Data Center Infrastructure Investing — The Picks and Shovels of the AI Buildout

• Introduction: Who Got Rich in the Gold Rush
• What a Data Center Actually Is: The Big Picture
• Value Chain 1: Real Estate and REITs
  • Wholesale Versus Colocation
  • Equinix Versus Digital Realty in Depth
  • FFO and REIT Valuation Metrics
• Value Chain 2: Power
  • Distribution Equipment Supply Chain and Lead Times
  • Diversifying Power Sources
• Value Chain 3: Cooling
  • PUE, Cost, and Adoption Stage Compared
• Value Chain 4: Networking
  • InfiniBand, Ethernet, and Optical Transceivers
• Value Chain 5: Servers and Semiconductors
• Location: Power, Water, Land
  • Regional Location Constraints Compared
• Hyperscaler Capex Trends
• Diverse Perspectives: The Bull and Bear Cases
  • The Core Bull Case
  • The Core Bear Case
• Cycle Risk: The Shadow of Overinvestment
  • Lessons From the Dot-Com Fiber Glut
  • An AI Demand-Slowdown Scenario
• Comparing Investment Approaches
• Risks and Checkpoints
  • Additional Checkpoints: Layer-Specific Detail
• Correcting Common Misconceptions
• Closing
• References

Introduction: Who Got Rich in the Gold Rush

During the 19th-century California gold rush, most people who rushed in to dig for gold never struck it rich. The ones who made steady money were the merchants selling picks, shovels, jeans, and groceries. That story has become a kind of proverb in the investing world, captured in the phrase picks and shovels investing.

Since 2023, as generative AI has grown explosively, a similar question keeps coming back. No one knows who the ultimate winner of the AI gold rush will be. Which model, which application survives is genuinely uncertain. But there is one thing every AI model has in common: it needs an enormous amount of computational capacity housed in physical space, which means data centers and the infrastructure that keeps them running.

This article breaks down the data center infrastructure value chain step by step, looking at where attention is concentrating, which companies are being discussed in each layer, and what the risks are. The goal is not to tell anyone to buy or sell a particular stock, but to explore in a balanced way why each area attracts interest and what makes it risky.

This article is for informational and educational purposes only and is not investment advice or a recommendation. Investment decisions and their consequences are your own responsibility. Consult a qualified professional when needed.

What a Data Center Actually Is: The Big Picture

A data center is not just a warehouse with servers stacked inside. It is a complex facility that combines stable power delivery, continuous cooling, ultra-fast networking, physical security, and the operational systems that keep all of this running without interruption. Facilities used for AI training and inference are particularly power-dense, which makes their design fundamentally different from traditional data centers.

A rough structure looks like this.

                  [ External grid / power contract ]
                            |
                  +---------v---------+
                  |  substation /     |
                  |  distribution     |
                  |  (UPS, generator) |
                  +---------+---------+
                            |
        +-------------------+-------------------+
        |                   |                   |
   +----v----+         +----v----+         +----v----+
   | cooling |         | servers |         | network |
   | (air /  |<------->| (GPU/CPU|<------->| (switch/|
   |  liquid)|         | storage)|         |  fiber) |
   +----+----+         +----+----+         +----+----+
        |                   |                   |
        +-------------------+-------------------+
                            |
                  +---------v---------+
                  | building / site   |
                  | (REIT, land, water|
                  +-------------------+

Each box in this diagram can be seen as one investment area: owning and leasing real estate, handling power, providing cooling, supplying networking gear, and building servers and semiconductors. Market attention usually fixates on the flashiest GPU chips, but through a picks-and-shovels lens, the key insight is that the entire surrounding infrastructure grows together.

Value Chain 1: Real Estate and REITs

The most foundational layer of a data center is ultimately the building and the land. The companies most often cited in this area are data center REITs (real estate investment trusts). A REIT is a company structured to own real estate and distribute rental income to investors as dividends.

Among data center REITs, the names most frequently mentioned in the market are Equinix and Digital Realty. Equinix is known for operating interconnection-focused data centers distributed across major cities worldwide, while Digital Realty is known for handling large wholesale facilities alongside hyperscale customers.

The characteristics of the REIT model can be summarized as follows.

A REIT pursues both bond-like steady dividends and the growth potential of real estate, but its sensitivity to interest rates matters a great deal. When rates rise, borrowing costs climb and the relative appeal of steady dividends falls, so the share price tends to come under pressure, as many outlets have analyzed.

Wholesale Versus Colocation

Data center leasing can be divided broadly into two models.

Wholesale                  Colocation
-------------------        ----------------------
large single customer      many small customers
megawatt-scale leasing     rack or cage scale
hyperscaler-centric        mixed enterprise/network
long, large contracts      relatively flexible
lower but stable margins   higher margins, heavy sales

Wholesale leases large blocks at once, so revenue is large and stable, but dependence on a small number of customers can rise. Colocation spreads customers and lowers risk, but requires more sales and operational effort. Neither model is inherently superior; the difference is which market the company targets.

Equinix Versus Digital Realty in Depth

Though grouped together as data center REITs, the two companies have quite different business models, as many outlets have analyzed. The key differences can be summarized as follows.

Interconnection refers to a structure where many customers connect directly to one another inside a single facility. Once many companies gather in one place, new customers also want into that ecosystem, creating a network effect, as is well known. This is often cited as the moat of retail colocation. Hyperscale wholesale, by contrast, is a model of leasing enormous capacity wholesale to a single cloud operator: contracts are large and long, but they carry price competition and customer-concentration risk.

FFO and REIT Valuation Metrics

REITs are often evaluated with metrics different from ordinary companies. The most frequently used is FFO (Funds From Operations).

REIT key metric flow (concept)

 net income
   + depreciation (real estate tends to hold/gain value)
   - gains on property sales
   = FFO (proxy for cash-generating power)

 further adjust FFO for maintenance capex etc. -> AFFO
 dividend sustainability = dividend / AFFO (lower is safer)

Ordinary companies are judged on net income (EPS), but real estate often carries large accounting depreciation even as the asset value holds or rises. So FFO, which reverses depreciation, is seen as a better gauge of a REIT's cash-generating power. If dividends are too high relative to AFFO, dividend sustainability comes into question, so checking this ratio is standard when looking at a REIT.

Value Chain 2: Power

When talking about AI data centers, the hottest topic lately is unquestionably power. The high-density server racks used in AI training consume far more electricity than those in conventional data centers. Where a typical server rack once drew a few kilowatts, the latest AI racks have reportedly exceeded tens of kilowatts and in some cases more than 100 kilowatts.

The International Energy Agency (IEA) has, across several reports, projected that data center power demand will grow rapidly over the coming years. Such projections have drawn broad market attention to power infrastructure companies, power equipment makers, and even power generators.

Companies frequently mentioned in the power area include Eaton and Vertiv. Eaton is known for handling a broad range of power management and distribution equipment, while Vertiv is known for specializing in power and cooling infrastructure for data centers.

Power is frequently cited as the bottleneck of data centers. Even when a site is secured and a building can be erected, the facility cannot operate without enough power delivered on time. In some regions, grid interconnection wait times have reportedly stretched to several years. As a result, the very ability to secure power is emerging as a competitive edge in the data center business, according to some analyses.

Distribution Equipment Supply Chain and Lead Times

The power area is not just about power plants. The equipment that takes electricity in and distributes it safely throughout a facility is central, and shortages and long lead times for this gear have been frequently reported lately.

Vertiv, a company specializing in data center power and cooling, has reportedly seen its order backlog grow sharply during the AI demand phase. Eaton and Schneider Electric are large companies handling a broad range of power management and distribution gear, and are seen as positioned to ride both the broad industrial electrification trend and data center demand at once. The key variable in this area, however, is lead time. When delivery of equipment like transformers and switchgear stretches out, data center completion itself can be delayed. This is an order tailwind for equipment makers, but it also acts as a constraint on the speed of the buildout as a whole.

Impact of power-equipment lead times (concept)

 GPUs secured OK -----+
 site/building OK ----+---> but transformer/switchgear delays
 power contract OK ---+          |
                                 v
                  go-live pushed back -> revenue recognition delayed
                  (equipment orders rise but sites wait)

Diversifying Power Sources

As power demand surges, the discussion around power sources is active. Some large technology companies have reportedly shown interest in nuclear power, particularly small modular reactors (SMRs), to secure stable, low-carbon electricity. At the same time, approaches combining solar, wind, and energy storage are being discussed. That said, these newer generation technologies still carry significant uncertainty in terms of commercialization and regulation, which deserves a balanced view.

Data center power source options (conceptual)

Natural gas ████████  stable, but carbon
Solar       ██████    cheap but intermittent
Wind        █████     location dependent
Nuclear     ████      stable, long build time
SMR         ██        high potential, early stage
* Bars are a simple sketch of discussion frequency, not actual generation share.

Value Chain 3: Cooling

When power rises, heat inevitably rises. A large portion of the electricity consumed by AI servers ultimately turns into heat, and if that heat cannot be removed, equipment fails or performance degrades. That is why cooling is becoming an ever more important area in data center infrastructure.

Traditional data centers relied mostly on air cooling. But as the power density of AI racks has risen, many have pointed out the limits of air alone, and liquid-based cooling methods are drawing attention.

Cooling methods are broadly distinguished as follows.

Direct-to-chip liquid cooling attaches a cold plate with cold liquid flowing over high-heat chips. Immersion cooling is the more radical approach of submerging entire servers in a special non-conductive liquid. Immersion cooling is highly efficient but raises the barrier to adoption because it fundamentally changes data center design and operation, by some assessments.

Air vs liquid cooling concept

[air]
 cold air -> servers -> hot air -> CRAC unit -> cold air again
 (fan dependent, noisy, power heavy, limits at high density)

[direct-to-chip]
 coolant -> cold plate (on chip) -> heat exchanger -> coolant
 (handles high density, leak management needed)

[immersion]
 entire server submerged in dielectric -> liquid absorbs heat
 (highest density, design/maintenance paradigm shift)

In the cooling area, beyond the previously mentioned Vertiv, a variety of specialist companies and component suppliers are discussed. Cooling was once treated as a secondary element, but in the AI era many see it elevated to a core variable that determines data center performance.

PUE, Cost, and Adoption Stage Compared

A metric that comes up often when comparing cooling methods is PUE (Power Usage Effectiveness). It is total facility power divided by the power IT equipment actually uses; the closer to 1.0, the less waste in cooling and other overhead power.

Here the PUE tendency and cost simplify the generally discussed direction; actual figures vary greatly with design, location, and climate. The key point is that as racks move toward high-density AI, air alone hits growing limits, and direct-to-chip cooling is rapidly establishing itself as the realistic mainstream alternative. Immersion cooling has the greatest efficiency potential, but because it requires overhauling the entire design, operations, and maintenance paradigm, it is still seen as adopted only in limited ways.

Cooling method adoption curve (concept)

 adoption
  high |  air  ████████████████ (mature)
       |  DLC  ██████████░░░░░░ (spreading)
       |  imm. ███░░░░░░░░░░░░░ (early)
  low  +----------------------------> time
 * Bars are a simple sketch of relative adoption stage.

Water use is also tied directly to cooling. Some cooling methods consume large amounts of water through evaporative cooling towers, and in water-scarce regions this has been reported as a source of environmental regulation and community pushback. So closed-loop liquid cooling that uses little water is also drawing attention as a way to ease location constraints.

Value Chain 4: Networking

For the many GPUs inside a data center to cooperate in training a giant AI model, an ultra-fast network linking them is essential. No matter how fast a single server is internally, slow communication between servers drags down overall training speed. That is why internal data center networking is considered a hidden core of AI infrastructure.

A company frequently mentioned in this area is Arista Networks. Arista is known for high-performance data center switches and network software, and is known for holding many hyperscale customers. Beyond that, various component makers producing optical transceivers, fiber cables, and network interface cards are discussed as well.

The competition over networking standards is an interesting thing to watch as well. Several outlets have reported that a certain proprietary technology and an open Ethernet-based technology are competing for leadership in AI data center networking. Which becomes the standard could reshape the landscape of beneficiary companies.

InfiniBand, Ethernet, and Optical Transceivers

In the back-end network of AI training clusters, the picture is often drawn as a contest between proprietary high-performance technology represented by InfiniBand and the open Ethernet camp.

A key variable is the hyperscalers' motivation to reduce vendor lock-in. As reports grow that they are actively pushing open Ethernet to lower single-supplier dependence, competition between the two camps is intensifying. Which side prevails could reshape the beneficiary landscape not only for switch makers but for the entire component ecosystem.

A component worth special attention is the optical transceiver. As GPU counts rise and data transfer speeds climb, demand for transceivers that send and receive optical signals can grow even faster than GPU growth itself, by some analyses, because each accelerator carries several high-speed optical links. So optical-component and transceiver suppliers are sometimes called the picks and shovels within the picks and shovels. That said, this area too sees fast generational technology shifts and fierce price competition, which should be kept in view.

Value Chain 5: Servers and Semiconductors

The best-known layer is ultimately servers and the semiconductors inside them. In the GPU field that is the heart of AI computation, Nvidia is widely known to sit at the center of the market. That said, dependence on a single GPU company and shifts in the competitive landscape are always variables to watch carefully.

In the area of assembling and integrating servers for delivery to customers, Super Micro is frequently mentioned. Super Micro has reportedly drawn attention for supplying AI servers, particularly high-density server systems with liquid cooling. Because this area is growing fast, issues such as margin pressure, component-supply dependence, and accounting transparency tend to be discussed alongside it.

This is the flashiest part of the picks-and-shovels metaphor, but it is also the area with the highest valuations and the most expectation already priced in. As flashy as it is, the volatility from disappointment can be large too, which deserves a balanced acknowledgment.

Location: Power, Water, Land

Where to build a data center is a decision that can make or break the business. A good location is generally said to satisfy three conditions at once.

The three conditions of a good data center site

   [ power ]          [ water ]          [ land ]
 ample and cheap     for cooling, or     wide, expandable
 stable supply       alternative cooling regulation-friendly
      \                  |                  /
       \                 |                 /
        +---------> suitable site <--------+
                 (plus network connectivity)

Power, as emphasized earlier, is the key bottleneck. Water is needed for cooling, but in water-scarce regions it can become a source of environmental and community conflict, as has been reported. Land must be wide and expandable, and local government regulation and incentives matter too. Add network connectivity and latency, and sites that satisfy every condition are growing scarce, by some analyses. This scarcity can favor operators who lock up good sites early, but it can also be a cause of new-supply delays.

Regional Location Constraints Compared

The regions where data centers cluster each carry different strengths and constraints, as many outlets have analyzed. Conceptually they can be laid out as follows.

These constraints show that data center siting is not decided simply by where land is cheap. Power availability, water regulation, community acceptance, and network connectivity all act together. In particular, with reports of growing community pushback and tightening regulation over data centers' power and water consumption, securing locations is becoming more difficult, by some analyses.

Hyperscaler Capex Trends

The largest source of data center infrastructure demand is ultimately the capital expenditure (capex) of large cloud operators, the hyperscalers. Reports have continued that the capex Microsoft, Amazon, Google, and Meta are pouring into AI infrastructure keeps rising. Because their spending flows directly into revenue across the entire data center value chain, the capex trend is considered one of the most important leading indicators for this theme.

Hyperscaler combined capex trend (concept sketch)

 scale
 large |                          ████
       |                    ████  ████
       |              ████  ████  ████
       |        ████  ████  ████  ████
 small +-----------------------------------> time
        past    ->    ->    ->    recent
 * Bars are a simple sketch of an upward trend, not actual figures.

This trend has two sides. Bulls hold that as long as capex guidance keeps being revised up, infrastructure demand is firm. Bears warn that this spending cannot rise forever, and that if capex growth slows or rolls over at some point, the whole value chain could take a simultaneous hit. So the capex guidance hyperscalers give in quarterly earnings, and shifts in its tone, are closely watched as a key variable that sets the mood for infrastructure names as a whole.

Also, even if capex rises, if it pivots toward in-house chip development or efficiency, it can become a headwind for certain outside suppliers. The hyperscalers' moves to expand their own accelerators are often discussed in this context. In other words, not just total capex but the composition of where that spending is directed must be watched together.

Diverse Perspectives: The Bull and Bear Cases

There are both bull and bear cases for data center infrastructure investing in the market. Seeing only one side is risky, so let us lay them out side by side.

The Core Bull Case

The bull case rests on the view that the AI buildout is only just beginning and that demand is structural.

The last argument in particular is the heart of the picks-and-shovels frame. Even without picking the ultimate AI winner, investing in the infrastructure they all commonly depend on can capture broad benefit, the logic goes.

The Core Bear Case

The bear case warns that current investment enthusiasm may be overheated and that late-cycle overinvestment is a risk.

The bear case in particular often cites the past telecom infrastructure bubble. The experience of the late 1990s, when fiber-optic cable was laid far ahead of future demand and then suffered a long oversupply, is often invoked as grounds for caution toward the AI data center buildout.

Cycle Risk: The Shadow of Overinvestment

The most recurring risk in infrastructure investing is the cycle. When demand is strong, everyone expands at once, and as a result, a few years later supply outpaces demand, and prices and profitability fall sharply.

Infrastructure investment cycle (simple sketch)

 demand seen -> big investment -> supply surge -> oversupply -> price drop
     ^                                                            |
     |                                                            v
     +<------------- investment pullback / restructuring <--------+

No one can say with certainty where in the cycle the data center buildout currently sits. Bulls see it as still early; bears see overheating signals already. What matters is acknowledging the cycle exists and not betting on a single-direction scenario.

Another point worth noting is the difference between backlog and actual revenue. News of large orders comes frequently, but there is no guarantee that every announced contract will be executed as planned. If the economy or financing environment changes, some plans may be delayed or scaled back.

Lessons From the Dot-Com Fiber Glut

The historical analogy the bear case loves most is the telecom infrastructure bubble of the late 1990s. Under projections that internet traffic would explode, many carriers laid fiber-optic networks all at once. In the end, the demand outlook itself proved right over the long run, but short-term supply ran so far ahead of demand that huge amounts of fiber sat unlit (so-called dark fiber) for a long time, as is well known.

How overinvestment hits profitability (concept)

 high utilization + strong prices  =  high profitability  (early boom)
        |
        v  (everyone expands at once)
 supply surge -> falling utilization + weak prices = collapsing profit (glut)
        |
        v
 depreciation stays the same -> risk of swinging to losses

Two lessons stand out from this analogy. First, even if the long-term demand outlook is right, short-term overinvestment can absolutely happen. Second, infrastructure assets, once built, keep generating depreciation like a fixed cost, so when utilization falls, profitability deteriorates quickly. Whether this analogy applies directly to AI data centers is debatable, but viewing utilization, depreciation, and price trends together is clearly valid.

An AI Demand-Slowdown Scenario

For balance, it helps to map out concretely the path by which the bear scenario could materialize.

If these triggers overlap at once, capex falls and utilization drops, and the overinvestment cycle seen earlier can kick in. Conversely, bulls counter that inference demand takes over from training demand and grows structurally, so efficiency gains do not directly lead to lower demand (the so-called Jevons paradox). Which side is right cannot be known in advance, which is exactly why avoiding a one-direction bet matters.

Comparing Investment Approaches

Even within the same data center theme, the character of risk and return changes greatly depending on which layer of the value chain you are exposed to. Laid side by side, the main approach areas look like this.

Broadly, the lower the value-chain layer (real estate, power), the lower the volatility and the more dividend-like; the higher the layer (servers, semiconductors), the greater both growth and volatility tend to be. No layer is the right answer; the question is which fits your risk tolerance and time horizon.

Risk-return character by value-chain layer (concept)

 high vol.  | servers/semis ●
            | networking      ●
            | cooling/power      ●
 low vol.   | REITs/real estate    ●
            +-------------------------> expected growth
 * Positions are a simple sketch of general tendencies.

This comparison is also an exercise in reapplying the bull and bear cases layer by layer. For example, under the same bear scenario (capex slowdown), equipment and semiconductor names that depend on new orders tend to react more sensitively than REITs locked into long leases. Conversely, under a rising-rate scenario, high-debt REITs may be more vulnerable. In other words, recognizing that the same risk works differently in each layer is the starting point of diversification.

Risks and Checkpoints

If you are interested in this area, it tends to help more to run through the following checkpoints yourself rather than following stock recommendations.

Here are some general principles worth adding too.

General review flow for infrastructure-theme investing

 1. Where in the value chain is this company
 2. What are the barriers to entry and how durable are they
 3. Is demand structural or a temporary boom
 4. How exposed is it to cycle risk
 5. Does the price already reflect the good scenario
 6. Does it fit my time horizon and risk tolerance

Diversification and time horizon matter too. Rather than concentrating on a single stock or a single area, approaches that diversify across the whole value chain or use related exchange-traded funds (ETFs) are commonly discussed. But whatever the approach, you should first assess whether it fits your situation and whether you fully understand the costs and risks.

Additional Checkpoints: Layer-Specific Detail

Beyond the general checkpoints above, here are items worth special attention by value-chain layer.

The items in this table are not to be checked once and forgotten; they have value in tracking which direction the trend moves each quarter. For example, if an equipment maker's backlog peaks and begins to slow, that can be an early signal of a cycle turn. Conversely, if the share of liquid-cooling adoption keeps rising, it may be worth re-examining the companies that benefit from that trend.

In addition, it helps to watch macro gauges that take the temperature of the whole theme. The direction of hyperscalers' combined capex guidance, the trend in power and equipment lead times, and indicators like new data center groundbreakings and permits help gauge where the cycle as a whole stands, beyond individual stocks.

Correcting Common Misconceptions

Here are a few misconceptions that often appear when this topic is discussed.

First, the idea that since AI is growing, related infrastructure stocks will rise no matter what. Even as an industry grows, a particular company's share price can fall depending on valuation, competition, and the cycle. Industry growth and individual-stock returns are not the same thing.

Second, the idea that picks and shovels are safe. Infrastructure may tend to be less volatile than the end product, but it is by no means risk-free. When an overinvestment cycle arrives, infrastructure companies take a big hit too.

Third, the impatience that if you do not get in now you will be late. There are always new opportunities in the market, and impulsive decisions rarely beat careful ones. The process of gathering enough information and judging by your own criteria matters.

Closing

The AI buildout is clearly a current that defines an era. And within that current, data center infrastructure, the value chain running from real estate to power, cooling, networking, and servers, occupies the position of picks and shovels. The appeal of this frame is that you can aim for broad benefit even without picking the ultimate winner.

But at the same time, as with all infrastructure investing, the shadows of the cycle, overinvestment, and high expectations follow along. Understanding both the bull and bear cases, and checking your position in the value chain and the risks rather than the stock itself, will likely help more over the long term.

The most important thing is that none of this discussion makes the decision for you. The process of gathering enough information, comparing both perspectives, and judging by your own criteria is itself the most important infrastructure an investor can have.

To emphasize once more: this article is for informational and educational purposes only and is not investment advice or a recommendation. Investment decisions and their consequences are your own responsibility. Consult a qualified professional when needed.

References

• Reuters - data center and AI infrastructure coverage
• Bloomberg - Technology section
• CNBC - Technology section
• The Wall Street Journal - Tech
• Financial Times - Technology
• International Energy Agency - power and data center analysis
• Gartner - IT infrastructure research
• Yahoo Finance - stock and market data
• Equinix - official company site
• Digital Realty - official company site