). JLL 2026; T&T DCCI 2025.
Construction cost inflation is decelerating (+5.5% in 2025 vs +9% in 2024). But lead times are not improving — schedule risk and cost risk are moving in opposite directions.
6 — Equipment Lead Time Progression
Switchgear 80 wk, transformers 50 wk, generators 78 wk — all worsening. Average +50% from 2020. Cost inflation decelerating; lead times are not. McKinsey (2026); JLL 2026.
Switchgear (80 wk) and generators (78 wk) are the binding programme constraints, not civil works. Cost per watt is stabilising; schedule slip is not. Interconnection queue is a separate and often longer constraint.
7 — Supply Chain Risk Matrix
Lead time (x) vs $/W (y). Bubble size = OEM margin %. Red dashed = architecture risk. Top-right = primary underwriting constraints.
Top-right quadrant = highest underwriting risk. CRAH (red dashed) is structurally obsolete above 41 kW/rack. Liquid cooling CDU (blue) is mandatory for AI but in a less consolidated market.
8 — Component $/W: Non-AI vs AI vs Colo
Facility-only $/W for key powered shell components across three facility types. Lean Research (Mar 2026); T&T 2025.
CRAH is the starkest shift: ~$0.40/W in non-AI, ~$0.15/W in AI as liquid cooling displaces it. Liquid cooling CDU is the inverse — new in AI, near-zero in non-AI. The architecture has changed at the physics level.
9 — Shell $/W by Market
14 markets ranked. Baseline = NoVA $11.5/W (used in Cost Build tab). Tokyo +34% premium; Charlotte −16%. T&T DCCI 2025.
NoVA (baseline) understates costs by 34% in Tokyo and 18% in Silicon Valley. It overstates by 16% in Charlotte — the fastest-growing US data center market by new supply.
10 — Hyperscaler Capex 2022–2026E
Big Five combined: ~$162B (2022) → ~$345B (2025, company rows) → ~$B (2026E, range $–B). 2026 values read from EE_SYNC.capex.hyperscalerCapex2026; each company is on its own disclosure basis (FY guide, CY guide or TTM actual) — not like-for-like. Microsoft $175B is the post-restatement CY26 guide (Jul 29 2026: useful life 15→25 yr, ~$15B of reported capex removed, no change to buildout); the prior $190B print predates it. 72% CAGR since Q2 2023. Q4 2025: $140.6B. Note: 2025 aggregate uses Big Five company rows ($345B); EE Forecast 2040 shows ~$410B for 2025 because it also counts neoclouds — a scope difference, not a conflict. Prior Futurum/Introl $448B estimate was a broader market figure.
Big Five 2026E: ~$B (range $–B; Microsoft $175B post-restatement CY26 guide, Jul 29 2026; Amazon $220B post Jul 30 2026 raise; Oracle $55.7B FY26 actual on a May year-end). 2025 Big Five company rows sum to $345B (Forecast 2040’s ~$410B for 2025 includes neoclouds — different scope). 72% CAGR since Q2 2023. Q4 2025: $140.6B in a single quarter. This is the demand engine behind every supply chain constraint in the other charts.
11 — Liquid Cooling Market Growth
CDU/DLC market: ~$0.4B (2020) → ~$3B (2025) → ~$7B (2029E). Nearly doubled in 2025. Fastest-growing component category. Dell'Oro Group (Jan 2026).
CDU/liquid cooling doubled in 2025 to ~$3B (Dell'Oro), driven by AI rack densities exceeding the 41 kW/rack air cooling ceiling. Dell'Oro forecasts ~$7B by 2029. Incumbents and specialists are competing simultaneously.
12 — Rack Density by Facility Type
Air cooling ceiling: 41 kW/rack (physics limit). Enterprise AI: 30–50 kW. Hyperscale AI: 120–150 kW. GB200 NVL72: 120–130 kW. Rubin Ultra: ~600 kW forward.
The 41 kW/rack air cooling ceiling is a physics limit, not an economic preference. GB200 NVL72 at 120–130 kW is already 3× above it. Rubin Ultra at ~600 kW will require purpose-built CDU infrastructure from the ground up.
IT / Hyperscaler
Stable / growing
Moderate risk
High risk / obsolescence
EPC / civil
All charts derived from primary sources in respective tabs. Lead times: McKinsey (2026) · JLL 2026. Margins: public earnings filings (Q1 2026). Build cost: JLL 2026; T&T DCCI 2025. Liquid cooling: Dell'Oro Group (Jan 2026). Hyperscaler capex: Epoch AI / Futurum / company guidance. Not investment advice. electroneconomics.substack.com
Electron Economics · Source-Based Market Intelligence  ·  Last updated: Aug 2026 · Company guidance as of Q1 2026 earnings

Hyperscaler & Neocloud Capex Forecast

The demand side of the cost table. Big Five hyperscalers guided ~$B combined 2026 capex (range $–B; each company on its own disclosure basis — see the basis line on every card). Neoclouds and AI-native builders add $50-60B+ on top. Bars scaled to Amazon 2026 ($200B = 100%). All figures from public earnings guidance and named company announcements. Jun 2026.

Reading this tab: 2024 and 2025 = reported or preliminary; 2026 = company guidance as of Jun 2026. Hyperscaler capex includes IT equipment, facility, and leasehold — not directly comparable to the facility-only $/W in the Cost Build tab. Neocloud spend-to-revenue ratios of 2–3× are common during buildout phase. Not revenue forecasts.
JLL 2026 supercycle read-through: $3 trillion total investment required through 2030. 100 GW new capacity; global capacity doubles to 200 GW. Key inflection 2027: AI inference overtakes training as dominant workload — shifting demand from centralised hyperscale clusters toward distributed regional hubs. This changes the geographic demand pattern: secondary and tertiary markets gain share. Americas retain ~50% of global capacity. APAC: 32 GW → 57 GW. 77% of current construction pipeline pre-committed; global occupancy 97%. Lease rates CAGR ~5% through 2030.
Source: JLL 2026 Global Data Center Outlook (Jan 2026) · DCD (Mar 2026) · Bisnow (Jan 2026)
These five numbers are not like-for-like. Each 2026 figure carries a different disclosure basis — fiscal-year guide, calendar-year guide, or trailing-twelve-month actual — and the basis is printed on every card and every table row. Big Five 2026E = ~$B (range $–B), read from EE_SYNC.capex.hyperscalerCapex2026 so the Capex Stack, Forecast 2040 and RV Model cannot drift apart.
Microsoft restatement — Jul 29 2026 (FY26 Q4 call): data-centre useful life extended from 15 to 25 years and some leases reclassified from finance to operating, cutting reported capex by ~$15B with no change to physical buildout. The prior $190B figure shown in this dashboard predates that restatement; $175B is the post-restatement CY26 guide.
Scope note vs EE Forecast 2040 — 2025 is a scope difference, not a conflict. This tab shows Big Five 2025 = $345B (AMZN + MSFT + GOOGL + META + ORCL only). EE Forecast 2040 shows ~$410B for 2025 because its aggregate also includes neoclouds (CoreWeave, xAI, Stargate). Neither number is wrong; they count different sets of buyers. Do not net them against each other.
Entity Type 2024 capex 2025 capex 2026 guided YoY 25-26 Key focus / note Source
Sources: Amazon FY2026 implied guide; $173B TTM actual - Alphabet FY2026 guide $195-205B - Microsoft CY2026 guide, post Jul 29 2026 useful-life restatement (15 to 25 yr) - Meta FY2026 guide $130-145B - Oracle FY2026 guidance - CoreWeave Q4 2025 earnings call (Mar 2026) - xAI Southaven $20B announcement (Jan 2026); CNBC (Jun 2026) - Futurum Group (Feb 2026) - Introl (Jan 2026) - Goldman Sachs hyperscaler forecast - Epoch AI via Visual Capitalist (Apr 2026) - JLL 2026 Global Data Center Outlook (Jan 2026): $3T supercycle, 100 GW, 2027 inference inflection. All 2026 figures = company guidance or analyst consensus as of Jun 2026. Not investment advice. electroneconomics.substack.com
Electron Economics · Source-Based Market Intelligence · Last updated: Aug 2026

OEM Intelligence — Competitive Positioning by Component

Who owns what share of each powered shell component market. Two display modes: revenue-anchored share where primary data exists (CDU, generators, PDU, transformers), and competitive positioning where segment revenue is not separately disclosed. All figures from public earnings filings unless marked otherwise. Confidence tier shown on each card.

Powered shell TAM est.
~$400–500M
per 100 MW AI facility · Lean Research 2026
OEM margin range
20–30%
vs EPC/GC 4–8% on same project · public filings Q1 2026
Highest margin OEM
29.5%
Cummins Power Systems EBITDA · Q1 2026 record
Fastest growing segment
CDU / Liquid
AAON/BASX +186.8% TTM · Dell'Oro $3B→$7B by 2029
⬤ Primary Revenue figure from named public filing
◎ Triangulated Derived from multiple sources; directional
○ Positioning Qualitative assessment; no primary revenue denominator
Leader Gaining Holding At risk Emerging
Architecture risk read-through: The three components most at risk from structural transitions are CRAH units (physics obsolescence above 41 kW/rack — air cooling ceiling), UPS systems (800VDC architecture migration compresses traditional UPS content; Schneider + ABB both partnering Nvidia on 800VDC), and chillers (liquid cooling reduces chiller content per MW in AI-dense builds). Components with no architecture risk but supply constraint: switchgear (80 wk lead, GOES steel), transformers (50 wk lead, Prolec GE + Korean challengers), generators (18+ mo lead, Cummins order book to 2028).
Revenue figures: Vertiv Q1 2026 8-K · Cummins Q1 2026 8-K · Legrand Q1 2026 · ABB Q1 2026 · Eaton Q1 2026 · GE Vernova Q1 2026 · AAON Q1 2026 · Dell'Oro Group (Jan 2026) · Gartner UPS market share (2024, triangulated). Market TAM estimates for UPS/switchgear/chillers are EE triangulations from segment revenue disclosures — not primary TAM figures. Segment revenue where not separately disclosed is noted as positioning only. Not investment advice. electroneconomics.substack.com
Electron Economics · Source-Based Market Intelligence · Last updated: Aug 2026

Changelog — What Changed and Why

Every data update, structural change, and source addition since the first version. All changes are sourced from public earnings filings, named analyst reports, or EE channel checks. No version removes data — superseded figures are noted with their replacement.

NewFirst time this data appears
UpdateExisting figure replaced with newer filing
FixError corrected
StructureUI / tab / layout change
v22 Sep 12 2026 Update
Oracle’s FY2027 capex guide is carried structurally, with the gross-versus-net distinction that trade press keeps collapsing.
  • Oracle reaffirmed FY2027 capex of US$90–95bn gross on its 10 September Q1 call — a reaffirmation of the guide set on 10 June, not a revision. Q1 actual capex was US$28.5bn, free cash flow negative US$5.4bn, and RPO reached US$664bn.
  • The US$70bn figure circulating in trade press is the NET number — “not more than $70 billion” after customer prepayments, supplier financing and bring-your-own-hardware. Gross and net are two numbers for one company, both correct on their own basis, and they are routinely swapped.
  • No other hyperscaler revised guidance in the window. Microsoft holds ~US$175bn CY26, Amazon US$220bn, Alphabet US$195–205bn, Meta US$130–145bn, Tesla US$25bn. Verified individually rather than assumed.
  • One trap worth naming: coverage reporting Google as having “raised 2026 spending to US$205bn” is restating the top of the existing range, not announcing a raise.
  • The Cushman & Wakefield 2026 cost guide at US$17.6m/MW was already carried from 7 September, with its denominator-basis caveat intact. Re-checked and unchanged.
Oracle also disclosed a utilisation figure that must not enter any efficiency model
Oracle reported 97.9 per cent GPU utilisation in Q1 FY2027. That is almost certainly fleet or capacity utilisation — the share of deployed capacity under contract or in use — and not model FLOPs utilisation, which is an order of magnitude lower and which this platform carries at 42 per cent in the Compute Price Model. No source states Oracle’s denominator. The two numbers look alike, describe entirely different things, and confusing them would overstate delivered compute by roughly 2.3 times.
Oracle Q1 FY2027 earnings call, 10 Sep 2026
v21 Sep 07 2026 FixUpdate
Two hyperscaler figures were wrong, not stale, and the Big Five total was typed into four places that all kept saying 762.
  • Amazon moves from US$200bn to US$220bn. It was raised on the Q2 2026 call of 30 July 2026, attributed to elevated memory chip costs, which is three weeks before this dashboard’s own August vintage. That is an error, not drift.
  • Oracle moves from US$50bn to US$55.663bn. The US$50bn was carried as an FY2026 estimate after the actual had already been reported on 10 June 2026. FY2027 guidance is about US$70bn on Oracle’s new project-net-cash-outlay basis or US$90–95bn gross — two numbers, one company, both correct on their own basis.
  • Big Five 2026 total moves from US$762bn to US$788bn and is now read from EE_SYNC in all four places it appears rather than typed. NVIDIA’s CFO independently put top-five 2026 capex at nearly US$800bn on 26 August.
  • The total now carries the warning it always needed: it sums four calendar-year figures with one May-ending fiscal year, and one of the four is derived from a June-ending fiscal guide on a changed accounting basis. It is a convention, not a measurement.
  • The Amazon US$173bn trailing-twelve-month actual is withdrawn. It could not be re-verified against any source and carrying an unsourced figure is worse than carrying none.
  • An independent per-MW cost check is added and deliberately not reconciled: Cushman & Wakefield put all-in greenfield at US$17.6m/MW on 3 September 2026, up 21 per cent per MW since Q4 2024. Neither that figure nor this model’s stack states whether its denominator is critical IT or facility, and at typical PUE the two differ by more than the gap anyone would be explaining.
  • The 5.5 per cent build cost escalation is relabelled rather than re-fitted. It is a general non-residential construction rate, supported by Turner at +5.2 per cent, and it is not a data-centre-specific rate: Cushman implies about 16.5 per cent annualised. Both stand on the face of the block.
Amazon and Oracle were errors at the dashboard’s own vintage
Both figures were superseded before 18 August 2026 and neither was picked up. The staleness audit could not have caught them either — it measures whether a page cites recent dates, not whether the numbers it cites are the current ones.
Amazon Q2 2026 call, 30 Jul 2026 · Oracle Q4/FY2026 results, 10 Jun 2026
Powered land is now the fastest-moving line in the stack
Cushman & Wakefield put powered land in primary US markets at US$584,000/MW year to date, 35 per cent above the five-year average and 51 per cent higher year on year. Power infrastructure is 21 per cent of total greenfield cost, the largest single category.
Cushman & Wakefield 2026 Data Center Development Cost Guide, 3 Sep 2026
v20.1 Aug 18 2026 Fix
Big Five 2026 capex rises to US$762 billion, and the all-in cost per watt is now stated on the right basis.
  • The US$32.75 per watt all-in figure is per critical IT watt, not per facility watt. Applying it to facility megawatts without adjusting for cooling and distribution losses overstated the 2040 build requirement by about US$2 trillion.
  • Big Five 2026 capex moves from US$745 billion to US$762 billion, with Alphabet at US$200 billion, Meta at US$137 billion and Microsoft at US$175 billion.
  • Microsoft's step-down is an accounting change rather than fewer buildings. Data center useful life went from 15 years to 25, cutting reported capex by roughly US$15 billion.
  • Second-quarter 2026 supplier margins now read the same everywhere a company appears: Eaton Electrical Americas 27.5 percent, Vertiv 22.6 percent, ABB Electrification 20.2 percent.
  • The five bases behind those capex numbers are different accounting measures. The tracker now says which is which rather than summing them as if they were comparable.
Fix
Analytics OEM-margin chart was still on Q1 2026 while the OEM Margins tab was on Q2 — same company, same page, up to 190 bps apart. Chart constants advanced to the Q2 2026 prints already carried in the OEM Margins view: Eaton Electrical Americas 25.6% → 27.5% (8-K, Jul 30 2026), Vertiv Total 20.8% → 22.6% (8-K, Jul 29 2026), ABB Electrification 23.5% → 20.2% (Jul 16 2026). Each entry’s note now states the Q2 basis and the filing date. Cummins Power Systems (29.5%), Trane Americas CHVAC (~25%), Legrand DC (20.7%), Schneider (~19%) and EPC/GC (4–8%) already matched and are unchanged. Same defect found in a third place and fixed: the OEM Intelligence tab carried margin fields on the Q1 print sitting directly beside notes that already quoted the Q2 print — Vertiv 20.8 → 22.6 (2 records), ABB 23.5 → 20.2 (2 records), Eaton 25.6 → 27.5 (4 records), plus the switchgear and UPS component notes, which still cited Eaton 30% / ABB 23.2% / Vertiv 22.3% from Q1 2025.
Eaton 8-K Jul 30 2026 · Vertiv 8-K Jul 29 2026 · ABB Q2 2026 results Jul 16 2026
Update
Big Five 2026 capex reconciled with EE Forecast 2040 and re-based per company — $745B → $762B. All five figures now read from EE_SYNC.capex.hyperscalerCapex2026: Alphabet $180B → $200B (FY26 guide $195–205B), Amazon $200B → $200B (FY26 implied; $173B TTM actual, now labelled), Microsoft $190B → $175B (CY26 guide, post Jul 29 2026 useful-life restatement), Meta $125B → $137B (FY26 guide $130–145B), Oracle $50B unchanged. Total $745B → $762B, range $713–785B. The disclosure basis is now rendered per company on the card and in the table — these are not like-for-like.
Company FY26/CY26 guidance · EE_SYNC v1.0 (Aug 18 2026)
New
Microsoft useful-life restatement noted next to the Microsoft entry. On the Jul 29 2026 FY26 Q4 call Microsoft extended data-centre useful life from 15 to 25 years and reclassified some leases from finance to operating, cutting reported capex by ~$15B with no change to physical buildout. The dashboard’s prior $190B figure predates the restatement; the accounting change, not a build cut, explains most of the $190B → $175B step-down.
Microsoft FY26 Q4 earnings call · Jul 29 2026
New
2025 Big Five total labelled as a scope difference, not a conflict. This dashboard shows $345B for 2025 (AMZN+MSFT+GOOGL+META+ORCL). EE Forecast 2040 shows ~$410B because its aggregate also counts neoclouds (CoreWeave, xAI, Stargate). Neither figure changed; a scope note now sits on the Demand Forecast tab and in the Analytics chart caption so a reader comparing the two dashboards is not misled.
EE cross-model reconciliation · Aug 18 2026
New
$/W basis stated explicitly on the Cost Build tab — load-bearing. The $32.75/W midpoint is per critical IT watt, not per facility watt. Applying it to a facility-basis MW figure overstates capex by the PUE factor; EE Forecast 2040 had been doing exactly that. The new basis strip states the rule (divide by PUE first — 1.56 in 2025 → 1.18 in 2040) and that figures are 2026 dollars with no escalation applied.
EE_SYNC.capex.basis · EE_SYNC.forecast.pue
Structure
Single source of truth for the stack, the palette and the nav. (1) The Sankey layer widths and title now derive from EE_SYNC.capex.layersPerW and allInPerW — $32.75 is written once, not restated in the chart title. (2) PAL2 is now built at init from getComputedStyle on the :root tokens and the ~170 inline hex literals in the chart code were replaced with PAL2.* references, so canvas charts follow the tokens instead of drifting from them. (3) Nav model version and sync stamp now render from EE_SYNC.capex.modelVersion / EE_SYNC.version / EE_SYNC.synced, removing the hardcoded “v20 · Aug 2026” duplicates. (4) Invalid #detailView colgroup fixed — 16 <col> elements across two mismatched colgroups (one unclosed) reduced to the 10 that match the table’s 10 columns.
EE house-style sync pass · Aug 18 2026
v20 Aug 17 2026 Update
Every cost stack and supplier margin figure rechecked against second-quarter 2026 filings, and none needed revision.
Update
Verification pass — all cost-stack and OEM-margin data confirmed current (Aug 17). Re-checked every electrical and thermal OEM against live sources: Eaton Electrical Americas 27.5% Q2 margin, Vertiv, Cummins Power Systems record $2.3B / ~29.5% margin (Aug 4), Trane Americas CHVAC, ABB, Legrand, Schneider — all confirmed current through Q2 2026 earnings. No figures required revision. This dashboard's scope (powered-shell cost stack + electrical/thermal OEM margins) had no new filings since v19.
EE verification pass · Aug 17 2026
Update
Three-way sync date advanced to Aug 17 2026. Paired update with the Gas Tracker (v1.9, added Baker Hughes / Twenty20 Energy 250 MW Frame 5 order) and the Forecast 2040 (v1.9). Cummins Power Systems — the one company shared with the Gas Tracker — remains reconciled to the same Aug 4 2026 earnings print in both.
EE internal · Aug 17 2026
v19 Aug 16 2026 Structure
Now kept in step with the gas equipment tracker and the 2040 forecast, with Cummins reconciled to one earnings source.
Structure
Three-way sync with EE Modular Gas Tracker and US Data Center Forecast 2040. This dashboard now carries a sync marker in the header, aligned to the same Aug 16 2026 checkpoint as the other two platforms. The three are maintained as a paired set: this dashboard owns the powered-shell cost stack and electrical/thermal OEM margins (Eaton, Vertiv, Cummins, Trane, ABB, Legrand, Schneider); the Gas Tracker owns behind-the-meter gas equipment (GE Vernova, Baker Hughes, Siemens Energy, INNIO, Wärtsilä, Cat); the Forecast owns the 2040 capacity model. Where a company appears in more than one — Cummins Power Systems here and as a genset supplier in the Gas Tracker — the figures are reconciled to the same earnings source. No cost-stack figures changed in this version; alignment and marker only.
EE internal · Aug 16 2026
Update
Cross-checked Cummins Power Systems against Gas Tracker Q2 pass. Both dashboards now reference the same Cummins Q2 2026 print (Aug 4 2026): Power Systems record $2.3B revenue (+19% YoY), segment margin ~29.5% and the only segment expanding YoY, order book extending into late 2028. No change required — figures already matched.
Cummins Q2 2026 earnings Aug 4 2026
v18 Aug 2026 Update
Second-quarter 2026 results loaded for six suppliers, with Eaton at a record 27.5 percent margin and Trane at a US$12.1 billion backlog.
Update
Eaton Q2 2026 — EA record margin 27.5%; above guidance; Mobility RMT with Dana — Electrical Americas Q2: record $4.0B revenue (+18% organic), operating margin 27.5% (+190 bps sequentially; substantially above 22.6–23.0% guided). DC orders +85%, DC revenue +65% vs Q2 2025. Electrical backlog +103% YoY. Boyd FY2026 raised to $1.8B. Mobility: Reverse Morris Trust with Dana, $1.1B cash to Eaton pre-close, Q1 2027. FY2026: adj. EPS $13.40–$13.60, organic growth 11–13%, segment margins 24.1–24.5%.
Eaton Q2 2026, Jul 30
Update
Vertiv Q2 2026 — 22.6% margin above guidance; FY raised; ThermoKey closed; 2030 targets — Adj. operating margin 22.6% (+410 bps YoY), above 20.7–21.7% guide. Revenue $3.27B (+24%). Americas +29%, APAC +29%, EMEA returned to growth +2%. FCF $925M (+234% YoY). FY2026 raised: net sales ~$14.0B, organic +31%, adj. op. margin 23.8%, adj. EPS ~$6.70. ThermoKey closed Jun 12. 2030 targets: 20–22% CAGR, ~27% margin, $28B capital deployment.
Vertiv Q2 2026, Jul 29
Update
Trane Q2 2026 — record $12.1B backlog; applied bookings +130% 4th consecutive quarter — Applied bookings +130% YoY for 4th consecutive quarter above 100%; 2-year stack 4×. Record $12.1B backlog (+70% YoY); $6B committed 2027+. Americas CHVAC all-time high +50%. Enterprise organic bookings +37%. Revenue $6.4B (+11%). FY2026: ~9% organic, adj. EPS $15.20–$15.30.
Trane Q2 2026, Jul 30
Update
Cummins Q2 2026 — Power Systems record $2.3B; FY raised to +10–13% — Power Systems Q2: record revenue $2.3B (+19% YoY); only segment with margin expansion in Q2. Total company record $9.46B (+9%). Order book into late 2028. FY2026: +10–13% revenue (was +8–11%), EBITDA 18.0–18.5%.
Cummins Q2 2026, Aug 4
Update
ABB Q2 2026 — record orders $12B; 20.2% margin; Rotork ~$5.5B acquisition — Record orders $12.0B (+30%); revenues $9.5B (+14%); op EBITA 20.2% (+90 bps YoY — Q1 was 23.5%; step-down mix/investment-driven). Record backlog $30B (+28%). Americas +53% (US +62%). Rotork: ~$5.5B — valve actuators; margin-accretive from year 2.
ABB Q2 2026, Jul 16
Update
Legrand H1 2026 — DC now 32% of revenue; guidance raised to +16–19% — H1 sales €5.4B (+13.1% YoY, +17% excl. currency). DC organic growth >30%; DC now 32% of total revenue — first disclosure at this granularity. US revenue +26.7% (DC approaching half of US turnover). Seven 2026 acquisitions adding ~€450M annual revenue: Keydak, TES, SRS Power Engineering (Malaysia, €90M), Girtz Industries (US), others. FY2026 raised: +16–19% (was +10–15%).
Legrand H1 2026, Jul 28

v17 Jun 2026 New
New view ranks suppliers component by component, from cooling units to switchgear, with a risk flag on each.
New
OEM Intelligence tab — 8 component cards, competitive positioning by component — New 8th tab organised by component (CDU, generators, transformers, PDU, UPS, switchgear, chillers, racks) rather than by company. Revenue-anchored share where primary data exists (⬤ Primary / ◎ Triangulated); competitive positioning where segment revenue is not separately disclosed (○ Positioning). Architecture risk flag per component. Player bars with position badges (Leader / Gaining / Holding / Niche / Emerging). Hover for full source note.
EE analysis — Q1 2026 primary filings

v16 Jun 2026 Update New
Big Five capex raised to US$745 billion after Microsoft confirmed US$190 billion, and GE Vernova's backlog reached US$163 billion.
Update
GE Vernova Q1 2026 — Prolec GE + $163B backlog + 17.8% EBITA — Total backlog $116B → $163B (+$13B sequentially, incl. $5B Prolec GE). Gas turbine: 83 → 100 GW under contract; target ≥110 GW by year-end 2026. Electrification DC equipment orders: $2.4B in Q1 2026 alone — more than all of 2025 combined. Electrification EBITDA margin +590 bps to 17.8%. Prolec GE (transformer mfr): $5B backlog, ~$3B 2026 revenue, 20%+ EBITDA. Updated: substation drawer, transformer drawer.
GEV Q1 2026 8-K, Apr 22
New
AAON/BASX — added as specialist CDU challenger — Q1 2026: BASX-branded DC cooling sales +72.4% to $228.6M; total backlog $2.13B (+107% YoY, +16.5% sequentially); liquid cooling sales +186.8% TTM; book-to-bill >2× for 4th consecutive quarter. CEO targets ~$1B BASX revenue 2026. FY2026 guidance raised to +40–45% revenue growth — fastest organic growth rate in the CDU segment. Added to: liquid cooling source drawer, OEM tab CDU competitive note.
AAON Q1 2026 8-K, May 7
Update
Vertiv — Q2 2026 guidance + ThermoKey acquisition + raised FY guide — Q2 2026 adj. operating margin guided 20.7–21.7% (+270 bps YoY). FY2026 net sales raised to $13.5–14.0B (organic +29–31%). Americas Q1 organic +44%. ThermoKeyacquisition closed Jun 12 2026 — EMEA dry coolers, heat-exchange + mfg capacityEPS guidance +51% YoY at midpoint.
Vertiv Q1 2026 8-K, Apr 22
Update
Big Five capex total — $688B → $745B — Microsoft confirmed $190B CY2026 at Q1 2026 earnings (was $133B in prior version — a $57B gap). Big Five aggregate updated accordingly. Microsoft demand card corrected.
Microsoft FY26 Q3 earnings call (Apr 2026); internal QA
New
Eaton Beam Rubin DSX platform — Eaton debuted the Beam Rubin DSX platform in collaboration with NVIDIA at Q1 2026 earnings. Purpose-built power distribution for GB200/Rubin-generation AI racks. Validates Eaton as a grid-to-chip player for the next GPU architecture cycle.
Eaton Q1 2026 earnings call, May 5
Fix
Cummins — Power Systems vs total company EBITDA clarified — Card previously showed 29.5% EBITDA as if it were a company-wide figure. Clarified: 29.5% is Power Systems segment EBITDA (Q1 2026 record); total company EBITDA guided at 17.0–18.0% for FY2026. Power Systems segment FY2026 guidance: 25–26% EBITDA.
Cummins Q1 2026 8-K, May 5; internal QA

v15 Jun 2026 Update New
Liquid cooling priced for the first time at US$4.5 to 5.2 million per megawatt, and Schneider added as a seventh supplier.
New
JLL $3T supercycle (Demand tab) — New aggregate card: $3T total investment 2026–2030 ($1.2T real estate + $870B debt + $1–2T IT fit-out). 100 GW new capacity; global capacity doubles to 200 GW. 2027 inflection: inference overtakes training, shifting demand toward distributed regional hubs. AI workloads: 25% (2025) → 50% (2030).
JLL 2026 Outlook
New
Schneider Electric — 7th OEM margin card — Group-level adj. EBITA guidance 19.1–19.4% FY2026. Q1 2026: record €9.77B revenue (+11.2% organic); Energy Management €8B (+12.8% organic). Secure Power (UPS + PDU) still not separately disclosed — noted explicitly.
Schneider Q1 2026, Apr 30
Update
Eaton — Q2 2026 guidance added to margin card — Q2 Electrical Americas guided 22.6–23.0% (below Q1's 25.6% — further dip before H2 recovery). Boyd Thermal raises DC addressable market to $3.4M/MW. Mobility segment spin-off targeted Q1 2027, immediately margin-accretive post-close.
Eaton Q1 2026, May 5
Update
Legrand — updated to Q1 2026 + two DC acquisitions — Q1 2026: +18% total sales (+9.3% organic); 20.7% margin unchanged. Keydak (China rack mfr, €60M+) and TES (UK power distribution, €85M, majority DC revenue) acquired Q1 2026. FY2026 target: +10–15% sales growth. DC revenue: €0.7B (2020) → €2.4B (FY2025) → >€2.6B est. 2026E.
Legrand Q1 2026, Apr
New
CDU cost anchor — $4.5–5.2M/MW (Liquid cooling drawer) — First primary-sourced $/MW figure for the CDU layer specifically. Air-cooled heat rejection: ~$1.8M/MW. Liquid-cooled AI halls: $4.5–5.2M/MW including CDUs, secondary pumping loops, and dry coolers. AI-optimised facilities at 40–80 kW rack density: ~$20M/MW (excl. IT + land).
Archdesk / JLL Apr 2026

v14 Jun 2026 Fix Structure
Chip silicon is 53 percent of all-in cost, not 60 percent, and every market now carries an underwriting rating.
Fix
IT share math corrected — Previous versions stated "GPU silicon ~60% of all-in." Corrected to ~53% at the $32.75/W midpoint (46–64% across the $27.50–38/W source range). The 60% figure was only valid at the $27.50/W lower bound.
Internal QA
Structure
Executive / Component Detail toggle (Cost Build) — Default view is now an 8-row executive summary (layer-level). Full component breakdown available via "Component Detail" toggle. Reduces cognitive load on first open.
UX review
New
Basis dictionary — Four explicit definitions added: All-in $/W, Facility-only $/W, Shell-only $/W, and Tenant fit-out. Replaces the thin basis badge that previously confused readers mixing facility-owner and tenant cost bases.
UX review
Fix
Geography sort bug fixed — sortDir was toggling after sorting, causing the sort direction indicator to lag by one click. Now updates before sorting.
Code review
New
Market attractiveness score (Geography tab) — Each market now has a derived underwriting signal: Strong (Charlotte, Columbus, Atlanta), Good (Frankfurt, Paris, Portland, Madrid, Phoenix), Caution (NoVA, Chicago, Dublin, Singapore), Constrained (Tokyo, Zurich, Silicon Valley, Amsterdam). Sortable column.
EE analysis
New
Geography Shell / AI all-in toggle — Switch the primary $/W column between shell-only (T&T 2025 baseline) and all-in AI estimate (shell + $17.50/W IT + $9.50/W contingency).
UX review
Structure
Chart interpretation lines (Analytics tab) — One-sentence analytical read-through added below each of the 12 canvases. The most important: rack density chart — "The 41 kW/rack air cooling ceiling is a physics limit, not an economic preference."
UX review
Structure
Demand classification badges — Each demand entity now shows its status: Committed (Stargate JV), Company guided (hyperscalers, CoreWeave), Announced (xAI). All dated Jun 2026.
UX review
Structure
URL hash routing + resize handler — Tab state now reflected in URL (#cost, #oem, #analytics etc.) so tabs are shareable links. Charts redraw on window resize.
Bug fix
Structure
OEM tab headline — H1 changed to: "The powered shell is ~15% of capex — but where 20–30% OEM margins live." The previous title described the section rather than stating the thesis.
UX review
Structure
Source buttons renamed ? → Src — "?" implied uncertainty; "Src" signals a deliberate source reference. ARIA labels added to all 12 canvases and all source buttons.
Accessibility / UX

v13 Jun 2026 New Update
New demand view puts Big Five 2026 capex at US$688 billion, and the charts section grew from two to twelve.
New
Demand Forecast tab — Big Five hyperscaler 2026 guided capex: Amazon $200B, Alphabet $180B, Microsoft $133B, Meta $125B, Oracle $50B. Total $688B (+88% vs 2024). Neoclouds: CoreWeave $32B, xAI $20B, Stargate JV $100B. Goldman Sachs: $5.3T Big-4 FY2025–2030. Epoch AI: $448B combined 2025 at 72% CAGR since Q2 2023.
Company guidance / Futurum / Goldman / Epoch AI
New
Analytics tab expanded from 2 → 12 charts — Added: T&T cost category shift, OEM margin vs EPC gap, OEM backlog, build cost trend 2020–2026E, lead time progression, component $/W by facility type, 14-market $/W ranked bar, hyperscaler capex stacked area, liquid cooling market growth, rack density range. All 12 watermarked ELECTRON ECONOMICS.
EE analysis / primary sources

v12 Jun 2026 Update
First-quarter 2026 margins refreshed across all suppliers, with Cummins Power Systems at a record 29.5 percent and Eaton down to 25.6 percent.
Update
Q1 2026 earnings — all OEM margin cards refreshed — Cummins Power Systems: 22.7% (FY2025) → 29.5% EBITDA (Q1 2026 record, +430 bps). ABB Electrification: 23.2% (Q1 2025) → 23.5% Q1 2026 (+320 bps); $4.6B revenue. Trane: $7.8B → $10.7B backlog (+70% YoY); applied bookings +160% Q1 2026 (3rd consecutive quarter >100%). Vertiv: 22.3% (Q3 2025) → 20.8% Q1 2026; backlog $9.5B → $15B+. Eaton: 30% → 25.6% Q1 2026 (temporary; 12-factory ramp; exits 2026 above 30%).
Q1 2026 8-Ks, Apr–May 2026
New
Eaton Boyd Thermal acquisition (Mar 2026) — Eaton enters liquid cooling. Boyd Q1 revenue 2×+ YoY; backlog doubled in 6 months. Added to CDU competitive set alongside Vertiv, Schneider, Trane (Stellar), JCI.
Eaton Q1 2026, May 5
New
Cummins — fuel cell business sold to Alstom (Q1 2026) — Cummins is now a pure generator play. Bloom Energy remains the BTM fuel cell wildcard. Updated architecture callout and generator OEM competitive note.
Cummins Q1 2026, May 5
New
Eaton CEO market data — 32 GW of DC capacity under construction in US (70% AI); 228 GW total DC pipeline = "12 years of backlog at 2025 build rates." Added to reconciliation row source drawer.
Eaton Q1 2026 earnings call

v11 Jun 2026 New
Twelve Electron Economics articles now sit beside the data they explain, on every major view.
New
12 Substack article links across all tabs — In-line ↗ chips added to source drawers linking relevant EE Substack pieces: The Diesel Genset Decade, The Price of a GPU Hour, CoreWeave's Contract Machine, Bloom Energy × Compute, Equinix Part 2, The Tariff Lottery, OpenAI's Energy Problem, Gas Turbine Reservations. Related reading boxes on OEM, Geography, Vintage, Analytics, and Demand tabs.
electroneconomics.substack.com

v9–v10 Jun 2026 Structure
Rebuilt around four cost layer cards and a sortable market table, with the first two analytical charts added.
Structure
Executive UX rebuild (v9) — New headline: "AI data center capex is not one cost curve — it is three stacked markets." Four KPI layer cards above the table (IT $17.50/W / Shell $4–5/W / Civil $1–1.5/W / Gap $9.50/W). Basis badge strip. Source [?] collapsible drawers replacing dense inline source cells. Expandable footnote registry. Geography: cards replaced with JS-sortable table. Eaton HTML bug fixed (missing closing div). Sticky first column removed (mobile double-scroll).
UX review
New
Analytics tab — first 2 charts (v10) — Chart 1: Capex flow and margin pool (Sankey-style). Chart 2: Supply chain risk bubble matrix (lead time × $/W × OEM margin). Both Canvas-based, no external libraries, ELECTRON ECONOMICS watermark.
EE analysis

v8 Jun 2026 Update New
ABB and Legrand added as fifth and sixth suppliers, and networking sizing replaced with Arista's roughly US$9 billion actual revenue.
New
ABB and Legrand — OEM margin cards added (5th + 6th) — ABB Electrification: 23.2% op. EBITA Q1 2025; $8.69B record backlog; 800VDC partnership with Nvidia. Legrand Data Center: 20.7% adj. operating FY2025; DC €2.4B (26% of group); +40% organic 2025; CAGR +19% since 2019.
ABB Q4 2025 / Legrand FY2025
Update
Networking row — Arista FY2025 ~$9B primary source — Previously showed "~$30B DC" (third-party TAM). Replaced with Arista Networks FY2025 implied revenue ~$9B (Q4 $2.488B, +28.9% YoY); GAAP gross margin 63–65%.
Arista Q4 FY2025, Feb 12
Update
PDU row — Legrand DC €2.4B primary source — Previously had no credible primary source. Legrand FY2025 DC revenue is now the primary market proxy for the white space layer (PDUs, busbars, cable management).
Legrand FY2025, Feb 12
Update
Trane — $10.7B backlog, +160% applied bookings (Q1 2026) — Previous: $7.8B backlog, +120% Q4 2025. Third consecutive quarter of applied bookings >100%. Stellar Energy: $1B backlog; CEO: "$1B business in 2–3 years."
Trane Q1 2026, Apr 30

v6–v7 Jun 2026 New Structure
Build costs traced from US$7.7 million to US$11.3 million per megawatt since 2020, and fourteen markets priced from Charlotte to Tokyo.
New
Vintage Trend tab — 2020–2026E build cost trajectory ($7.7M → $11.3M/MW; 7% CAGR). Lead time worsening: switchgear 40 → 80 wk; generators 9 mo → 18+ mo. Liquid cooling CDU market growth: $0.4B (2020) → $3B (2025). T&T inflation YoY: +15% (2022), +9% (2024), +5.5% (2025) — decelerating. Key thesis: cost inflation decelerating; schedule risk is not.
JLL 2026 / T&T DCCI 2025 / McKinsey 2026
New
Geography tab — 14 markets with T&T 2025 $/W. Tokyo $15.2/W (most expensive globally) through Charlotte $9.5/W. NoVA/DFW used as baseline in Cost Build tab — Geography tab shows the delta.
T&T DCCI 2025 / CBRE 2025
Structure
Light mode / warm beige palette (v7) — Previous versions used a near-black background. Changed to warm beige (#faf8f4) across all surface tokens, text tokens, accent colours, and pill backgrounds. Dark mode explicitly rejected.
Design decision

v1–v5 Jun 2026 Structure
First release set the cost stack at US$32.75 per watt, added four supplier margin cards, and barred unverifiable market sizing.
New
Initial build — Cost Build tab — Three-column table: non-AI / AI / colo. Components across Layers 1–3 with $/W, $M/100MW, confidence pills (Primary / EE est. / EE↑). Source discipline: MarketsandMarkets, Mordor Intelligence, GMInsights, and all market research TAM firms excluded — methodology not disclosed, figures not reproducible. Only public earnings filings, T&T, JLL, Dell'Oro, Lean Research, and Epoch AI accepted as primary sources.
Lean Research / Epoch AI / T&T / JLL
New
OEM Margins tab — initial 4 margin cards — Eaton Electrical Americas 30% (Q1 2025 record), Vertiv 22.3% (Q3 2025), Cummins Power Systems 22.7% (FY2025), Trane Americas CHVAC ~25%+. All from named 8-K filings. Core insight: equipment OEMs earn 22–30% margins on products sold into projects where EPCs earn 4–8%.
Eaton / Vertiv / Cummins / Trane 8-Ks
Fix
Math reconciliation established (v5) — IT ($17.50/W) + Powered shell ($4.50/W) + Civil ($1.25/W) = $23.25/W. Gap to all-in: $9.50/W = 29% of $32.75/W midpoint. Named explicitly as structurally embedded overhead — not a rounding error. Derived from Epoch AI $38B/GW and Lean Research $27.50/W (400 MW) anchors.
Epoch AI / Lean Research
All version changes sourced from public earnings filings, named consultant reports, or EE channel checks. No version removes historical data — superseded figures noted with their replacement source. For methodology and full source registry, see the expandable footnote section on the Cost Build tab. Not investment advice. electroneconomics.substack.com
Electron Economics · Track record

Accuracy Ledger

Every figure this platform has published and later changed, across all seven models, with the size and direction of each change and what caused it. Research products ask you to trust their numbers. This is the evidence for how often those numbers have moved and by how much.
Revisions by model
Split between our own errors and figures that moved because new information arrived. A model with more revisions is not a worse model; it is usually a more heavily used one.
Our errorNew information
What actually goes wrong
Cause of our own errors only. New-information updates are excluded, because they are not mistakes.
Every revision
Rows marked with a green edge belong to the model you are looking at. Sorted by size of change.