AI Power Infrastructure Investment Map: From Order Beneficiaries to Infrastructure Platforms

A deep industry research report on data center power, liquid cooling, electrical equipment, construction, and nuclear PPAs

Analysis date: 2026-06-01 · Data cutoff: 2026-06-01 company market and valuation snapshot

Chapter 1: Opening: This Trade Is Not Really Betting That “AI Needs Power”

Covered names: VRT, ETN, GEV, POWL, HUBB, NVT, MOD, PWR, FIX, TLN, VST, CEG, NRG, as well as MU, TSM, GLW, APH, FLEX and the Taiwan supply chain as demand corroboration. Company market and valuation snapshot as of June 1, 2026. This article is industry and investment research and does not constitute buy or sell advice.

The market already knows that AI needs more electricity. The real disagreement lies downstream: how many announced data center projects can actually break ground, energize, purchase equipment, complete acceptance, and ultimately turn into operating cash flow for suppliers.

This industry chain cannot be summarized with a single “AI power” label. Liquid cooling, power distribution, and rack equipment closest to the cabinet monetize revenue the fastest and command the highest valuations, but competitive capacity expansion and customer price pressure arrive earlier. Switchgear, transformers, gas turbines, and grid-interconnection assets closer to the power grid monetize more slowly, but capacity is harder to add and pricing power lasts longer. Generation and long-term power purchase agreements are more like contract-duration assets, with valuations shaped by regulation, capacity prices, and hedging arrangements.

This report focuses on only four variables:

In one sentence, this rally has moved from “is there demand?” to “who can realize it?” In the next phase, stock-price dispersion will be driven more by cash conversion than by order headlines.

To read this industry chain, first separate three types of companies

AI power is not a single industry, but a set of businesses with completely different realization speeds and risks. If readers only remember that “data centers need electricity,” it is easy to mix liquid cooling, switchgear, construction, generation, and long-term nuclear contracts into the same trade. A more useful classification is to look at how close a company is to the rack, how long its delivery cycle is, and when cash comes back.

All three types of companies can benefit from growth in AI electricity demand, but the investment language is completely different. In near-rack equipment, the trade is realization speed. In grid equipment, the trade is scarce duration. In generation assets, the trade is long-term contracts and power-price optionality. All company comparisons that follow are built around this layering.

Chapter 2: Core Judgments

The market has accepted one fact: artificial intelligence needs electricity, and electricity is becoming scarce. The problem is that this correct fact is being traded in a dangerous way. Every company tagged with “AI power,” from liquid cooling to nuclear power and from switchgear to field construction, is being packaged into the same thematic trade, bought together, and bid up together. When a group of stocks fundamentally driven by the same underlying assumption is re-rated collectively, their drawdowns will also be collective and synchronized.

The real question is not “who benefits from AI,” but a two-step judgment. First: how much of the promised data center load will truly be financed, started, and connected to the grid? Second: among the portion that is actually built, who sits on a physical bottleneck deep enough and durable enough to turn orders into high-margin, repeatable, hard-to-replicate cash flow? In other words, who can upgrade from an order beneficiary in this cycle into an infrastructure platform for the next decade?

Around these two judgments, four core conclusions emerge.

First, demand is not as certain as the consensus assumes. In 2026, the announced U.S. data center pipeline is about 16 gigawatts, while only about 5 gigawatts is actually under construction, implying that 30% to 50% of projects face delays or cancellations. Total North American data center supply increased by about 36% year over year in 2025 to roughly 9,432 megawatts, yet vacancy still fell to about 1.4%. Capacity under construction was about 5,994 megawatts and was lower than at the end of the previous year, precisely because of constraints in power, permitting, zoning, and equipment delivery. What is really blocking construction progress is not capital, but transformers, switchgear, battery backup, and available grid power. Demand will be speed-limited by physical bottlenecks on the supply side, and speed-limited demand first hurts the companies closest to the rack, with the fastest revenue realization but the shallowest moats.

Second, value is distributed extremely unevenly across the industry chain, while market valuation multiples appear inverted relative to cost share. But only part of this inversion is a mismatch; the other part is precisely the market's correct pricing of cash-flow conversion speed. In the construction cost of an AI data center, electrical systems account for 40% to 45%, the single largest component, while cooling accounts for 15% to 20%. The pieces that can truly command pricing and truly hold up the schedule are the small subset of electrical systems with the longest lead times and the greatest difficulty in expanding capacity: transformers, gas-turbine slots, and medium-voltage switchgear. Companies selling this small subset (GE Vernova, Eaton, Powell) have the largest cost share and the deepest bottlenecks, yet the lowest valuation multiples. Companies selling liquid cooling, which has a smaller cost share and higher substitutability (Modine at about 155x, Vertiv at about 79x), have the highest valuation multiples.

It is not enough to simply attribute this phenomenon to market misjudgment. For the same order, a liquid-cooling company can manufacture, install, recognize revenue, and collect cash within a few months, while transformers and gas turbines take three to six years from order to delivery, pushing revenue and profit into financial statements years in the future. In other words, although scarce core power components have the deepest bottlenecks and the longest pricing power, their extremely long delivery cycles prevent them from rapidly converting orders into current-period cash flow. Equipment close to the rack has a shallower moat, but because it turns faster, it can convert demand of the same size into cash earlier and more densely. Part of the reason the market is willing to give liquid cooling a higher multiple is exactly its faster cash generation. The high multiple buys not only growth, but also the time value of cash return. Therefore, the real question is not “is electrical expensive or is liquid cooling expensive,” but how to put the two on the same ruler using turnover and the cash conversion cycle. Whoever can turn demand into retained cash the fastest and cleanest has a basis for a high multiple. Whoever has a high multiple that comes only from a growth narrative without support from cash conversion is the true mismatch. Later sections will use each company's turnover and cash conversion data to test this.

Third, technology is shifting along three main lines: AI from training toward inference and agents, racks from air cooling toward liquid cooling, and power distribution from AC toward 800-volt high-voltage DC. Each line is reallocating beneficiaries rather than allowing all “AI power” companies to benefit together. Training reinforces centralized, high-density factories; inference reinforces distributed regional nodes and retrofits of existing capacity. Liquid cooling reinforces cooling suppliers close to the rack, but they are also the easiest for customers to pressure by broadening supplier pools and developing in-house solutions. 800-volt DC reinforces high-voltage protection, insulation, busbars, and system integration, while reducing the value of some low-voltage copper cabling and on-site customization. When these three lines are mapped onto each company, the direction of benefit diverges, as detailed in Section 5.

Fourth, and the judgment that ties the above three points together: pricing power has duration, and that duration is precisely bounded by each company's capacity release timing. New switchgear capacity comes online in the second half of 2027, new transformer capacity comes online in 2028, and gas-turbine slots are already booked into 2029 and 2030. Liquid-cooling suppliers, by contrast, are expanding capacity as quickly as possible and will be the first to drive their own scarcity premium to zero. Companies that can expand capacity rapidly and meaningfully have high supply elasticity and quickly expiring pricing power. Companies whose capacity additions are constrained by electrical steel, production-line certification, and skilled labor, and therefore recover supply slowly, retain pricing power longer. This compresses the core mismatch into one sentence: the market has assigned the highest valuations to liquid-cooling companies with the most aggressive capacity expansion, the highest supply elasticity, and the shortest pricing power, while assigning the lowest valuations to grid companies with the slowest capacity expansion, the lowest supply elasticity, and the longest pricing power.

All company comparisons that follow revolve around the same question: realization speed and pricing-power duration often move in opposite directions. Segments where revenue arrives quickly also attract competition quickly. Segments where revenue arrives slowly may instead lock in a longer profit cycle. The valuation mismatch is hidden between the two.

---

Chapter 3: Disaggregating the Industry Chain: What Each Segment Sells and How Much It Represents

To judge who is valuable, we first need to break the broad “AI power” theme into the physical structure of an AI data center from chips to the grid, and assign an approximate share of construction cost to each segment. Without this sense of proportion, no order amount or growth rate can be compared across companies.

An AI-optimized data center costs more than about $20 million per megawatt to build, far above the roughly $11.3 million for a traditional data center. The total construction cost of a gigawatt-scale AI campus can reach $45 billion to $55 billion. This spending is distributed along a physical chain: from compute and power supply inside the rack, to cooling and power distribution in the data hall, then to campus medium-voltage electrical equipment and substations, and finally to transformers, grid interconnection, and generation outside the campus.

If construction cost (excluding IT compute equipment itself, land, and software) is divided into major segments, electrical systems account for 40% to 45%, the single largest block. Inside that are rack-level power distribution (PDUs, meaning rack power distribution units, busbars, and power shelves), white-space critical power (UPS and battery backup), and medium-voltage electrical equipment (switchgear and electrical houses). Cooling accounts for 15% to 20%, including rack-level liquid cooling (CDUs, meaning coolant distribution units, manifolds, cold plates, and rear-door heat exchangers) and room-level heat rejection. Structure, civil works, and on-site mechanical/electrical installation make up most of the remainder. Commissioning and functional testing are embedded in the segments above, but they are also the source of repeatable service revenue for platform-type companies. Further outside the campus, transformers, substations, grid interconnection, and generation are not counted in data center construction cost; they are separate grid and power-source investments. But they are prerequisites for whether the whole project can be energized, with the longest lead times and the greatest risk of delaying progress.

This breakdown corrects three common misjudgments. First, electrical systems at 40% to 45% are much larger than cooling at 15% to 20%. The market's attention to liquid cooling is disproportionate to its cost share. Liquid cooling is the most elastic and most closely watched theme, but it is the second-largest cost, not the largest. Second, redundancy directly magnifies the absolute dollars spent on electrical and cooling systems. Tier IV facilities cost 25% to 40% more than Tier III because every layer requires duplicated redundancy, and AI factories generally require high redundancy. This is the structural reason electrical equipment companies' revenue can grow faster than the number of data centers. Third, cost share is not the same as value share. Electrical accounts for 40% to 45% of cost, but value attribution depends on bottleneck depth and substitutability. The parts that can truly command pricing are the small subset with the longest lead times and the greatest difficulty in expanding capacity: transformers, gas-turbine slots, and medium-voltage switchgear.

Using this breakdown, the differences in value capture become immediately clear once each company is placed into its product segment.

Rack-level power distribution and power shelves (PDUs, busbars, power shelves, 800-volt DC): Vertiv, Eaton, Schneider, Delta, Lite-On, and Flex participate directly, while nVent participates through cable management and protection components. This segment is closest to GPUs, the most standardized, and also the easiest for open rack standards and customer self-procurement to erode. Rack-level liquid cooling (CDUs, manifolds, cold plates, rear-door heat exchangers): Vertiv, Modine, nVent, Auras, AVC, and CoolIT participate. This segment has the greatest elasticity, but customers are rapidly expanding their supplier pools. White-space critical power (UPS, battery backup): Vertiv, Eaton, and Schneider participate, with system integration and service giving them stickiness above pure components. Medium-voltage electrical equipment (switchgear, electrical houses, busbars): Eaton, Powell, Schneider, Hubbell, and nVent's prefabricated electrical houses participate. This segment has deep bottlenecks and is hard to internalize, making it the most stable point of value capture. Transformers, gas turbines, substations, and grid interconnection outside the campus: GE Vernova, Eaton, and the construction chain's Quanta participate. This segment has the longest lead times, is the least replaceable, and has the longest pricing power. Generation and long-term contracts: Talen, Vistra, Constellation, and NRG, which are not synchronized with equipment shipments and move with capacity markets and regulation.

---