Are AI Data Centers Becoming a Power Stock Story?

The AI trade is no longer only about GPUs, models, or who has the best chatbot. It is becoming a power stock story.

That does not mean Nvidia, AMD, Broadcom, Microsoft, Amazon, Alphabet, Meta, and Oracle suddenly stop mattering. They still sit at the center of the AI value chain. But the market is starting to notice a harder constraint underneath the whole boom: data centers need electricity, cooling, transformers, switchgear, land, permits, grid connections, and financing before a single AI model can run at scale.

The simple version is this: AI demand is digital, but AI capacity is physical. Investors who only follow chips may miss the second-order winners. The more AI usage expands, the more the bottleneck moves into power equipment, data-center real estate, utility infrastructure, and cloud capacity. That is why Vertiv, GE Vernova, Eaton, Quanta Services, Equinix, Digital Realty, Constellation, Vistra, and the hyperscalers deserve to be analyzed together.

The numbers explain why. The IEA says global data-center electricity demand grew 17% in 2025, AI-focused data-center electricity demand grew 50%, and the sector is projected to roughly double from 485 TWh in 2025 to about 950 TWh by 2030. The same report says capital expenditure from the largest technology companies exceeded $400 billion in 2025 and is expected to jump another 75% in 2026.

This is the key investor question: if AI data centers are now a power problem, which stocks actually turn that problem into profit?

The first answer: data centers are already a huge global footprint

There is no perfect census of data centers. Governments do not require one central registry, private enterprise facilities are often undisclosed, and some directories count campuses differently from individual buildings. That matters, because a single gigawatt campus can be more economically important than dozens of small colocation sites.

Still, the best live directories show the scale. Data Center Map lists 11,414 data centers across 179 countries. Its own methodology says the count is based on unique facilities and that coverage is strongest for colocation, cloud, connectivity, hyperscale, selected crypto, and selected large operators, while enterprise and government coverage is more limited.

For country-level comparison, Cloudscene data cited by Euronews puts the United States far ahead with 5,427 data centers in 2025. Germany follows with 529, the United Kingdom with 523, China with 449, Canada with 337, France with 322, Australia with 314, the Netherlands with 298, Russia with 251, and Japan with 222.

The United States is not just leading by count. It is leading by market structure. It has the hyperscalers, the deepest capital markets, the largest cloud customer base, and the biggest AI chip deployment pipeline. That is why data-center power demand has become a U.S. equity market story instead of only an infrastructure story.

But the country table also carries a warning. Facility count is not capacity. Germany or the U.K. can have many facilities without matching the U.S. hyperscale footprint. China can look smaller by facility count while still being strategically important in compute, power planning, and state-led AI infrastructure. Investors should treat country counts as a map of footprint, not a clean ranking of AI compute.

The second answer: data creation keeps rising

The demand side is not only AI training. It is video, enterprise cloud migration, software telemetry, connected devices, streaming, search, e-commerce, cybersecurity logs, synthetic data, and now AI inference at consumer scale.

Industry data trackers estimate that roughly 221 zettabytes of data could be generated in 2026, according to Exploding Topics. That implies about 18.4 zettabytes per month, or about 0.61 zettabytes per day. The exact number should be treated as a forecast, not a utility-meter reading, but the direction is not controversial: data creation is still compounding.

The more important stock-market point is that created data does not equal stored data. Some data is temporary. Some is deduplicated. Some is compressed. Some is processed and discarded. But AI changes the economics because useful data is increasingly copied, labeled, embedded, searched, trained against, cached, and used again for inference. Data centers therefore become both storage assets and compute factories.

That is why memory, optical networking, power conversion, cooling, real estate, and cloud leasing can all move together when AI demand accelerates. We already covered the hardware side through Corning, but the broader story is that power is becoming the gating layer above chips.

How many more data centers are required?

The honest answer is that the world does not need a fixed number of additional buildings. It needs powered capacity.

That is the distinction most generic data-center articles miss. An old 5 MW facility and a planned 300 MW AI campus are both “one data center” in casual language, but they are completely different investment objects. For investors, megawatts matter more than addresses.

JLL projects that the global data-center sector will add 97 GW of capacity between 2026 and 2030, effectively doubling to about 200 GW by 2030. It also estimates the Americas will represent roughly half of global capacity, with the U.S. accounting for about 90% of the Americas capacity base.

Translate that into investor math. Ninety-seven gigawatts equals 970 new 100 MW data-center equivalents. If the average future AI campus is closer to 250 MW, the same capacity is roughly 388 campus equivalents. If capacity comes through smaller 50 MW expansions, the number rises to nearly 1,940 equivalents.

That does not mean 970 brand-new sites will be built. It means the world needs the equivalent of that much powered capacity through new campuses, expansions, leased halls, hyperscaler-owned facilities, and grid upgrades. The limiting factor is no longer just real estate. It is whether electricity, cooling, and equipment arrive on time.

The U.S. pipeline shows the same tension. Pew found more than 3,000 operational U.S. data centers and more than 1,500 in development as of February 19, 2026. Pew also found that 67% of planned U.S. data centers are in rural areas, while 87% of current operating sites are urban. That shift tells investors where the next friction point is likely to appear: power lines, land-use politics, water, local rates, and rural community opposition.

Delays are already visible. Axios reported that up to 11 GW of 2026 global data-center capacity remained announced but not under construction, while nearly 6 GW came online in 2025 and 5 GW was already under construction in 2026. In plain English: demand is huge, but the construction queue is not frictionless.

The companies making money right now

The first profit pool is cloud capacity. This is where the hyperscalers already monetize the boom.

Amazon is the cleanest example. Amazon reported Q1 2026 AWS sales of $37.6 billion, up 28% year over year, and AWS operating income of $14.2 billion. The same release says Amazon’s free cash flow fell sharply because purchases of property and equipment rose by $59.3 billion year over year, primarily reflecting AI investments. That is the AI data-center trade in one sentence: high-margin cloud profits on one side, giant physical capex on the other.

Alphabet is also converting AI infrastructure into reported cloud profit. Alphabet said Google Cloud revenue rose 63% to $20.0 billion in Q1 2026, while Google Cloud operating income reached $6.6 billion. Alphabet also said cloud backlog nearly doubled quarter on quarter to more than $460 billion. That is why our earlier Google capex question matters: the spend is enormous, but the profit engine is no longer hypothetical.

Microsoft reports the story differently, but the pattern is the same. Microsoft said Microsoft Cloud revenue reached $54.5 billion in fiscal Q3 2026, up 29%, and Azure and other cloud services revenue rose 40%. Its cloud segment notes that Intelligent Cloud revenue increased 30%, operating income increased 24%, and cost of revenue rose because of AI infrastructure investments. The takeaway: AI capacity is driving growth, but it is also pressuring gross margin through depreciation and buildout costs.

Oracle is the more aggressive AI-cloud backlog story. Oracle reported fiscal Q2 2026 cloud infrastructure revenue of $4.1 billion, up 68%, and total remaining performance obligations of $523 billion. Oracle also said it had more than 211 live and planned regions worldwide and was building multicloud data centers inside Amazon, Google, and Microsoft cloud environments. That is a direct bet that demand for GPU clusters and database-adjacent AI workloads will remain capacity-constrained.

The second profit pool is equipment. Vertiv is the obvious cooling and power-management winner. Vertiv reported Q1 2026 sales of $2.65 billion, up 30%, with operating profit up 51% and adjusted operating profit up 64%. The Americas region grew 44% organically, driven by data-center demand. Vertiv also raised full-year guidance to $13.5 billion to $14.0 billion in net sales.

GE Vernova is where AI meets the grid. GE Vernova reported Q1 2026 orders of $18.3 billion, up 71% organically, revenue of $9.3 billion, and $2.4 billion of Electrification equipment orders to support data centers in the quarter. That single data-center order figure exceeded all of last year, according to the company. This is why TECHi’s earlier GE Vernova piece framed it as the real AI data-center trade.

Eaton sits in the electrical distribution layer. Eaton reported record Q1 2026 Electrical Americas sales of $3.6 billion and operating profit of $922 million, while total backlog in that segment was up 44% from March 2025. Eaton also closed acquisitions including Boyd Thermal, which supplies thermal systems used in data centers and other markets.

Quanta Services is the grid-and-large-load execution story. Quanta reported Q1 2026 revenue of $7.87 billion, adjusted EBITDA of $686.4 million, and total backlog of $48.5 billion. Management also described a $2.4 trillion addressable market through 2030 across utility, generation, and large-load markets. That is not pure AI, but AI data centers are now one of the large-load forces pulling that backlog forward.

The third profit pool is data-center real estate. Equinix and Digital Realty do not make GPUs or turbines; they own and operate the neutral facilities, interconnection hubs, and hyperscale platforms where demand shows up as leases.

Equinix reported Q1 2026 revenue of $2.44 billion, adjusted EBITDA of $1.25 billion, and AFFO of $1.07 billion. The company said roughly 60% of its largest deals were AI-related, and it raised full-year guidance. Digital Realty reported Q1 2026 revenue of $1.6 billion, adjusted EBITDA of $920 million, and $707 million of annualized GAAP base-rent bookings at 100% share, including the largest hyperscale lease in its history.

This is why the most profitable data-center companies are not all in the same sector. AWS and Google Cloud generate operating income from compute sold to customers. Vertiv and Eaton generate profit from power and cooling equipment. GE Vernova and Quanta sell into the grid buildout. Equinix and Digital Realty convert scarcity into recurring rent and interconnection revenue.

Why the power trade may be more durable than the first AI trade

The first AI stock trade was straightforward: buy the companies that sell chips and servers. That worked because GPU demand exploded faster than supply.

The second trade is less obvious but potentially more durable. A GPU can be ordered faster than a transmission line can be permitted. A model can be released faster than a substation can be upgraded. A cloud customer can sign a contract faster than a gas turbine or transformer can be delivered. That timing mismatch is the investment setup.

The IEA demand report says data-center electricity consumption is projected to grow about four times faster than electricity consumption from all other sectors between 2024 and 2030. It also says accelerated servers, mainly driven by AI, account for almost half of the net increase in global data-center electricity consumption.

That means the power trade is not a slogan. It is an earnings bridge. More AI workloads require more accelerated servers. More accelerated servers require denser racks. Denser racks require better cooling, higher-voltage distribution, batteries, switchgear, transformers, backup power, and larger grid interconnections. Each step creates revenue for a different stock-market cohort.

The story also connects to financing. Data centers are capital-heavy assets. We saw that in the bond market through Alphabet bonds, and we saw it again in the neocloud story through CoreWeave. High backlog is powerful, but it can become dangerous if debt, depreciation, or customer concentration rises faster than cash flow.

That is the investor trap. The AI power trade is real, but not every company exposed to data centers is equally attractive. The best businesses either own scarce capacity, have pricing power in critical equipment, or can convert capex into high-margin recurring revenue. The weakest ones may only own the debt.

The power winners are not the same as the compute winners

For cloud companies, the key metric is utilization. An AI data center is expensive before it is profitable. If capacity fills quickly with paying customers, the hyperscaler earns an infrastructure return. If demand disappoints, depreciation becomes a margin headwind.

For equipment suppliers, the key metric is backlog quality. Vertiv, Eaton, GE Vernova, and Quanta can all benefit from AI infrastructure, but investors should separate contracted orders from loose pipeline commentary. The difference matters when supply chains tighten or customers delay projects.

For REITs, the key metric is lease spread and funding cost. Equinix and Digital Realty can benefit from scarce powered shells, but their model still depends on capital markets. Higher rates, equity issuance, or delayed lease commencements can dilute the AI benefit.

For utilities and power producers, the key metric is who pays. Data centers can be attractive load if contracts are long-term, creditworthy, and structured so residential customers do not subsidize upgrades. They can become political liabilities if local power bills rise or if projects consume water and land without clear community benefits.

This is where the public debate can hurt the stocks. Data-center construction is moving into rural areas, and local pushback is increasing. A power story becomes a political story when residents see higher bills, water demand, or land-use fights before they see jobs and tax revenue.

That is why the best version of this trade is not “buy every power stock.” It is more selective: own the companies with order visibility, critical products, disciplined capital allocation, and pricing power tied to real bottlenecks.

What makes this story different today

The market has already covered “AI needs more data centers.” That is not enough anymore.

The fresh story is that the data-center boom is creating a new hierarchy inside the AI trade. GPUs remain the heart. But power availability is becoming the oxygen. Without power, GPU clusters are inventory. Without cooling, dense racks throttle. Without transformers and switchgear, campuses cannot energize. Without grid construction, signed cloud contracts become delayed revenue.

That is why a stock screen built only around “AI software” misses the real infrastructure winners. The better basket has four layers:

1. Cloud monetizers: Amazon, Alphabet, Microsoft, Oracle.
2. Equipment suppliers: Vertiv, Eaton, GE Vernova, Schneider Electric, Siemens Energy.
3. Grid builders: Quanta Services and related transmission contractors.
4. Capacity landlords: Equinix, Digital Realty, and select private data-center platforms.

The private market matters here too. OpenAI, Anthropic, xAI, sovereign AI projects, and enterprise AI deployments all add demand that may not appear cleanly in public-company revenue until it flows through GPU orders, cloud commitments, colocation leases, or power-equipment backlog. That is why the public stocks can react before the end customer has a stable AI profit model.

The DG Matrix story shows the same theme at the component level: once the data-center problem becomes a power-density problem, even smaller power-system companies become part of the AI value chain.

The risks investors should not ignore

The first risk is overbuilding. The IEA’s 2026 update says data-center investments have grown too large to be funded only from company balance sheets and will require capital markets. If AI monetization slows, financing costs rise, or hyperscalers cut guidance, capacity plans can be delayed.

The second risk is local resistance. Pew’s data shows the U.S. development pipeline is moving toward rural counties, while Axios reported delays tied to power constraints, supply chains, and community opposition. The easier urban sites are not where the next wave is necessarily going.

The third risk is supply-chain concentration. The IEA says high-bandwidth memory shortages are expected to persist through at least the end of 2027, and AI data centers require specialized power electronics, transformers, batteries, and cooling systems. If one link is constrained, the whole project can slip.

The fourth risk is valuation. Many AI infrastructure stocks already price in years of demand. A company can beat earnings and still fall if backlog growth, margin, or guidance fails to match expectations. The data-center theme is real, but a real theme can still be overpriced.

The fifth risk is policy backlash. Data centers can improve tax bases and create infrastructure investment, but they can also trigger electricity-rate debates. The IEA explicitly warns that data centers can create affordability challenges in systems that need new generation and grid investment.

Investor takeaway

Yes, AI data centers are becoming a power stock story. But the winning angle is not “electricity demand goes up.” The winning angle is scarcity.

Scarce grid connections create value for owners of powered land and transmission expertise. Scarce cooling systems create value for Vertiv-style suppliers. Scarce transformers and switchgear create value for Eaton and GE Vernova. Scarce AI capacity creates value for AWS, Google Cloud, Azure, Oracle Cloud, Equinix, and Digital Realty. Scarce project execution creates value for Quanta.

The best stocks in this theme are not just exposed to data centers. They are positioned where the bottleneck becomes revenue.

That is the difference between a headline trade and an investable trade. AI data creation may be measured in zettabytes, but stock-market returns will be measured in margins, backlog conversion, free cash flow, and the ability to secure power before competitors do.