Data Center Stocks: What Is Inside the Data Center Power Theme

The theme applies one test: does the company sell hardware, contracting, or materials that scale directly with data center buildouts? In practice that means power and cooling equipment, electrical infrastructure construction, mechanical and electrical contracting, site development, optical fibre, and the industrial gases consumed at hyperscale fabs and facilities.

The words doing the work are scale directly with buildouts. The test deliberately does not ask whether a company benefits from AI, because almost everything does in some diluted sense. It asks whether revenue moves with megawatts energised and square footage built. That is a narrower question, and it is the reason the roster contains contractors and equipment makers rather than the far more famous companies at either end of the chain.

The framing behind the test is that the binding constraint on AI capacity has moved. For most of the early cycle the scarce resource was accelerators. Increasingly the scarce resources are electricity, transmission capacity, skilled electrical labour, and the ability to carry heat away from a rack that draws an order of magnitude more power than a conventional one. A theme built around the current constraint owns the companies paid to relieve it.

The second structural choice is that the theme spans layers rather than picking one. No listed company covers the full stack from the grid connection to the chip, so each name captures a different slice of the same campus budget on a different schedule. Owning only equipment makers is a bet on product cycles. Owning only contractors is a bet on backlog conversion. Holding both means the roster is exposed to the same demand through different business models. For the general idea, see thematic investing.

A data center campus starts as dirt with a power problem. Before a single rack arrives, someone has to grade the site, pour the foundations, run the water and sewer tie-ins, and then get a transmission connection and substation capacity large enough to serve a load measured in hundreds of megawatts. This is the layer where the theme's core claim shows up most literally: the constraint on AI capacity is no longer how many accelerators can be manufactured, it is how quickly a site can be energised. Work here is contracted, priced per project, and lands in backlog long before the building opens.

How this layer relates to the rest. Nothing above this layer happens without it. Site preparation and grid interconnection set the calendar for the entire build, so when permitting, transmission queues, or skilled labour slow down here, every layer above sees the delay a few quarters later. It is also the layer most tied to utility capital spending rather than to hyperscaler spending directly, which gives it a slightly different funding clock from the rest of the theme.

Once a shell exists and power is at the property line, the work moves indoors. Mechanical, electrical, and plumbing contracting is the trade that wires the electrical distribution, installs the chilled-water and coolant loops, and physically connects the cooling equipment to the power equipment so the two operate as one system. It is labour-intensive, it is priced as construction rather than as product, and it is the part of the buildout that scales with square footage and megawatts rather than with chip shipments. On a modern AI campus the mechanical and electrical scope is a far larger share of the budget than it was in a conventional server hall.

How this layer relates to the rest. This layer is where the site layer's delivered power and the rack layer's equipment are joined together. It depends on the grid connection existing and on the equipment being available, and it is the layer that most directly converts a hyperscaler's announced megawatts into billed hours of skilled labour. Because it is contracted work, its economics are about backlog conversion and margin on execution, not about product cycles.

The last few metres are the hardest. A conventional server rack drew a few kilowatts and could be air cooled; a dense AI training rack draws an order of magnitude more, and air can no longer carry that much heat away. That forces a shift to liquid cooling, direct-to-chip cold plates, coolant distribution units, and rear-door heat exchangers, alongside high-density power distribution and uninterruptible supplies sized for the load. This layer sells equipment rather than labour, so its revenue is measured in units shipped and its economics look like manufacturing, not construction.

How this layer relates to the rest. This layer is the point of the whole chain. Everything below it exists to deliver electricity to this equipment, and everything above it in the compute stack depends on this equipment working. It is also the layer that changes most when rack density changes, which is why it responds to a technical shift in chip design faster than the contracting layers do.

Every layer above consumes physical material by the truckload: wire, cable, switchgear, busways, conduit, and the optical fibre that ties racks and campuses together. Somebody has to source it, stock it, and get it to a site on the day the crew needs it, and somebody has to manufacture the fibre that carries traffic between accelerators. This layer is less glamorous than the rack and less visible than the contractors, and it is the one that gets paid on almost every project regardless of which contractor won the bid or which equipment vendor was specified.

How this layer relates to the rest. This layer sits underneath all the others as a supplier rather than beside them as a peer. It has the broadest exposure to the buildout and the least exposure to any single project, which makes it the theme's quiet ballast. Its trade-off is symmetrical: it participates in every build, but it captures a thinner slice of each one.

One step further back sits the industrial gas industry. Semiconductor fabs consume large volumes of specialty electronics gases and ultra-high-purity chemistries as a continuous input to making advanced logic and memory, and large industrial sites, including data center campuses and their supporting plants, consume industrial gases too. This is adjacent exposure rather than construction exposure, and the theme labels it that way on purpose: these companies are paid on volume consumed over long contracts, not on facilities completed.

How this layer relates to the rest. This layer is the theme's slowest-moving one and it behaves least like the rest. Industrial gas revenue is typically contracted over long periods with on-site supply arrangements, so it does not spike when a campus breaks ground or fall when one is cancelled. That is precisely why it is here: it is the part of the roster least tied to the construction calendar.

Read top to bottom, the theme follows a construction sequence. A hyperscaler commits to a site, Sterling moves earth and runs the utility tie-ins, Quanta and the utilities bring transmission and substation capacity to the fence line, Comfort Systems and EMCOR wire the building and install the cooling loops, WESCO supplies the electrical material all the way through, and Vertiv's power and thermal equipment lands in the data hall near the end. Corning's fibre ties the racks and halls together, and Linde and Air Products sit further upstream at the fabs that made the silicon in the first place.

The consequence is that the nine names are paid at different points on the same calendar. Site work books at project start, contracting converts through the middle, equipment ships toward the end, and distribution earns throughout. A quarter that looks weak for one layer can be a quarter where another layer is at peak billing on the very same campus, which is why the theme reads better as a sequence than as a list.

The strongest argument for the theme is that it is a way to own AI demand without owning AI technology risk. A coolant distribution unit, a run of busway, or a substation is bought because a rack is hot and hungry, not because a particular model or a particular accelerator architecture won. If the leading lab changes, if the winning chip vendor changes, if inference displaces training as the dominant workload, this roster is largely indifferent. That is a genuinely different exposure from holding the semiconductor names.

The honest other half is that the theme trades one risk for another rather than removing risk. Every name here depends on the buildout continuing at something like its current pace, and that spending is concentrated in a handful of customers who can revise their plans without warning. A capital spending pause would hit the whole roster at once, because the layers are separated by schedule but not by demand source. The theme is diversified across business models and correlated on the thing that funds all of them. Both halves are true, and understanding both is more useful than any ranking of the nine.

A membership test is only credible if it excludes things. These are the names people most often expect to find here, and the specific reason each one does not qualify.

• NVIDIA and the accelerator designers. They are the source of the demand, not the physical plant that answers it. Including them would import exactly the risk this theme is built to avoid: which architecture wins, which foundry node arrives on time, which model provider keeps buying. They belong to the semiconductors and AI infrastructure themes, where that risk is the thesis rather than a contaminant.
• The hyperscalers (Microsoft, Amazon, Alphabet, Meta). They fund the buildout rather than get paid for it. Data center capital spending is a cost line inside enormous businesses driven by advertising, software, and retail, so their share prices respond to that spending as an expense, not as revenue. A theme about who gets paid to build cannot hold the people writing the cheques.
• Data center REITs (Digital Realty, Equinix). They own the finished building and collect lease income, which is a landlord business governed by leasing demand, financing costs, and interest rates. That is real exposure to data centers and a genuinely different one, so it sits in the REIT theme rather than here. This theme's test asks who is paid to construct, power, and cool the facility, not who owns it afterwards.
• Constellation Energy, Vistra, and the regulated utilities. They generate and sell the electricity. That is the layer immediately upstream of this theme's boundary, and it behaves differently: rate cases, power purchase agreements, and fuel costs drive it rather than construction backlog. Generation exposure is available in the utility theme, and the nuclear and small modular reactor developers courting data center customers sit in the nuclear and SMR theme.
• Eaton and GE Vernova. Both are real electrification names and both appear in the companion roundup of widely held data center stocks, which is a deliberately broader list. They are not constituents here because GE Vernova's centre of gravity is power generation equipment sold to utilities, which crosses the generation boundary above, and Eaton's data center work sits inside a much larger diversified electrical and industrial portfolio. Eaton is held in the copper and electrification theme instead, where broad electrification is the point rather than a segment.

Four of those exclusions point at neighbouring Walnut themes rather than at nothing, and the boundaries are worth naming directly. Compute and silicon sit in AI infrastructure and semiconductors. Electricity generation and its regulated economics sit in utilities, and the reactor developers courting data center offtake sit in nuclear and small modular reactors. This theme is the band in between: after the electricity is generated and before the chip does any work. A company can be an excellent business and still be the wrong expression of a given theme.

The same nine names, grouped by the layer they occupy rather than ranked, so the shape of the theme is visible at a glance.

Four of the 9 are contractors paid for labour and project execution, three sell equipment or materials by the unit, and two are upstream industrial gas suppliers on long-term contracts. That mix of business models against a single demand source is the theme's central design decision, not an accident of what happened to be listed.

The passive route runs into a problem here: there is no clean pure-play fund for this category, because index providers have not defined a tight enough rule set around it. The theme names XLU, the utilities sector fund, and XLI, the industrials sector fund, as its ETF proxies, and both are diluted by construction. XLU holds regulated generators and grid operators rather than the contractors and equipment makers on this roster. XLI does hold several of these names, at small weights inside hundreds of unrelated industrials. Some broader infrastructure and smart-grid funds include Quanta and Vertiv, again at low weights.

A theme inverts the trade. You know exactly which nine names you own, which layer each one represents, and what weight each carries, and you accept that nine names is a narrower roster than a fund holds and that you are responsible for the research. Neither is automatically better. The fund is the simpler instrument and gives broad sector exposure in one ticket, the theme is the more deliberate one, and plenty of people hold a broad fund as a core with a small thematic tilt beside it.

A list of nine names is an input, not a portfolio. What turns one into the other is structure: which layers you want exposure to, what weight each name carries, and whether the concentration you end up with was chosen or inherited.

• Decide the layer mix first, then the names. The split between direct equipment exposure and diversified contractors changes the character of the position far more than swapping one contractor for another. Comfort Systems and EMCOR are close substitutes; Vertiv and Linde are not.
• Set target weights that sum to 100. Equal weighting across nine names is a choice, and so is tilting toward the pure-plays. Both are defensible. Not deciding is what leaves you concentrated by accident after one name runs.
• Remember the correlation underneath. The layers bill on different schedules but they are funded by the same handful of customers, so this roster will not diversify away a capital spending slowdown.
• Frame it against the S&P 500. A narrow thematic position should be judged against a broad benchmark, because the extra concentration has to be buying you something. XLU or XLI is a useful second comparison.
• Revisit as weights move. Thematic positions drift fast when the constituents have this much dispersion between them, and this theme rerated sharply as the thesis took hold.

This is what Walnut is built for. You describe the thesis, the AI assistant proposes constituents and weights you can edit, the portfolio tracks as one performance line against the S&P 500, and you place trades you approve yourself at your own broker. Walnut is informational and does not tell you which stocks to buy.

For the companion view of which data center names are most widely held and discussed, including the REITs and chip names this theme deliberately excludes, see best data center stocks. For the generation layer immediately upstream of this one, see best utility stocks.

The data center power and cooling theme is nine companies across five layers of a single construction sequence, and the sequence is the whole idea. Sterling and Quanta prepare the site and bring power to it. Comfort Systems and EMCOR wire and cool the building. Vertiv delivers power and removes heat at the rack, which is the hardest few metres in the chain. WESCO and Corning supply the material and the fibre that every stage consumes. Linde and Air Products sit upstream at the fabs, on contracted volume rather than project awards.

Understood as a flat list of nine industrials, the theme looks like an odd collection of contractors and gas suppliers. Understood as a way to own AI demand through the physical plant instead of through the technology, it is a specific and coherent structure: paid regardless of which model or which chip wins, and dependent on a capital spending cycle that can slow. Both of those are true at once, and the structure is what you are deciding whether to own. Nothing here is a recommendation, and Walnut is not an investment adviser.