Research cutoff: November 19, 2026. Structural analysis is intended to remain useful beyond one reporting season; all financial and demand figures below are explicitly dated. A forecast is not an observed outcome; a megawatt of generation capacity is not a megawatt-hour of electricity delivered; an equipment order is not recognized revenue; and not all data-center electricity use is AI-specific. Company figures are sourced from the cited reporting entity. This is educational research, not investment advice.
Executive summary: five conclusions for investors
The AI data center power supply chain comprises more than utilities and nuclear-power stocks. It includes electricity generation and procurement, transmission networks, regional interconnection, transformers and switchgear, facility distribution, uninterruptible power supply (UPS), backup systems, electrical monitoring and cooling-related power consumption. These layers create different business models, customer contracts and financial risks.
Five conclusions stand out:
- Electricity demand is growing, but the forecast is not a sales ledger. The International Energy Agency's 2026 baseline projected global data-center electricity consumption to rise from 485 TWh in 2025 to about 950 TWh in 2030, or around 3% of global electricity demand. Those are all data centers, not just AI workloads, and the 2030 value is a scenario projection. IEA
- Connection timing can matter more than theoretical generation supply. A data-center campus may have sufficient funding and hardware but still lack a viable grid interconnection, transmission upgrades, firm service agreement or completed on-site electrical systems. In the United States, the EIA discusses how large-load growth interacts with regional grid planning and generation choices. EIA
- The value chain is not one trade. Generators sell energy or contracted capacity; grid and electrical-equipment suppliers sell engineered products and services; critical-power vendors supply UPS, switchgear and associated infrastructure. All can benefit from greater activity, but their revenue timing, pricing power, capex requirements and valuation drivers differ.
- Reported order momentum is not identical to realized AI revenue. Eaton disclosed strong data-center orders and revenue growth, while Vertiv reported fast consolidated net-sales and organic-order growth. These are encouraging operating signals, but neither should be relabeled a market-wide AI power revenue number. Eaton · Vertiv
- Contracts and commissioning must be separated. Constellation reported long-term Amazon and Google agreements involving specified megawatts and future investment. The existence of a power purchase agreement (PPA) does not mean every associated megawatt is already additional operating generation capacity. Constellation
Investment question: Where does projected data-center electricity demand become a firm commitment, then an installed asset, then recognized revenue and cash flow—and what price is the equity market already paying for that conversion?
1. Why AI data centers need more electricity
Training and serving large models require accelerators, CPUs, storage, network switches and other IT equipment. These consume power and produce heat. Cooling, power conversion, pumps, fans, security and supporting infrastructure add additional electrical demand. Larger clusters can raise the absolute load even when individual chips become more energy-efficient.
Data-center electricity demand also includes cloud, enterprise and traditional internet workloads. Therefore, data-center load growth is not automatically a measurement of AI revenue. A forecast can rise because of increased AI adoption, conventional cloud demand, workload mix, low facility efficiency or greater installed infrastructure. A stock thesis must isolate what matters to the company's products and billing arrangements.
A 2026 source-backed demand snapshot
| Measure | Reported value | Correct scope | Source |
|---|---|---|---|
| Global data-center electricity consumption, 2025 | 485 TWh | IEA estimate; all data-center workloads | IEA 2026 |
| Global data-center electricity consumption, 2030 | ~950 TWh | IEA central projection; not realized, not AI-only | IEA 2026 |
| Global electricity share, 2030 | ~3% | IEA central projection for data centers | IEA 2026 |
| U.S. data-center electricity share by end of decade | 9.5%–15.3% | U.S. DOE/LBNL scenario range, not a realized or global share | DOE data center resource hub |
A reader should not average these forecasts: they reflect different geographic boundaries, models, years and scenario assumptions. Nor can they be multiplied by an arbitrary electricity price to derive revenue for a particular listed utility. Retail rates, wholesale prices, load factors, customer location and contract structure vary materially.
MW, MWh, TWh and PUE: four metrics that should not be confused
- MW (megawatts) is instantaneous power or capacity. A 100 MW service request describes a rate of power use or reserved capability, not annual consumption.
- MWh (megawatt-hours) is energy consumed or generated over a period. If an illustrative facility actually draws 100 MW continuously for 24 hours, its consumption is 2,400 MWh. The calculation is arithmetic, not a prediction that a specific facility will operate that way.
- TWh is one billion kWh, commonly used to describe annual electricity demand at a national or global scale.
- Power Usage Effectiveness (PUE) is facility power divided by IT-equipment power over the same measurement period. A PUE of 1.25 implies 25% additional facility electricity beyond measured IT electricity in that example; it is not a company-specific industry average.
The relevant investor distinction is requested utility capacity vs energized infrastructure vs actual annual consumption vs billable revenue.
2. How the AI data center power supply chain works
A conceptual electricity-delivery map:
ENERGY SOURCES AND PROCUREMENT
Natural gas / nuclear / renewables / other generation
|
Wholesale markets / utility procurement / power contracts
|
TRANSMISSION & INTERCONNECTION
Bulk transmission grid → substations → grid capacity studies
|
DISTRIBUTION & ELECTRICAL EQUIPMENT
Transformers → switchgear → breakers → monitoring & controls
|
DATA CENTER CRITICAL POWER
UPS / batteries / power distribution / backup generation
|
IT + COOLING LOAD
Servers / AI accelerators / networks + chillers / fans / pumps
|
RELIABLE TRAINING AND INFERENCE CAPACITYConceptual map only. This is not a single regulatory or ownership chain, a universal site design, or proof that a named vendor supplies a named data center. Utilities, power producers, campus owners and equipment vendors operate under different arrangements.
The economic layers
| Layer | What is sold | Potential revenue driver | Main constraints |
|---|---|---|---|
| Power generation | Electricity, output or contracted capacity | Energy prices, PPAs, operating availability | Plant economics, fuel, regulation, time to operate |
| Utilities and networks | Regulated delivery and system upgrades | Approved rate base, load growth, allowed returns | Interconnection, cost recovery, regulatory approval |
| Grid equipment | Transformers, high-voltage equipment, switchgear | Orders and deliveries for grid/campus upgrades | Production slots, skilled labor, lead times |
| Facility electrical | Medium/low-voltage distribution, breakers, controls | Installed distribution capacity | Design, project schedule, commissioning |
| Critical power | UPS, batteries, power management, backup systems | Installed MW, reliability and service contracts | Qualification, component delivery, service execution |
| Power management services | Maintenance, efficiency, controls, monitoring | Installed base and service attach | Competition, contract economics |
Important: Transmission and distribution utilities should not be grouped mechanically with independent power producers. Their return frameworks and risk allocation differ. A regulated utility may recover costs through approved rates, subject to regulatory tests; a merchant generator faces different exposure to pricing, hedging and contracting.
3. Generation: baseload, flexible power and procurement
Data-center operators need dependable electricity, but the way they meet that need varies by grid, market and contract. They may procure electricity through a utility tariff, bilateral agreement, virtual or physical PPA, dedicated supply arrangements, onsite generation or combinations of these. “24/7 power” is not the same as “24/7 carbon-free energy”: reliability and hourly emissions matching are distinct issues.
Nuclear generation and life-extension projects
Existing nuclear plants can offer large volumes of continuous generation, and uprates may add capacity without building an entirely new site. But uprates still require engineering, investment, regulatory approval and a schedule. Power contracts can improve visibility into financing; they do not eliminate execution risk.
In its November 9, 2026 results, Constellation Energy (NASDAQ: CEG) reported:
- A 20-year, 690 MW Amazon PPA announced in September 2026, associated with Calvert Cliffs and a related retail supply arrangement.
- A 20-year Google PPA announced in October 2026, intended to support over $4.3 billion of investment across 11 nuclear units and approximately 890 MW of incremental generating capacity, with the first uprate expected in 2028.
These are company-reported contractual commitments and future plans, not a claim that the entire 890 MW was operating in November 2026. Contractual and development milestones must be tracked separately. Constellation Q3 2026
Natural gas and flexible generation
Natural-gas power plants can provide dispatchable capacity and support grid reliability where fuel infrastructure and permitting allow. Nevertheless, construction, turbine availability, fuel-price exposure, emissions rules, interconnection and project financing can limit realized returns. GE Vernova's Power segment is relevant to gas-power equipment; the segment should not be equated with ownership of the electricity those turbines eventually produce. GE Vernova
Renewable energy, storage and grid integration
Wind and solar add electricity supply, but hourly output patterns do not always match the load of an AI campus. Batteries, demand management, diversified procurement and dispatchable generation can help balance reliability, costs and emissions. Storage's economics depend on market rules, duration, cycling and grid conditions. The precise optimal mix is regional; no single energy technology should be declared the universal AI-data-center winner.
4. Grid interconnections: why more generation does not guarantee a connection
A planned campus may request hundreds of megawatts, but connecting that load can require new substations, transformers, transmission lines or regional upgrades. The relevant parties may include utilities, grid operators, regulators, equipment suppliers and landowners. A planning queue can contain speculative or duplicated projects; requested load must not be mistaken for commissioned demand.
The EIA noted in 2026 that faster large-load growth can influence the location and type of future generation. EPRI likewise warned that announced nominal MW should be treated as a pipeline indicator rather than a near-term peak-load forecast, since non-IT loads, ramp schedules, load flexibility and on-site assets all affect the grid impact. EIA · EPRI
A diligence checklist for any announced data-center site:
- Is the quoted MW a request, a reserved allocation, a signed service agreement or already energized load?
- Is it IT load or whole-facility load? Is PUE included?
- Has an interconnection or capacity study been approved? What upgrades are required?
- Which company pays for the substation, transmission extension and system upgrades?
- Is the timetable firm or dependent on permitting, equipment delivery and construction?
- Who bears stranded-asset risk if the campus is delayed or canceled?
These questions matter for equipment demand, utility returns and the credit risk of developers.
5. Transformers, switchgear, UPS and backup power
The physical data-center power system spans several voltage levels and equipment classes:
Transformers change voltage so electricity can be transmitted and distributed. They are often engineered for specific capacity, operating environment and compliance standards. Their delivery schedules can affect both grid upgrades and campus construction.
Switchgear and circuit breakers isolate faults and route power safely through substations and facilities. Higher-density campuses may need more complex configurations, redundancy and protection systems.
Uninterruptible power systems (UPS) keep critical IT systems operating through interruptions or transitions to backup generation. They may include batteries, inverters and controls, and their commercial economics differ from those of wholesale generation.
Backup generators and onsite systems provide resilience but are subject to emissions, fuel availability, operating limits and site-specific design. A backup-generator order should not be counted as a sale of baseload electricity.
Cooling power must also be considered. Fans, pumps and chillers draw electricity; hardware or mechanical efficiencies can improve site PUE while total facility demand still grows.
Eaton, Schneider Electric and Vertiv are important references across these electrical and critical-infrastructure categories, though their portfolios and reported segments differ materially.
6. Key listed companies: roles, not a buy list
| Company | Identifier | Role in the power value chain | Most relevant reported evidence |
|---|---|---|---|
| Eaton | ETN · NYSE | Electrical distribution, protection and power management | Electrical segment orders, backlog, data-center revenue, margins |
| GE Vernova | GEV · NYSE | Gas power equipment, grid technologies and electrification | Power/ Electrification orders, backlog, operating profitability |
| Vertiv | VRT · NYSE | Critical digital infrastructure, UPS, power and thermal systems | Net sales, organic orders, backlog, service/margin trends |
| Constellation Energy | CEG · Nasdaq | Power generation and contracted nuclear output | Actual generation, PPAs, uprate investments, regulatory timing |
| Schneider Electric | SU · Euronext Paris | Energy-management equipment and automation | Energy Management revenue, demand, backlog, margin profile |
| Generac | GNRC · NYSE | Backup generation and power systems | Commercial and industrial orders, project timing, cash conversion |
These companies operate different parts of the value chain. Constellation is not a transformer supplier; Vertiv is not the owner of every grid it serves; GE Vernova equipment sales are not the same thing as retail electricity revenue. This role map does not assert contractual relationships with any particular hyperscaler unless separately sourced.
A company may have mixed exposure
Data-center growth may be concentrated in one division of a diversified company. Schneider Electric's revenue covers customers and end-markets beyond AI; Eaton serves aerospace and industrial customers alongside electrical infrastructure; GE Vernova operates Power, Electrification and Wind. A company-level sales growth rate cannot automatically be converted to an AI-exposure growth rate.
7. What 2026 financial filings actually show
The table is a dated evidence snapshot, not an estimate of AI-power market size. Fiscal periods, business definitions, order calculations and currencies differ. Do not add the figures together or rank stocks solely on the largest reported number.
| Company, report date | Verified company-reported figure | What it measures and what it does NOT measure |
|---|---|---|
| Eaton, Nov. 3, Q3 2026 | Electrical Americas US$4.2B sales, up 22% organically YoY | A regional electrical segment, not AI-only sales. Eaton |
| Eaton, Q3 2026 analyst slides | Electrical Sector data-center revenue ~65% higher YoY; rolling-12-month data-center orders >100% higher | Data-center exposure indicators; order metric and revenue metric use different bases, neither is segment-level total growth. Eaton slides |
| Vertiv, Oct. 28, Q3 2026 | Consolidated US$4.313B net sales, up 61% YoY; organic orders up more than 100% | Company-wide quarter revenue and company-defined organic orders, not power-equipment-only revenue. Vertiv |
| GE Vernova, Oct. 28, Q3 2026 | Consolidated US$12.6B revenue, up 27% YoY, 18% organically | Includes Power, Electrification and Wind, not solely AI data centers. GE Vernova |
| Schneider Electric, Oct. 16, Q3 2026 | Consolidated €11.460B revenue, up 16.9% organically | Company-wide revenue, not a data-center-only figure. Schneider IR |
| Constellation, Nov. 9, Q3 2026 | 690 MW Amazon PPA; Google-related ~890 MW planned incremental capacity | Contracted power and future project scope; not a quarterly revenue number or 890 MW already online. Constellation |
The table illustrates the different transmission mechanisms from industry demand to shareholder economics:
- Eaton: increased orders can flow into backlog, production slots, deliveries and cash. Look at organic growth, mix, operating margin and whether delivery costs erode earnings.
- Vertiv: rapid order and revenue growth may indicate deployment urgency, but supply-chain costs, service execution, acquired growth and working capital determine realized value.
- GE Vernova: power equipment and grid demand can support backlog; investors still need division-level profitability, project execution and cash conversion.
- Schneider Electric: data-center demand is a growth driver across a diversified energy-management portfolio; company-wide revenue is not a precise AI measure.
- Constellation: contract duration and pricing may support investment, but operating availability, project completion, hedging and capital expenditures determine returns.
Orders vs backlog vs revenue vs cash
Order: a commercial commitment under company-specific definitions. It may be cancellable, reschedulable or conditional.
Backlog: the outstanding value of orders or contracts yet to be fulfilled; definitions and cancellation provisions vary by company.
Recognized revenue: the amount reported under applicable accounting rules when performance obligations are satisfied.
Cash flow: the timing and amount of cash after working capital and investments. A higher backlog can coexist with weak near-term free cash flow.
No analyst should combine orders, backlog and revenue into a single “AI power sales” headline.
8. Where the bottlenecks and profit pools may be
Grid connection and approvals. An attractive plant site can remain stranded without grid access. Track signed utility agreements, schedule changes, construction permits and approved funding.
Power-equipment manufacturing. Transformers, switchgear and other engineered components may face production constraints and long qualification schedules. High backlog can support utilization, but unit economics still depend on pricing, labor, inputs and delivery quality.
Generation and gas turbine supply. Dispatchable generation capacity may require equipment manufacturing slots, fuel infrastructure and regulatory approval. Long lead times can defer when electricity actually reaches a campus.
UPS and facility reliability. More installed IT capacity increases the amount of critical-power equipment that must be qualified, installed and serviced. Reliability requirements can create recurring service revenue, but new designs and competition also create risk.
Cooling and the electric load. Cooling constraints can cap usable IT capacity even if nominal power is available. Power and thermal technologies overlap at the facility level, but investors should not double-count the same project as two independent end-markets.
A company with a visible bottleneck is not necessarily the best-performing stock. The investment case depends on realized pricing power, margins, capex, market expectations and dilution or financing needs.
9. Investment implications: who benefits, and under what conditions?
There is no single investment outcome for the full AI power infrastructure ecosystem. The same AI buildout can produce a backlog surge at an equipment supplier, a long-dated PPA for a generator, a regulated capital program for a utility and substantial financing needs for a developer.
| Scenario | Supporting evidence to monitor | What could change the conclusion |
|---|---|---|
| Sustained buildout and healthy economics | Firm load contracts, equipment deliveries, disciplined margins, energization milestones, positive cash conversion | Delays, rising costs, lower utilization, price competition |
| Growing projects but delayed monetization | High orders and permit activity, slow grid work and revenue conversion | Faster approvals or canceled/downsized campuses |
| Demand normalization or oversupply | Slower hyperscaler capex, downward load forecast revisions, excess manufacturing capacity | Strong realized utilization or new applications |
These are research scenarios, not probability estimates, model-generated price targets or backtested signals.
Seven risks an investor should investigate
- Valuation risk: the stock price may already assume years of accelerating equipment demand or contracted generation earnings.
- Project timing: grid interconnection, land, zoning, construction, fuel and equipment deliveries rarely share a single schedule.
- Contract quality: PPAs can differ by tenor, price indexation, deliverability, termination rights and counterparty credit; their announced MW do not disclose the full economics.
- Ratepayer and regulatory risk: utility investments can face scrutiny over who pays for network upgrades, tariffs and affordability.
- Energy-system risk: weather, fuel prices, grid congestion, generation outages and transmission limitations affect reliability and margins.
- Capital intensity and cash conversion: rising capex, inventory and working capital can offset high revenue growth.
- Demand overstatement and technology efficiency: proposed campuses may not be completed, and better model/hardware efficiency can change how much energy is used per workload; overall demand may still rise through rebound effects.
10. A quarterly investor monitoring dashboard
A good industry dashboard records every figure's source, reporting date, geographic scope, denominator and status (observed, contracted, planned, forecast).
| Measure | Why it matters | Recommended verification |
|---|---|---|
| Regional utility large-load interconnection milestones | Converts announced MW into usable load | Utility/regulator and grid-operator records |
| Actual energized MW and facility utilization | More reliable than announced project capacity | Operator filings where disclosed |
| Energy use, TWh and electricity pricing | Describes demand and procurement economics | IEA, EIA, utility filings, market data |
| Generation / equipment contract terms | Makes future cash-flow visibility assessable | Company filings and confirmed PPAs |
| Orders, backlog and cancellation rights | Estimates conversion pipeline and downside | Eaton, GE Vernova, Vertiv, Schneider filings |
| Organic sales growth and margins | Distinguishes acquired growth from core demand and value capture | Financial statements and reconciliations |
| Capital expenditure and free cash flow | Tests investment returns and funding needs | Cash-flow statements, regulatory filings |
| Data-center power density and PUE | Affects site power footprint and equipment design | Operator disclosures and technical standards |
| Customer concentration and delivery delays | Indicates contract, supply and execution risks | Annual/quarterly risk disclosures |
| Valuation, dilution and cost of capital | Separates strong industry activity from an attractive stock | Dated stock data and explicit forecasts |
Simple repeatable research workflow: map the company's role → select the correct disclosed operating metric → determine whether demand is forecast, contracted or realized → assess margin and free cash flow conversion → compare against valuation → record the evidence that would invalidate the thesis.
11. Frequently asked questions
How much electricity do AI data centers use?
Reliable public statistics often cover all data centers, not AI alone. The IEA estimated global data-center electricity use at 485 TWh in 2025 and projected about 950 TWh in 2030 in its 2026 central scenario. Neither figure is an AI-only consumption statistic. IEA
Which public companies are exposed to AI data center power demand?
Examples of different roles include Eaton (ETN), GE Vernova (GEV), Vertiv (VRT), Constellation Energy (CEG), Schneider Electric (Paris: SU) and Generac (GNRC). They are not interchangeable and this list is not a recommendation to buy their shares.
Why is electricity a bottleneck for AI data centers?
Even if electricity generation exists in the wider grid, a specific campus may need transmission capacity, substations, transformers, service agreements and critical-power systems before it can operate. Time, place and reliability are as important as annual energy volume.
Do power purchase agreements guarantee investment returns?
No. PPAs differ in pricing, settlement, counterparty obligations, credit quality, performance conditions and capital requirements. A contract is evidence of commercial interest, not a guaranteed profit or a promise that planned generation is already commissioned.
What is the difference between MW and MWh for data centers?
MW measures power at a moment, while MWh measures electricity delivered or consumed over time. A 100 MW load running continuously for one hour uses 100 MWh. Annual use depends on the actual load profile and operating hours.
Is nuclear power the only way to supply AI data centers?
No. Grid supply can draw on nuclear, natural gas, hydro, wind, solar, storage and other technologies. Reliability, regional availability, permits, contracting, cost and carbon objectives determine the mix.
Are Eaton and Vertiv direct competitors?
They overlap in selected power-distribution and critical-infrastructure categories but have different product portfolios, end-market mixes and financial reporting. Comparisons require specific products and customer use cases, not an assumption that every dollar of their revenue competes directly.
Does stronger AI infrastructure demand mean these stocks will outperform?
Not necessarily. Shares reflect expected future growth and profitability. The relevant question is whether incremental orders become sufficiently profitable free cash flow relative to the valuation and investment risk.
Continue research on BasisPilot
Related industry hubs (publish these links only when the destination exists):
Existing dated Insights (their URLs should not be changed):
- Schneider Electric Q3 2026: Data Center Power
- Amazon–Constellation: Calvert Cliffs PPA
- Google–Constellation: Nuclear Uprates and Power
- Constellation Q3 2026: Power Contract Economics
- Data Center Generator Contract: Generac
Stock research: Use existing /stocks/{ticker} pages only after validating route availability. Do not fabricate a published stock-page URL for non-US listings.
Sources, definitions and update policy
Sources below prioritize the IEA, U.S. agencies, company investor-relations releases and securities disclosures. The company figures represent their explicit date and financial category. No invented revenue allocations, site-utilization percentages, capacity additions, valuations, target prices or supplier contracts are included.
- IEA 2026, Key Questions on Energy and AI, executive summary — 2025/2030 global data-center electricity demand and scenario context.
- EIA, Data center server energy use grows across the commercial building stock — U.S. demand trends and modeling scope.
- Eaton Q3 2026 results — Electrical segment sales, orders and backlog context.
- Eaton Q3 2026 analyst presentation — company-defined data-center order and revenue growth metrics.
- Vertiv Q3 2026 earnings — consolidated sales, organic orders and growth definitions.
- GE Vernova Q3 2026 SEC earnings exhibit — consolidated financial reporting and segment context.
- Schneider Electric Q3 2026 investor-relations results — quarterly revenue and data-center demand context.
- Constellation Q3 2026 earnings — Amazon/Google contracts and expected nuclear uprate timelines.
- EPRI Powering Intelligence 2026 — region-specific grid constraints and limitations of announced MW.
- EIA, fossil generation and faster data-center load growth — generation and grid implications under demand scenarios.
- GE Vernova Q3 2026 investor update — company characterization of utility and data-center demand.
- U.S. DOE Data Center Resource Hub — U.S. demand scenarios, not realized all-world figures.
Editorial review policy: Recheck the IEA/EIA forecast when new versions are issued, company financial rows after each new quarterly filing, and project capacity when contracts are amended or actual energization is reported. Keep structural explanations distinct from dated numbers; explain substantive changes rather than refreshing timestamps cosmetically.
Disclaimer: This report is for education and independent research, not personalized investment advice or an instruction to trade a security. Company revenue, growth and signed contracts do not establish stock valuation or future shareholder returns.