AI Supply Chain Deep Dive: The Energy Constraint Inside the Compute Build-Out

Core Conclusions

• The impact of the AI compute build-out on energy has already moved from a "concept theme" into a "power-system constraint" phase. The IEA estimates global data-center electricity use at about 415TWh in 2024, rising to 945TWh by 2030 under its base case; data-center power use jumped another 17% year over year in 2025, with AI-type data centers growing faster still. Lawrence Berkeley National Laboratory estimates that U.S. data centers already consumed 176TWh in 2023 and could reach 325-580TWh by 2028; EPRI further raised its 2030 U.S. data-center figure to 380-790TWh, equivalent to 9%-17% of total U.S. electricity. This means the real constraint on AI is no longer "are there chips," but "can the power actually be connected."

• The first to benefit, and along the shortest path to realization, are not the generators but "data-center electrical equipment + cooling + medium/high-voltage distribution gear." AI rack power density has risen from about 25kW/rack in 2022, to 40kW in 2023 and 72kW in 2024, up to roughly 132kW/rack for the GB200 design in 2025, with the GB300 reference design already at 142kW/rack, and the industry even beginning to plan for 1MW/rack. This directly lifts the per-rack dollar content of UPS, PDUs, busways, switchgear, CDUs, liquid cooling, chiller plants, distribution systems, and medium-voltage equipment.

• The true near-term bottlenecks are concentrated in three "long-lead-time" links: large gas turbines, transformers/high-voltage gear, and grid-interconnection approvals. GE Vernova's 2025 investor materials show that its new-equipment orders are already sold out through 2028, with remaining sellable 2029 capacity below 10GW; Hitachi Energy has explicitly said the global transformer shortage is intensifying and added a 250 million dollar investment in 2025 to address it; NERC has likewise listed large data-center load as a new long-cycle reliability risk.

• Over the medium term, natural gas and existing/restarted nuclear are the most realistic options to "fill the gap" in AI power; renewables and storage are necessary conditions but usually not a standalone instant solution. The IEA explicitly notes that natural gas and coal together will carry more than 40% of new data-center power demand through 2030; at the same time, the tech industry is still pushing nuclear and SMRs and accelerating the signing of long-term power-purchase agreements. The IEA also notes that, constrained by slow grid interconnection, many U.S. projects are advancing on-campus natural gas generation, and because AI data-center load swings quickly, battery storage is becoming a key technology.

• The group of companies that benefit most directly is the one whose revenue, orders, or customers already clearly come from AI data-center power demand. Typical examples include Vertiv, Eaton, Schneider Electric, nVent, Legrand, Delta Electronics, GE Vernova, and Siemens Energy, along with Hitachi Energy/Hitachi, ABB, and Prysmian on the high-voltage equipment and transformer side. Their common trait: their products sit directly on the "grid-distribution-rack-cooling" path, and they already have AI data-center project orders or management that explicitly names data-center demand.

• Utilities and independent power producers will also benefit, but you have to separate "certainty of benefit" from "speed of realization." Constellation comes closer to a "direct beneficiary" through long-term nuclear supply agreements with the likes of Microsoft and Meta; Vistra, Talen, and Dominion benefit from load growth and improving power prices, but their earnings realization is more constrained by regulation, transmission bottlenecks, site location, load curves, and capex, making them a "moderate-to-high beta but not fully certain" group.

• The real opportunity in the Chinese market lies in "compute-power coordination," green-power trading, liquid cooling, and data-center power equipment, not a simple bet on traditional generation. The National Data Administration and the National Energy Administration have continued to advance compute-power coordination, green-power direct connection, and demand-response mechanisms across 2025-2026; the green-power share target for newly built data centers at national hub nodes has been raised to 80%. This makes the more worthwhile angles in China liquid cooling, UPS/distribution, power modules, data-center transformers, and storage/backup power, rather than a blanket "all power stocks benefit."

• The base figure for Chinese data-center power demand itself differs by definition, which shows that "the easiest thing to get wrong in industry research is guesswork extrapolation." The IEA estimates Chinese data-center power use at about 100TWh in 2024 while explicitly flagging a large data gap for China; meanwhile, under CAICT's definition, China's "computing centers" used about 166 billion kWh of electricity in 2024, or 1.68% of total social electricity use. This is a reminder to first distinguish, in investment research, whether the figure includes traditional IDC, edge nodes, carrier equipment rooms, crypto mining rigs, and where the boundary of "computing center" is drawn.

• The popular but logically weak "pseudo-beneficiaries" fall into three main groups: first, companies that simply own generation assets but are not near AI load centers or cannot sign new high-value load contracts; second, equipment makers that talk up "AI" but whose revenue is not broken out, whose customers are not disclosed, and whose orders do not show up; third, thematic names mapped onto AI merely because "nuclear/storage/green power" is a hot concept, but with no AI data-center customers or order evidence yet. The IEA, Uptime, and multiple equipment makers all emphasize that what truly limits the build-out today is distribution, cooling, approvals, transformers, and interconnection capacity -- not every energy company will benefit in lockstep.

• The biggest risk is not that demand disappears, but that "project execution runs slower than the market expects." The supply chain and approval chain have become decisive for the pace of revenue recognition: NERC stresses that large-load interconnection challenges system planning; reporting from Hitachi Energy and Reuters shows large-transformer lead times can run into years; GE Vernova and Siemens Energy are also adding capacity while production schedules stay tight. In other words, AI demand may keep growing fast, but listed companies' revenue realization will be reshaped by "equipment scheduling + grid interconnection + EPC cadence."

Industry Chain Landscape and Demand Analysis

The AI Energy Supply Chain at a Glance

The Power Logic of AI Data Centers

The core of AI power demand is not the "number of servers" but per-rack power density, load utilization, cooling method, and PUE. Schneider's official materials show that the NVIDIA DGX SuperPod has moved from about 25kW/rack in 2022 to roughly 132kW/rack in the GB200 era, and on to 142kW/rack in the GB300 reference design; in Vertiv's 360AI reference design, 24 AI racks are configured at 132kW/rack, for a full 3.6MW solution with liquid cooling accounting for as much as 76%. Traditional data centers still commonly run racks in the 5-15kW/rack range, which is why, once AI penetrates, distribution, UPS, busways, cooling, and building electrical systems get rebuilt in tandem.

On PUE, the Uptime Institute's 2025 global sample shows an average PUE of 1.54, essentially flat for six straight years; but the average for facilities built within the last 5 years has fallen to 1.48, large data centers above 20MW average as low as 1.44, and some new facilities in North America and high-latitude Europe already reach 1.3 or better. This means total AI power demand will still rise meaningfully, but the competitive focus for new equipment makers and operators has shifted to "keeping non-IT energy use under control at higher power density." Liquid cooling is therefore not a simple "energy-saving theme" but a prerequisite for deploying high-density compute.

The IEA also notes that in 2024, accelerated servers already accounted for 24% of server power use and 15% of total data-center power use; and that in data-center power growth from 2025 to 2030, about 70% of the server-side increment comes from accelerated servers. This means the training side is no longer just the GPU companies' own story; it has materially propagated into power supply, cooling, cabling, transformers, and campus engineering.

Demand Shifts Globally and in Key Regions

Globally, the IEA estimates data centers consumed about 415TWh in 2024, with the U.S., Europe, and China together accounting for roughly 85% of the world total. Of that, U.S. data centers used about 180TWh in 2024, over 4% of total U.S. electricity; China used about 100TWh, though the IEA simultaneously flags clear gaps in China's definitions and data coverage; Europe used about 70TWh, slightly below 2% of total European electricity; and Japan used less than 20TWh. By 2030, under the IEA base case, the U.S. adds about 240TWh versus 2024, China adds about 175TWh, Europe adds 45TWh+, and Japan adds about 15TWh.

The U.S. is the most critical incremental market. LBNL research shows U.S. data centers already used 176TWh in 2023 and could rise to 325-580TWh by 2028, corresponding to 6.7%-12.0% of total U.S. electricity; in 2026 EPRI again raised its 2030 scenario to 380-790TWh and 9%-17%, noting that Virginia is currently the only state where data centers account for more than 20% of power use, potentially rising to 39%-57% by 2030. Materials Dominion Energy submitted to regulators even show that as of July 2025 it had signed 30,132MW of ELOA engineering-authorized capacity, with its official load forecast climbing to 17,596MW by 2046; Reuters reported in 2026 that Dominion's contracted data-center capacity had approached 51GW as of March 2026. This shows the U.S. East Coast, Texas, and parts of the Midwest have entered a stage of "grab the power first, negotiate the land next, and only then the data-center equipment."

China's policy direction is "compute-power coordination," not letting load pile up in the east unchecked. In 2025 the National Data Administration explicitly stated that the average PUE of data-center clusters at the eight major hub nodes had fallen to about 1.3, with the most advanced data centers as low as 1.04, and that it aims for a green-power share of 80% for newly built data centers at national hub nodes in 2025; in 2025 the NDRC and the National Energy Administration also issued green-power direct-connection policy, encouraging compute facilities to sign multi-year green-power trading contracts and participate in ancillary services and demand response. At the same time, CAICT estimates China's computing centers used about 166 billion kWh of electricity in 2024, or 1.68% of total social electricity use. For investment, this means the more realizable parts of the China chain are liquid cooling, data-center power, transformers, storage/backup, green-power trading, and campus microgrids.

The European, Japanese, and East Asian markets share a common trait: power constraints are more explicit, and new capacity leans more toward efficiency and low carbon. The EU has established a data-center energy-efficiency and sustainability reporting framework; European Parliament research notes that EU data centers account for about 3% of regional total electricity, with countries like Ireland higher still. Japan's METI launched a "Watt-Bit Collaboration" public-private mechanism in 2025, explicitly treating data-center expansion driven by AI and communications traffic as an important energy-policy topic; Schneider also disclosed in Q1 2026 that China and East Asia, driven by data centers and infrastructure, delivered double-digit growth in East Asia, while Japan and Taiwan were simultaneously driven by semiconductor expansion.

Power Procurement Strategies of Cloud Providers and Operators

The strategy of the large players is now very clear: lock in baseload first, then lock in green power, then add on-campus self-generation and storage. The IEA noted in 2026 that the tech industry accounted for about 40% of global corporate renewable PPA signings in 2025 and has become an important demand driver for nuclear and advanced geothermal; meanwhile, affected by slow grid interconnection, many U.S. projects have begun advancing on-campus natural gas generation, and because AI data centers exhibit rapid, large load swings, battery storage is becoming a key technology for the next generation of AI data centers.

Long-term nuclear contracts have moved from experiment to scale. Microsoft reached an agreement with Constellation to support restarting the Crane Clean Energy Center associated with Three Mile Island; Meta signed a 20-year nuclear agreement with Constellation to support continued operation of the Clinton plant in Illinois; and Google partnered with Kairos Power to advance advanced nuclear, planning to deploy 500MW by 2035. The significance of such deals lies not only in the volume of power per transaction, but in how they push AI power procurement from "annual green-power offsets" toward a "campus-level, decade-plus, financeable" long-term supply model.

Scenario Forecasts

Segment-Level Breakdown

Segment Matrix

What Is Direct Benefit, Indirect Benefit, and Pseudo-Benefit

Direct beneficiaries: Vertiv, Eaton, Schneider, nVent, Legrand, Delta, Envicool, Shenling Environment, GE Vernova, Siemens Energy, Hitachi Energy/Hitachi, HD Hyundai Electric. These companies either already explicitly name data-center/AI demand as an order driver in their financials, or their products go directly into AI data-center power and cooling systems.

Indirect beneficiaries: Constellation, Vistra, Talen, Dominion, Quanta Services, Prysmian, ABB, CGN Power, China Resources Power. They benefit from the systemic power-demand growth AI induces, but financial realization usually has to pass through intermediate links such as capacity contracting, regulation, line construction, PPA structuring, and interconnection.

Pseudo-beneficiaries: any company that mentions AI only in roadshow or media framing but discloses no corresponding customers, orders, capacity utilization, revenue share, or margin change should be parked in a Tier-D watch pool rather than placed directly on the "AI energy core beneficiary" list. This judgment matters especially in China and Asia, because a large number of power-equipment and energy companies will passively benefit from total power-use growth, yet only a few can directly share in the high value density of AI data centers.

Master Table of Investment Targets

Master Table of Listed Companies

Note: the table below prioritizes companies worth further research; where no company explicitly discloses "AI-related revenue share," it is uniformly marked "undisclosed/estimated." Valuation summaries are based on public data available as of 2026-05-16; for non-U.S. and some Asian companies, real-time valuation capture is incomplete, marked "needs further verification."

In-Depth Analysis of Key Companies

Vertiv

Company overview and business structure. Vertiv is a global leader in critical digital infrastructure, with a core business spanning critical power, thermal management, low/medium-voltage distribution, busways, and services. In its 2025 annual report the company lists data centers as its primary end market and explicitly names hyperscale/cloud, colo, and neocloud as core customer groups.

Direct exposure to AI energy demand. This is one of the most direct AI energy-chain names in the entire market. Its revenue path is very short: as AI rack power density rises, it directly lifts the value of UPS, busways, liquid cooling, thermal management, and services. Vertiv's collaboration with NVIDIA has moved from "selling equipment" to "jointly building AI factory reference architectures," and the annual report mentions a collaboration arrangement with Oklo.

Financials, orders, price, and margins. 2025 revenue was 10.230 billion dollars, up 27.7% year over year; net income was 1.333 billion dollars. The company's backlog doubled from 7.2 billion dollars at the end of 2024 to 15 billion dollars at the end of 2025, most of which will be delivered within 12-18 months. The high backlog and higher-density product mix mean future margins and operating leverage still have support.

Competitive advantage and peer comparison. Versus Eaton and Schneider, Vertiv is purer and has more "direct AI data-center exposure"; versus nVent, Vertiv's power + cooling + service integration is more complete. The trade-off is a clearly higher valuation.

Valuation and research conclusion. As of 2026-05-16, market cap is about 145.5 billion dollars, P/E about 93x, and PS roughly 14.2x on 2025 revenue. EV/EBITDA and FCF yield need to be worked out further with net debt and cash-flow line items. On balance, I rate it a strong beneficiary / growth + thematic investment, but the valuation margin for error is already thin; the indicators most worth tracking are new orders, the liquid-cooling share, the speed of backlog-to-revenue conversion, and cancellation/deferral rates.

Eaton

Company overview and business structure. Eaton is a global electrification platform company whose two major segments, Electrical Americas and Electrical Global, are simultaneously driven by data centers, distribution, and electrical upgrades. Compared with Vertiv's "pure data-center exposure," Eaton's benefit is broader, covering distribution, switching, protection, UPS, industry, and utilities.

AI exposure path. AI is not Eaton's only growth factor, but it is already one of the core incremental sources. The company's 2025 annual report explicitly states organic sales growth of 8%, with one important driver being the strong performance of data center end-markets across both electrical segments.

Financials and profit. 2025 revenue was 27.448 billion dollars, up 10% year over year, with organic growth of 8%; gross margin was 37.6%, and although down slightly from the prior year, overall earnings quality remained very high. Eaton's business-model advantage is that data-center demand drives not only one-time equipment sales but also channel, service, and refresh cycles.

Moat and comparison. Versus Vertiv, Eaton is less "pure AI" but more resilient; versus Schneider, Eaton's U.S. electrical channel and North American data-center project exposure are more direct.

Valuation and research conclusion. As of 2026-05-16, market cap is about 155.5 billion dollars, P/E about 39x, and PS roughly 5.7x on 2025 revenue. On balance, I rate Eaton a strong beneficiary / growth investment; its advantage is certainty higher than most pure-theme names; the indicators to track are the two electrical segments' data-center orders, M&A integration, gross margin, and delivery cycles.

GE Vernova

Company overview and business structure. GE Vernova holds gas-turbine, T&D, grid-equipment, and storage-related businesses simultaneously, making it one of the few platforms that can satisfy both "AI campus self-generation" and "external grid expansion." In its 2025 annual report and investor materials, the company itself explicitly lists AI data centers as part of a power-demand super-cycle.

Direct AI exposure. Its core benefit is this: once a data center cannot secure sufficient external grid power, it turns to gas for backfill; and once a campus is committed to construction, it procures medium/high-voltage gear, substation equipment, and storage. GE Vernova's investor materials show new equipment is already sold out through 2028, with remaining sellable 2029 capacity below 10GW, one of the strongest pieces of supply-constraint evidence today.

Financials and orders. 2025 total revenue was 38.068 billion dollars, with Adjusted EBITDA of 3.196 billion dollars. The company added 18GW of contracts in 2025 and expects about 80GW of gas-turbine contracts in hand at year-end. The biggest advantage of such long-cycle industrial assets is extremely high order visibility; the drawback is that project delivery, raw materials, and policy disruptions also get amplified.

Peer comparison. Versus Siemens Energy, GE Vernova's U.S. supply chain and the U.S.-equity "AI backfill" narrative are more favored by the market; versus utilities, GEV is more of a "shovel seller." But like all turbine companies, AI is only part of the demand growth, and not all of the increment can be attributed to AI.

Valuation and research conclusion. Current market cap is about 285.4 billion dollars, P/E about 30.7x, and PS roughly 7.5x on revenue. EV/EBITDA needs to be supplemented with net debt. The research conclusion is strong beneficiary / cyclical + thematic investment; keep tracking the turbine scheduling horizon, service orders, Electrification orders, and rare-earth/material supply risk.

Constellation Energy

Company overview and business structure. Constellation's core is large-scale, scarce U.S. nuclear and zero-carbon power assets. In the AI era, its value lies not in "total generation rising slightly," but in owning 24/7 baseload that tech companies can lock in through long-term contracts.

AI exposure path. The long-term nuclear supply arrangements associated with Microsoft, Meta, and others show that tech companies are willing to pay with longer terms and higher certainty to lock in zero-carbon power. This gradually moves Constellation from a traditional utility valuation framework into a "scarce baseload + long-term contract platform" framework.

Financials and valuation. The company's 2025 annual report has been published; current market cap is about 96.5 billion dollars, P/E about 26.7x. Because this round of research did not uniformly capture its full revenue, net debt, and FCF line items, PS, EV/EBITDA, and FCF yield still need further verification.

Risks and verdict. The main risks are nuclear-plant availability, regulation, power prices, and the market having already priced in an "AI + nuclear" premium. The research conclusion is strong beneficiary / dividend + thematic investment; the indicators worth tracking are long-term contracting scale, existing-nuclear uprates, plant capacity factors, and the regulatory environment.

Schneider Electric

Company overview and business structure. Schneider is an "electrification + automation + software + service" platform leader. Its role in AI data centers is not as a single-component supplier but as an integrated solutions provider across distribution, energy management, liquid-cooling control, DCIM, and services.

AI exposure. The company's full-year 2025 results explicitly state: Q4 2025 Energy Management organic growth of 11%, led by Data Center; in Q1 2026 it again disclosed double-digit growth in Data Center & Networks demand, with Pure Data Center demand also up double digits. By region, North America, China and East Asia, and multiple European markets were all driven by data-center projects.

Financial performance. 2025 revenue was 40.152 billion euros, Adj. EBITA was 7.520 billion euros, and free cash flow was 4.635 billion euros. The advantage of such a cash-flow-type industrial leader is that even if the macro slows, as long as Data Center and Grid stay strong, the valuation can be supported by quality.

Peer comparison and verdict. Versus Vertiv, Schneider is more diversified and usually carries more of a "quality premium," unlike a pure AI-theme stock; but in global regional coverage and software/service integration, Schneider may have the edge. Research conclusion: strong beneficiary / growth + quality investment. Track the Systems business growth rate, Data Center order quality, liquid-cooling control-system progress, and regional mix.

Siemens Energy

Company overview and business structure. Siemens Energy's AI benefit logic is likewise twofold: Gas Services on one end and Grid Technologies on the other. The 2025 annual report shows revenue of 39.1 billion euros, orders of 58.9 billion euros, and backlog of 138 billion euros; Q2 2025 single-quarter orders were 14.4 billion euros with backlog rising to 133 billion euros; and the 2026 IR page shows backlog has risen to 154 billion euros.

AI exposure. The company has repeatedly cited data centers in earnings communications as an important source of turbine and grid-equipment demand. Reuters reporting in 2026 also directly links its profit improvement to AI-driven data-center demand.

Strengths and risks. The strength is that both order book and product positioning are excellent; the risk is that wind businesses such as Siemens Gamesa can still disturb the overall valuation and sentiment. Versus GE Vernova, Siemens Energy's "pure AI backfill narrative" is slightly weaker, but the structural demand for Grid Technologies is very strong.

Research conclusion. Positioned as a strong beneficiary / cyclical + growth investment. Keep tracking new Gas Services orders, Grid Technologies margins, the convergence of the wind drag, and U.S. capacity-expansion progress.

nVent

Company overview and business structure. nVent was originally more of a supporting-component industrial company, but its data-center exposure clearly grew starting in 2025. The company's 2025 Q4/full-year materials show full-year revenue of 3.9 billion dollars, adjusted operating income of 786 million dollars, and ROS of 20.2%; data-center sales were about 1 billion dollars, up 50%+ year over year, with organic growth of about 40%.

AI exposure path. The company enters data-center distribution and infrastructure through busways, cabinets, electrical connections, and the acquired Avail infrastructure business; the 10-K explicitly states the newly acquired business is mainly oriented to utilities and data centers. Versus Vertiv/Eaton, nVent is more of a low-profile "distribution-side AI chain backfill" name.

Valuation and conclusion. Current market cap is about 27.7 billion dollars, P/E 28x, and PS 7.1x on 2025 revenue. The research conclusion is strong beneficiary / growth investment, worth tracking the data-center revenue share, Avail integration, and the penetration of busway and cabinet products in AI data halls.

Legrand

Company overview and business structure. Legrand is essentially a leader in low-voltage electrical and digital building infrastructure, but data centers have become one of its strongest growth poles in recent years. The company's full-year 2025 results explicitly state: 2025 revenue grew 13%, driven by datacenters and M&A, with an adjusted operating margin of 20.7% and free cash flow of 1.3 billion euros.

AI exposure. Reuters reporting in 2025 shows Legrand's 2024 data-center revenue already reached 1.6 billion euros; and in 2025 the company's framing was further upgraded to "record growth driven by data centers." This shows it is no longer "slightly affected" but materially benefiting.

Verdict. Legrand's issue is not the logic but whether the market treats it more as a "building electrical stock" or a "data-center electrical stock." The research conclusion is strong beneficiary / growth + quality investment; focus on the data-center sales share, North American growth, M&A synergies, and margin maintenance.

Delta Electronics

Company overview and business structure. Delta is one of Asia's most complete AI data-center power and cooling platforms, covering UPS, power conversion, thermal management, liquid cooling, container data centers, and microgrids. The chairman's report explicitly states that liquid cooling has become a key growth driver in 2025; and at GTC 2025 it launched next-generation power and cooling solutions for NVIDIA AI data centers.

AI exposure. Unlike many companies that "only make one component," Delta can supply both power and cooling, and its website discloses that its data-center solutions can help lower PUE and achieve up to 30% annualized energy savings. This gives it both "direct dollar content" and an "energy-saving ROI" selling point.

Risks and conclusion. Reuters reporting in 2026 notes that Delta is also facing cost pressure from oil prices and material shortages. The research conclusion is strong beneficiary / growth investment, worth tracking the liquid-cooling revenue share, North American AI data-center projects, gross-margin pressure, and U.S. manufacturing footprint.

Vistra

Company overview and business structure. Vistra's logic is "power asset portfolio + data-center load growth + asset M&A." In 2025, the company's adjusted EBITDA was 5.912 billion dollars; in 2026 it announced the acquisition of Cogentrix's gas assets to expand its footprint in key power markets.

AI exposure path. It does not sell equipment directly like Vertiv, but the high-load regional power demand that AI drives will lift the value of its generation assets, capacity revenue, and retail hedging capability, putting it in the "clearly benefiting but via a longer path" category.

Valuation and verdict. Current market cap is about 47.7 billion dollars, P/E 23.4x. Research conclusion: moderate-to-strong beneficiary / cyclical + dividend investment. Keep tracking PJM/ERCOT load expectations, asset-M&A returns, and the power structure near load centers.

Hitachi

Company overview and business structure. Hitachi is deeply positioned in transformers, HVDC, switchgear, and grid infrastructure through Hitachi Energy. The latter explicitly added a 250 million dollar investment in 2025 to address the global transformer shortage, then announced a historic 1 billion dollar U.S. manufacturing investment, of which 457 million dollars goes to a large power-transformer plant in Virginia.

AI exposure. This type of company makes money not from "equipment placed inside AI data centers" but from "AI campuses wanting to connect to the external high-voltage grid." Because large transformers and high-voltage gear are among the clearest shortage constraints today, Hitachi's position is very prominent.

Research conclusion. I define Hitachi/Hitachi Energy as a moderate-to-strong beneficiary / cyclical + infrastructure investment; its key tracking indicators are transformer capacity, North American lead times, HVDC orders, and the ramp of the U.S. plant.

Quanta Services

Company overview and business structure. Quanta is a classic "don't underestimate the difficulty of connecting power" name. The company's IR page shows 2025 revenue of 28.5 billion dollars, of which 70% comes from utility and power, plus another 13% from technology, manufacturing, and communications.

AI exposure path. It is not a pure AI-equipment stock but an EPC and general-contracting beneficiary of AI campuses, grid expansion, and substation and line upgrades. As data-center load enters U.S. states in a more dispersed way, engineering capability like Quanta's grows more valuable. NERC also warns that large-load interconnection adds new complexity to system planning.

Valuation and verdict. Current market cap is about 117.1 billion dollars, P/E 106x. The research conclusion is moderate beneficiary / growth + infrastructure investment; the valuation is not cheap, but in "that large stretch of engineering work between the plant and the data hall that no one wants to look at yet must be done," Quanta has real research value.

Private Companies, Tiering, and Scoring

Private Companies and Primary-Market Leads

Note: in the table below, Kairos Power is a directly verified lead this round; the others are listed as highly relevant research leads, but funding amounts, valuations, and IPO timing are mostly not uniformly verified this round, so they are marked "needs further verification" and not included in this round's quantitative scoring.

Company Tiering and Investment Priority

Tier A: core direct beneficiaries in the AI energy chain. Vertiv, Eaton, Schneider Electric, GE Vernova, Siemens Energy, nVent, Delta Electronics, Envicool. Their common trait is that demand has already turned directly into orders/revenue/product iteration, and they sit in the shortest links of the value chain: distribution, UPS, liquid cooling, turbines, and grid equipment.

Tier B: clear beneficiaries, but with cyclical or valuation risk. Constellation, Vistra, Legrand, Hitachi, Prysmian, HD Hyundai Electric, Shenling Environment, Jinpan Technology. They either have slightly longer benefit chains, or valuations that have already re-rated sharply, or project-recognition cadences that are more cyclical.

Tier C: possible long-term beneficiaries, but with weaker near-term financial beta. Quanta Services, CGN Power, China Resources Power, ABB, Gotion High-tech. The reason is that they are genuinely on the benefit chain, but either AI is only part of a large pie, or the gains show up more in medium-to-long-term capex and structural demand.

Tier D: strong AI narrative, but insufficient evidence of actual benefit. Any generalized energy and electrical company that discloses no AI customers, orders, revenue share, or capacity-utilization change, yet is simply labeled an "AI power beneficiary" by the market, should be parked in Tier D rather than placed directly in the priority pool.

Scoring Model and Ranking

Scoring model. I adopt the user-suggested weights: direct AI demand exposure 30%, industry position and moat 20%, financial quality 20%, valuation reasonableness 15%, and future catalysts 15%. The scores are research judgments based on this round's public materials, not investment advice.

Ranking logic. The top-ranked companies are not "the hottest concepts," but those that satisfy three things at once: first, already validated by AI orders; second, sitting in genuine bottleneck links; third, able to convert backlog into revenue and profit. Therefore equipment/system makers rank ahead of generators overall, and transformer and distribution leaders rank ahead of generalized energy companies. Conversely, names with long benefit paths, strong regulatory constraints, or no AI revenue evidence yet -- however hot the theme -- do not score very high overall.

Risks, Conclusions, and Research Checklist

Risk Analysis

Risk that AI power demand is overestimated. The IEA noted in 2026 that the energy use per AI task is falling fast, with an unprecedented pace of efficiency improvement; therefore "training-compute demand growth" does not automatically equal "a linear explosion in generation demand." If model compression, inference-efficiency gains, and the shift of inference to the edge happen faster than expected, the power-demand curve will be smoother than the market narrative.

Risk that data-center construction slows. Uptime stressed in 2025 that the industry faces capacity-planning uncertainty, power availability, supply-chain delays, and difficulty judging AI demand; Vertiv's annual report also explicitly warns that backlog carries cancellation, deferral, and rescheduling risk. Equipment makers benefit fastest but are also the first to feel changes when projects slow.

Risk of grid-interconnection and approval delays. NERC has listed large data-center load interconnection as a new long-term reliability challenge; China is also continuously advancing compute-power coordination and green-power direct connection, showing interconnection difficulty is not unique to the U.S. This risk hits "projects under construction" first, rather than existing assets in place.

Interest-rate and capex risk. Renewables, nuclear restarts, storage, and large campus engineering all depend heavily on capital costs; if the rate environment is unfavorable, PPA return rates and the pace of campus-project execution will be dragged down. Equipment makers are relatively better off, but will ultimately be affected by downstream capex cadence transmitting through.

Volatility in power prices, gas prices, uranium prices, and equipment prices. EPRI, LBNL, and the IEA all show that the choice of power equipment and generation paths is highly sensitive to overall cost; transformer prices have already risen significantly over the past several years, while the turbine and nuclear-related chains are more exposed to commodities and policy.

Policy and regulatory risk. The Talen/AWS case shows that FERC and state regulators can directly change behind-the-meter supply business models; meanwhile, China and Europe shape revenue models through efficiency, green power, and grid-feed rules. The closer to generation and power trading, the higher the regulatory sensitivity.

Customer-concentration risk. AI infrastructure orders are highly concentrated among a few hyperscalers and large-scale colos. Dominion's materials show that just 7 of 54 customers account for 72% of its YTD demand; this implies large-customer bargaining and project-delay risk for equipment makers, utilities, and engineering firms alike.

Valuation-bubble risk. The market has already given some "AI power shovel sellers" very high premiums, for example the high P/E levels of Vertiv and Quanta Services. Even if fundamentals keep improving, valuations could retrace once order growth shifts from "above expectations" to "in line with expectations."

Technology-substitution risk. If chips, packaging, system architecture, and software scheduling keep improving efficiency, or some inference shifts to the device side, the power intensity per unit of AI compute could fall; the IEA has explicitly flagged this uncertainty.

Geopolitical and supply-chain risk. Rare earths, copper, key electrical components, and large-equipment capacity could all become new constraints. By late 2025 GE Vernova had even partnered with the U.S. government to expand yttrium reserves, showing that the energy-equipment supply chain itself is entering a period of geopolitical contention.

Final Conclusion

The importance of this link in the AI supply chain. Energy is not a "supporting role" in the AI supply chain but a hard constraint that determines the speed, location, and capital returns of the AI compute build-out. The IEA has bluntly said "no energy, no AI"; and from the latest data, AI's impact on the energy chain is no longer confined to generation, but appears more profoundly as a full re-rating of distribution, cooling, grid interconnection, long-lead-time equipment, and long-term supply contracts.

The five most noteworthy segments. First, UPS/PDU/critical distribution systems; Second, liquid cooling and high-density thermal management; Third, transformers, switchgear, and medium/high-voltage distribution; Fourth, gas turbines and campus distributed generation; Fifth, existing-nuclear restarts and long-term nuclear/SMR supply contracts. These five directions correspond to the five hardest constraints in the AI build-out: power quality, cooling, interconnection, backfill, and baseload.

The ten listed companies most worth deep research. Vertiv, Eaton, Schneider Electric, GE Vernova, Siemens Energy, Constellation Energy, nVent, Legrand, Delta Electronics, Vistra. For a U.S./Europe priority pool, Hitachi and Quanta Services can be placed in the second tier.

The five private companies most worth tracking. Kairos Power, Crusoe, LiquidStack, Submer, Enchanted Rock. Among these, Kairos Power has the highest research value, because it already has a real demand anchor like Google; the others are better treated as primary-market leads rather than names to draw conclusions on directly.

The three points most easily misunderstood by the market. First, AI power demand is large, but the first to make money is not necessarily the generators; it is the distribution and cooling equipment makers. Second, nuclear is the long-term solution, but much of the near-term increment will still be carried by natural gas and existing power sources. Third, not all "energy stocks" benefit; whether a company is near a load center and whether it has interconnection/contracting capability matters far more than "whether it owns generation assets."

The indicators most worth tracking over the next 6-12 months. First, the interconnection capacity and interconnection wait times in the U.S. and major load centers; Second, the lead-time changes of large transformers and turbines; Third, the long-term power contracts and on-campus generation projects of hyperscalers/colos; Fourth, the penetration and per-megawatt dollar content of liquid cooling in newly built AI data halls; Fifth, the order growth, backlog, cancellation rates, and gross margins of equipment makers.

Research Checklist

Macro and industry materials to read first: The IEA's Energy and AI and the 2026 Key Questions on Energy and AI related press releases; LBNL's 2024 United States Data Center Energy Usage Report; EPRI's Powering Intelligence: Updated U.S. Data Center Scenarios; NERC's 2024 Long-Term Reliability Assessment; the Uptime Institute's Global Data Center Survey 2025. These materials form the foundation for demand estimation and bottleneck judgment.

Company financials and materials to read first: Vertiv 2025 annual report; Eaton 2025 annual report; GE Vernova 2025 annual report and 2025 Investor Update; Schneider Electric 2025 Full Year Results and 2026 Q1 Revenue Presentation; Siemens Energy 2025 annual report and Q2 2025 results; Hitachi Energy 2025-2026 transformer capacity-expansion announcements; nVent 2025 Q4 deck; Delta Electronics Chairman's Statement.

Data points to monitor first: Contracted data-center load in Virginia, Texas, and Midwest U.S. states; green-power share and compute-power coordination pilots at China's hub nodes; transformer and turbine lead times; tech companies' nuclear/PPA orders; liquid-cooling project orders and penetration.

Open Questions and Limitations

This report has tried to use the latest public materials available as of 2026-05-16, but three types of limitations still need to be made explicit. First, a large number of companies do not separately disclose their AI-related revenue share, so "direct exposure" must be cross-validated using customers, orders, products, and management statements. Second, valuation and segment definitions for Chinese and some Asian companies are not uniform enough in English-language public materials, so any move into live research should return to exchange annual reports and earnings-call transcripts for line-by-line verification. Third, some private companies can only serve as a lead pool, not a conclusion pool, and need a separate round of primary-market verification. These limitations do not change the report's main judgment: in the AI energy chain, the first, most certain, and most verifiable beneficiaries are still data-center power infrastructure, liquid cooling, transformers/switchgear, and the turbine/nuclear supply chain.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.