AGI Investment Tracks — Track 8

Data Center Physical Infrastructure

As of 2026-05-26 · Analyst: datacenter-infrastructure

1. Thesis

Given AGI is coming within a couple of years, the binding constraint on compute scaling shifts from chips to everything that wraps the chips: the substation, the transformer queue, the 500-ton cooling unit, the electrician on the roof. Hyperscaler capex is in the $400B+/yr range and the physical buildout is still in early innings — the 165% increase in DC power demand by 2030 (Goldman) implies roughly 100+ GW of net new capacity in the US alone, against a supply chain where large power transformers are now on ~128-week lead times and generator step-ups on ~144 weeks.

The honest catch: the picks-and-shovels trade is no secret. Vertiv is up ~241% in the last year alone (mkt cap $124B, +25x since 2022). Eaton, Quanta, Comfort Systems, GE Vernova have all run. The real edge in 2026 is finding the lagging links — transformer pure-plays still trading at industrial multiples (HD Hyundai Electric, Hammond Power Solutions, Mitsubishi Electric, HPS Korea), backup-power names that just inked gigawatt-scale deals (Bloom Energy via AEP), and the mechanical/electrical contractors with multi-year backlogs already locked but trading off short-term sentiment (MYR Group, MasTec power delivery segment).

Conviction core: Transformers > switchgear > on-site generation > mechanical EPC > cooling. Transformers are still the single most-cited bottleneck and have the fewest qualified global manufacturers. Vertiv-class cooling is now a crowded trade.

2. The Bottlenecks — Quantitative

Large Power Transformer Lead Time

~128 wks

~2.5 yrs (Wood Mackenzie, Q2'25). GSUs ~144 wks.

US Power Transformer Deficit

~30%

Distribution transformers ~10% deficit, worsening for 3-phase pad-mount.

Transformer Demand Surge (since 2019)

+119–274%

GSU demand +274%, power transformers +119%, distribution +34%.

DC Vacancy (Primary US Mkts)

1.9%

Record low. 6,350 MW under construction at end-2024 (2x 2023). 2026 still tight.

DC Power Demand by 2030

+165%

Goldman Sachs baseline; AI workloads are the marginal driver.

Equinix AI-Linked Deals Q1'26

~60%

8 of top 10 AI model providers actively expanding with EQIX.

The bottleneck stack (worst to least)

Large power transformers (500+ MVA): Fewer than five global manufacturers can produce 765 kV+ units. Lead times stretched from 6–12 months pre-2020 to 24–36 months. Prices up 77% since 2019. Grain-oriented electrical steel (GOES) — produced by a handful of mills globally, NLMK sanctioned 2022 — is the upstream chokepoint.
Electrical labor — linemen, journeymen, commissioning engineers: Quanta has a 12-month backlog of $28B against ~$35B in revenue, and is hiring/training aggressively but talent compounds slowly.
Switchgear (MV/HV): Lead times ~44 weeks. Eaton, ABB, Siemens, Hubbell capacity-constrained. Eaton announced $30M expansion for medium-voltage switchgear in Q1'26.
Backup power (diesel + gas gensets, fuel cells): Cummins Power Systems +19% YoY in Q1'26, management cites demand "outpacing expectations." Bloom Energy signed $2.65B / 1 GW deal with AEP to skip the grid queue entirely.
Land in power-rich corridors: Loudoun County saturated. ERCOT growth limited by transmission, not generation. Phoenix water-constrained. New builds shifting to Iowa, Ohio, Tennessee, Indiana.
Cooling (especially liquid for >50 kW/rack): Vertiv crowded; nVent, Modine still earlier in the story. Submer, Asetek, CoolIT mostly private.
Long-haul fiber/DCI optics: Coherent, Lumentum, Ciena ramping. Less binding than power.

3. Supply / Demand Gap Model (back-of-envelope)

Rough US-centric framing for new IT load coming online 2026–2030:

Component	Demand pull (2026–30)	Supply add capacity	Gap (qualitative)
Large power transformers (US fleet additions)	~100 GW new DC + grid replacement of aging fleet (~55% > 33 yrs old)	~$2B of US capacity investment since 2023 (Hitachi, Siemens, Eaton, Prolec, Hyundai, etc.) — bringing online 2026–28	Severe through 2027, easing 2028+
Gas turbines (peakers / behind-meter)	GE Vernova gas backlog 100 GW (target 110 GW by year-end '26)	Slot reservations already filled to 2028+ at GEV, Mitsubishi Power, Siemens Energy	Severe — slot reservation now the gating event
Diesel/gas gensets (backup)	~50 GW backup capacity for new DC builds	Cummins, Caterpillar, Generac ramping	Tight — lead times stretched but not 2yr+
Electrical contracting labor	Quanta $48.5B backlog, MasTec $20B, MYR record	~7% YoY headcount growth at top EPCs; apprentice pipelines lag	Severe — multi-year
Liquid cooling (CDUs, manifolds, rear-door HX)	All new GPU racks >50 kW need liquid; majority of 2026+ builds	Vertiv, nVent, Schneider, Modine adding capacity	Tight, easing
Switchgear (MV/HV)	Every MW of DC needs gear	Eaton, ABB, Siemens, Hubbell investing	Tight, ~44wk lead times
DC colocation MW	AI hyperscaler + neocloud absorption (CoreWeave, Crusoe, Lambda)	6.35 GW under construction (end-2024), more since	Tight — 1.9% vacancy; pricing power for EQIX/DLR/IRM

4. Investable Public Companies

Power equipment / transformers / switchgear

Ticker	Name	Mkt cap	Why it's a bottleneck	Risks	Priced in?
ETN	Eaton	~$157B	MV switchgear, UPS, electrical Americas; Q1'26 record orders/backlog; raised organic growth to 10%	Cyclical, fwd P/E 28x	Mostly
GEV	GE Vernova	~$140B (est)	Gas turbines (100 GW backlog), Grid Solutions, Prolec consolidation; Q1'26 orders +71% organic; $2.4B DC electrification orders in Q1 alone	Execution; cap-ex cycles	Mostly
267260.KS	HD Hyundai Electric	~$10B+ (volatile)	Pure-play transformer / switchgear; 24% op margin; $4.7B backlog (~2.4x sales = 3+ yrs locked); US factory ramping +30% by 2026	Korean tape, FX, KOSPI access	Less so
298040.KS	Hyosung Heavy	Small-mid cap	Korean transformer pure-play, US export ramp	Same access issues; smaller scale	Less so
HPS.A.TO	Hammond Power Solutions	C$3.7B	NA dry-type transformer pure-play, record backlog "driven by US/data center demand"; AEG Power acquisition pending	Small/mid cap; concentration	Partially
HUBB	Hubbell	~$22B	Utility T&D products, enclosures, gear; benefits from grid spend	Slower revenue mix to AI	Mostly
ABBN.SW	ABB	~$110B	Electrification, MV products, transformers — broad exposure	Big and diversified	Mostly
SU.PA	Schneider Electric	~$130B	Switchgear, DC reference designs, cooling (acquired Motivair)	Multiple re-rated; FX	Mostly
ENR.DE	Siemens Energy	~$80B	HVDC, grid solutions, gas turbines; massive backlog ramp	Wind exposure (Gamesa)	Partially
6501.T	Hitachi (Energy)	~$140B+	Largest investor in transformer capacity globally ($4.5B by '27)	Conglomerate discount	Less so
6502.T	Toshiba	Private/restructuring	Switchgear / transformers — limited liquidity	Hard to own	N/A

Cooling, racks, in-DC infrastructure

Ticker	Name	Mkt cap	Why it's a bottleneck	Risks	Priced in?
VRT	Vertiv	~$124B	Cooling + UPS + rack systems; FY26 organic growth 29-31%; $15B+ implied backlog; "AI buildout not yet peaked"	Up 241% in 1yr, 25x in 3yr; stopped disclosing quarterly backlog	Heavily
NVT	nVent	~$15B (est)	Liquid cooling, enclosures; Q1'26 record $2.6B backlog "led by AI DC"; sales $1.24B vs $809M YoY (+54%)	Hyperscaler concentration risk	Partly
MOD	Modine Manufacturing	~$6-7B	Data center cooling segment growing fastest; less covered	Auto/HVAC legacy	Less so
MTRSY	Munters	Mid cap	DC cooling specialist (Swedish)	Smaller, less liquid	Less so
KGP.IR	Kingspan (Tate raised floor)	~$14B	Raised access flooring near-monopoly; DC insulation	Building products cyclicality	Less so

Backup / on-site power

Ticker	Name	Mkt cap	Why it's a bottleneck	Risks	Priced in?
CMI	Cummins	~$50B	Power Systems +19% Q1'26; gensets sold out; "demand outpacing expectations"; raised FY guide to +8–11%	Diesel ESG headwinds; legacy truck mix	Partly
BE	Bloom Energy	~$12-15B (est)	$2.65B / 1 GW deal with AEP to feed AI DCs (Jan 2026) — fuel cells skip the grid queue	Cash burn, execution; one big customer	Partially
GNRC	Generac	~$10B	Backup power (more SMB/residential, but commercial growing)	Mostly outside hyperscaler scale	Less so — for good reason
FLNC	Fluence Energy	~$3B	Battery storage for DC peak shaving	Execution; tariffs	Off-thesis (mostly)

EPC / contracting / mechanical

Ticker	Name	Mkt cap	Why it's a bottleneck	Risks	Priced in?
PWR	Quanta Services	~$70-80B (est)	$48.5B record backlog, $28B 12-mo backlog; $2.4T TAM through 2030; electrical labor moat	Up massively; margin discipline	Heavily
FIX	Comfort Systems USA	~$30B+ (est)	Mechanical EPC; $12.45B backlog +81% YoY; hyperscale exposure huge	+300% in 12 months; hyperscale concentration	Heavily
MTZ	MasTec	~$13B (est)	$20.3B record backlog; Power Delivery double-digit growth; pipeline +91% (LNG/gas for DC power)	Pipeline cyclicality	Partly
MYRG	MYR Group	~$3B (est)	T&D-focused electrical contractor; record Q1'26 backlog; smaller/less-followed pure-play vs PWR	Small-cap volatility; project execution	Less so
EME	EMCOR	~$30B (est)	Mechanical/electrical; DC exposure growing	Diversified business mix	Partly
PWR/APG/ACM/J	AECOM, Jacobs, API	Various	DC design, commissioning, infrastructure consulting	Service businesses, capped margins	Partly

Data center REITs & networking

Ticker	Name	Mkt cap	Why it's a bottleneck	Risks	Priced in?
EQIX	Equinix	~$80B (est)	~60% AI deals; record Q1'26 bookings/backlog; raised FY guide; 26% cash-on-cash on stabilized	REIT rate sensitivity; capex heavy	Partly
DLR	Digital Realty	~$50B (est)	Hyperscale colocation; AI driven absorption	Slower than EQIX on retail/AI mix	Partly
IRM	Iron Mountain	~$30B (est)	Frame data center business growing; underappreciated	Legacy storage drag	Less so
ANET	Arista Networks	~$140B	Hyperscaler switching standard; AI cluster spine	Microsoft/Meta concentration	Mostly
ALAB	Astera Labs	~$15B (est)	Connectivity ICs for AI servers (PCIe/CXL retimers)	Single-product franchise risk	Mostly
CRWV	CoreWeave	~$50B (est)	AI neocloud, exposed to physical buildout but is a buyer not seller	Off-thesis for picks-and-shovels	N/A

Cable, conduit, fiber

Ticker	Name	Mkt cap	Why it's a bottleneck	Priced in?
PRY.MI	Prysmian	~$25B	HVDC cable, fiber; benefits from DC interconnection and offshore wind backlog	Partly
NEX.PA	Nexans	~$6B	HV cable, DC connections	Less so
ATKR	Atkore	~$2B (est)	Conduit, cable management — direct DC consumable	Cyclical, beaten down
BDC	Belden	Small/mid	Cabling, DC build-out	Less so
COHR / LITE	Coherent, Lumentum	Various	Optical transceivers for AI cluster interconnects	Mostly

5. Pre-IPO / Private to Watch

Switch (taken private by DigitalBridge / IFM, 2022) — possible re-IPO candidate, premium AI-grade colocation.
QTS (Blackstone, 2021) — could spin back or partial monetization.
CoreSite (American Tower, 2021) — embedded in AMT.
Submer, Asetek (public small), CoolIT — immersion / liquid cooling. CoolIT growing fast on hyperscaler wins.
Wilson Power, Howard Industries, Virginia Transformer Corp — private US transformer makers; M&A targets.
Zayo (DigitalBridge / EQT) — long-haul fiber, possible IPO 2026–27.
Crusoe Energy — flared-gas + DC; raising at growing valuations; private.
Applied Digital APLD (already public, sub-$2B) — HPC hosting; speculative.

6. Commodity / Physical Plays

Copper — every MW of DC needs ~25–40 tonnes of copper (busbar, cable). Plays: FCX (Freeport-McMoRan), SCCO (Southern Copper), TECK, ANTO.L. Already a known macro AGI trade.
Grain-oriented electrical steel (GOES) — transformer cores. Producers: Cleveland-Cliffs CLF (only US producer), JFE, Nippon Steel, Thyssen, Baoshan. NLMK (Russia) was a top global producer, sanctioned 2022 — structural shortage projected by Wood Mac through 2030. CLF is the most direct US-listed bet on the chokepoint.
Aluminum — busbars; rising tariff complications.
Nickel & cobalt — adjacent to backup-battery demand.
Silica/quartz, rare earths for transformers — niche; mostly via specialty miners.

7. People / Talent Concentration

The labor moat is real and underappreciated. In order of acuteness:

Transformer-design engineers — globally fewer than ~5,000 senior engineers capable of 765 kV+ unit design. Concentrated at Hitachi, Mitsubishi, Hyundai, Siemens, Prolec.
DC commissioning engineers — boutique firms: Cupertino Electric (private), Salute Mission Critical (private), Vert (private), DPR Construction (private).
Lineman / journeyman electrical labor — IBEW union pipeline cannot keep pace. Quanta, MYR, MasTec hoarding talent via subsidiaries and apprenticeship programs.
Mechanical EPC / chiller commissioning — Comfort Systems, EMCOR concentrate this; Holder Construction (private) dominant in DC GC.
Power transformer factory labor — coil winders, tank welders; cited by Hitachi as their #1 capacity constraint, more than steel.

8. Top Picks — Ranked by Conviction

1. HD Hyundai Electric (267260.KS) — Highest conviction

Why: Korean-listed pure-play transformer + switchgear maker. 24% op margin. Backlog of $6.7B is 2.4x annual sales — 3+ years of work locked. North America (highest-margin region) growing fastest, US factory ramping output +30%. Slot reservation system means they cherry-pick highest-margin hyperscaler orders. Trades at less demanding multiples than US-listed peers because of Korea discount + access friction.

Catch: Korean tape access (need to use Interactive Brokers or a Korea-capable broker). FX exposure. Stock is up ~233% but starting from a lower base than VRT/PWR.

Conviction: 9/10 — closest thing to a pure-play on the most-binding bottleneck.

2. Hammond Power Solutions (HPS.A.TO) — North America transformer pure-play

Why: Canadian-listed dry-type transformer specialist. Record Q1'26 backlog "driven by US and data center demand." AEG Power Solutions acquisition (closing Q2'26) adds power-quality breadth. C$3.7B mkt cap — still small enough to compound. Less crowded than VRT/ETN. RBC PT C$350 vs current trading.

Catch: Already up 233% YoY. Liquidity OK but lower than US large caps. Smaller scale = customer concentration risk.

Conviction: 8/10.

3. GE Vernova (GEV) — gas-turbine + grid combo, still under-owned vs scale

Why: 100 GW gas turbine backlog (target 110 by year-end '26), $2.4B of data center electrification orders in Q1 alone (exceeded all of 2025). Prolec GE consolidation in Q1 added $5B to backlog. Orders +71% organic in Q1'26. Free cash flow $4.8B in the quarter. Three-engine company: gas turbines (multi-year sold out), grid solutions (transformer/switchgear), electrification (DC equipment).

Catch: Stock already up substantially since 2024 spin from GE. M&A gains inflate reported net income.

Conviction: 8/10 — best US-listed combination of bottlenecks under one ticker.

4. nVent Electric (NVT) — under-followed liquid cooling + enclosures

Why: Record $2.6B Q1'26 backlog led by AI DCs. Sales went from $809M to $1.24B YoY. Smaller and less covered than Vertiv. Liquid cooling is the secular tailwind as racks move from 20 kW to 100+ kW.

Catch: Hyperscaler concentration (5 customers ~50%+). Vertical integration risk if hyperscalers in-source. Stock has run on the print.

Conviction: 7/10.

5. Cleveland-Cliffs (CLF) — chokepoint upstream play on GOES

Why: Only US producer of grain-oriented electrical steel (Butler Works, PA — heavily expanded for transformer demand). The single most-cited upstream chokepoint in transformer manufacturing. Stock has been beaten down by steel cyclicality, giving an unusual entry point for what's effectively a hidden AGI pick-and-shovel.

Catch: CLF is a diversified steel co — GOES is <15% of revenue. The thesis is the marginal value of GOES rises faster than the rest of steel falls.

Conviction: 6/10 — most "left-behind" of the picks, with the most steel-cyclicality noise.

Honorable mention — but call it priced-in: Vertiv (VRT), Quanta (PWR), Comfort Systems (FIX), Eaton (ETN). These are the right businesses for the thesis but the multiples now require continued upside surprise. If you don't already own them, the risk/reward is worse than the picks above. If you do own them, no reason to sell — the buildout is multi-year, not multi-quarter.

9. What Would Change My Mind

Hyperscaler capex pause — would hit EPC and equipment names first. Watch MSFT, META, GOOGL, AMZN capex guides quarterly. Any of them cutting by 20%+ would re-rate the whole stack.
Algorithmic efficiency breakthrough — if test-time compute or post-RSI inference becomes 10x cheaper without proportional usage expansion, demand for raw GW slows. (Counter: Jevons paradox suggests usage will fill any efficiency gain).
Transformer manufacturers ramping faster than expected — if Hitachi/Siemens/Hyundai/Prolec come online ~12 months earlier than guided, the transformer scarcity premium compresses. So far they're guiding production starts 2027–28.
Hyperscaler vertical integration — Google's TPUs were the analog. If Microsoft or Meta start manufacturing their own transformers, cooling distribution units, or switchgear (already happening for liquid cooling at modest scale), the value-add of VRT/NVT/Hubbell shrinks.
Distributed inference at the edge becomes dominant — would shift demand from central DC to edge buildout, hurting EQIX/DLR concentration plays but helping edge networking (CSCO, JNPR).
Permit/grid interconnect liberalization — FERC Order 2023 implementation accelerates; ERCOT/PJM queue clears faster. Hurts behind-the-meter generation thesis (Bloom, Cummins) but probably accelerates equipment demand overall.
Catastrophic recession — would hit cyclical EPC names first; transformers are stickier given multi-year backlogs.

10. Sources & Date

Data as of 2026-05-26. All Q1 2026 figures from company press releases / SEC filings unless noted. Estimates are clearly labeled.