MSME & Global Commerce · established evidence

The Server Farm That Drinks a City Dry

Last reviewed 2026-08-11. Written by Chandranshu Kumar, Founder, Raveneye Global. · 11 min read

Every dominant medium of an age has needed a physical plant to move it, and whoever owned that plant shaped the economy and the balance of power that followed, from the telegraph cable to the fibre backbone. The AI answer is the current medium, and its plant is the data center: a warehouse of processors, chillers, and a substation that behaves less like an office building and more like an industrial facility. Big Tech's capital spending on that plant is projected to reach roughly $725 billion to $760 billion in 2026, up from about $413 billion in 2025, funneled almost entirely into GPU clusters, custom AI chips, and the cooling systems needed to run them. The bill shows up beyond the balance sheet too: in electricity draws large enough to rank data centers among the world's biggest national consumers, and in water draws that compete directly with the towns around them. Because a handful of American companies decide where that plant gets built, they are now, in effect, allocating a share of the world's power and water, and governments have started competing for the privilege of hosting it.

The plant behind every information age

Every dominant medium of exchange has needed a physical plant to move it, and the plant has usually decided who profited from the age and who was merely served by it. The nineteenth century's medium was the telegraph, and its plant was a strand of copper laid across ocean floors; the handful of states and companies that owned the cables read the world's trade and diplomatic traffic first, and built the pricing and treaty advantages that followed from reading it first. The mainframe era of the mid twentieth century repeated the pattern at a smaller scale: a company or government that could afford a computing center, typically a single room of specialized machines, could out-calculate rivals who could not, and control of that scarce computing capacity shaped which firms and which state programs moved first.

The twenty-first century's dominant medium is the AI answer, generated on demand and delivered as a sentence rather than a link. Its plant is less romantic than a cable ship: a warehouse of racked processors, backup generators, chillers, and a substation, usually built somewhere unglamorous with cheap land and a spare grid connection. What has changed since the cable era is the size of the plant relative to the economy sitting next to it. A single AI data center campus can now draw as much power as a small city and consume water on an industrial scale, and the decision of where to put one has become an economic and diplomatic question in its own right.

That decision, who gets the compute, the power, and the water, and who gets asked to make room, is now made almost entirely by a handful of American companies. What follows is what the buildout actually costs, what it actually draws from the ground and the grid, and why governments have started treating a data center the way they once treated a port or a rail line.

This is the pattern this series traces throughout: whoever controls the physical plant of the dominant medium of an age shapes both its economy and its balance of power. The data center is that plant for the AI age, and its bill is now large enough to change how governments plan a grid, not just how a company plans a budget.

The six hundred billion dollar bet

Big Tech's spending on this plant has become one of the largest peacetime capital programs on record. Microsoft, Alphabet, Meta, and Amazon are together projected to spend roughly $725 billion to $760 billion on capital expenditure in 2026, up from about $413 billion in 2025, according to hyperscaler capex tracking reported by Statista and Tom's Hardware. That is a rise of nearly 80 percent in a single year, and the 2026 figure alone is larger than the annual economic output of most of the world's countries.

The guidance behind that total is public, company by company. Amazon has guided to roughly $200 billion in 2026 capital spending, Google to $175 billion to $185 billion, Meta to $115 billion to $135 billion, and Microsoft to roughly $190 billion, per the same industry tracking. Add Oracle, whose AI infrastructure commitments have grown fast enough that analysts now group it with the four hyperscalers as a "Big Five," and the combined 2026 figure reaches roughly $660 billion to $690 billion, close to double what the group spent in 2025, according to capex analysis published by Introl.

Set against national accounts, the number is a useful check on how large this program has become: the roughly $725 billion to $760 billion four companies plan to spend in 2026 would, on its own, rank among the thirty or so largest national economies by nominal GDP. None of that spending buys a product a household can hold; it buys chips, buildings, and the cooling systems to keep the chips running, which says something about where the money in this economy is now flowing.

Not all of it builds AI capacity specifically; hyperscalers also run conventional cloud and consumer businesses. But a large majority does. Roughly 75 percent of projected 2026 hyperscaler capex, on the order of $450 billion, is earmarked specifically for AI infrastructure: GPU clusters, the custom AI chips each company now designs in house (Google's TPUs, Amazon's Trainium, Meta's MTIA, Microsoft's Maia), and the buildings, power systems, and cooling plant needed to run them, according to a breakdown published by the Futurum Group. That is the clearest evidence that data centers have stopped being a background line item and become the thing Big Tech is now built around.

The power bill

The most direct way to see the scale of the buildout is the electricity meter. Data centers worldwide consumed roughly 448 to 490 terawatt-hours of electricity in 2025, according to industry and research analysis on data center energy demand summarized by the Brookings Institution. Treated as a single country, that consumption would place data centers around the eleventh largest electricity consumer in the world, ahead of nations with populations in the tens of millions.

AI workloads are not the whole of that figure, but they are a large and fast-growing share of it. AI accounted for roughly 15 to 20 percent of total data center electricity demand by the end of 2024, per research on AI-specific data center energy consumption. If AI's share of data center power keeps rising, toward an estimated 40 percent by 2030, the same research projects AI alone could then account for around 3 percent of total world electricity consumption, a level that would put a single computing workload on a comparable footing with the electricity use of a mid-sized industrial country.

What the raw terawatt-hour figures do not capture is concentration. A conventional data center draws power steadily and predictably. A gigawatt-scale AI campus, the kind now under construction across the US and increasingly abroad, can rival the peak electricity demand of a mid-sized city inside a single facility, and it wants that power now, not after a decade of grid planning. Utilities have responded by moving data center interconnection requests to the front of years-long queues, striking direct supply deals with hyperscalers, and in some regions delaying the retirement of older power plants to keep up. The cost of new capacity, and the risk of shortfalls, increasingly lands on the same grid that also has to keep the lights on for a city's households.

Some hyperscalers have responded to the power constraint by trying to secure their own generation rather than wait on a utility's build-out timeline, from long-term nuclear power-purchase agreements to on-site gas turbines. The approach treats electricity supply the way a company might once have treated a raw-material contract: not a service to be metered, but an input to be secured directly.

The water bill

Electricity is not the only resource the plant depends on. Most large AI data centers still rely on evaporative cooling to keep racks below the temperature at which processors throttle or fail, which means water, not just power, has become a hard constraint on where a facility can be built. US data centers expanded for AI workloads were consuming nearly one trillion liters of water annually by 2025, with a typical 100 megawatt AI facility using roughly 1.5 million to 3.0 million cubic meters of water a year for cooling alone, according to water consumption industry analysis.

Extended globally and forward in time, the figure grows further. A UN-affiliated report, summarized by Earth.org, projected that data centers worldwide could consume up to 9.3 trillion liters of water by 2030, with AI-driven cooling demand identified as a substantial part of that rise. The projection is a modeled scenario rather than a measured outcome, and should be read as such, but the direction it describes, sharply rising water demand concentrated in facilities that did not exist a decade ago, is not in dispute.

The strain is local before it is global. A data center's water draw competes directly with the municipal supply and the agricultural use of whatever county or district it sits in, and several of the regions hyperscalers have favored for cheap land and power, including parts of the US Southwest and other drought-prone areas, are places where that water is already contested. A facility that is invisible from the road can still be the single largest industrial water user in its county, which is why water rights, not just electricity tariffs, have become a line item in the negotiations that bring a data center to a given town.

The industry's response has not been static. Newer facilities increasingly favor closed-loop or air-based cooling systems that draw far less water than the evaporative designs common a few years ago, and several hyperscalers have committed publicly to becoming water positive, replenishing more than they draw, in the watersheds where they operate. Whether those commitments keep pace with the scale of new construction is a separate question from whether the intent is genuine, and it is the kind of claim this series treats as a scenario to watch rather than a settled result.

Siting as the new industrial policy

Because a data center needs cheap power, cheap water, and a low-latency route into the fibre network all at once, it cannot go just anywhere, and the sites that meet all three conditions are relatively few. Researchers studying the pattern have found that the concentrated siting of AI data centers is itself a source of regional power-system stress, since new gigawatt-scale campuses can rival the peak demand of a mid-sized city inside a single interconnection point, according to a 2025-2026 analysis published on arXiv. Ireland's grid operator has at points restricted new data center connections around Dublin specifically because AI-scale facilities were absorbing a growing share of national electricity growth, an early example of the tension this creates.

That concentration hands an unusual amount of allocation power to a small number of companies and the governments hosting them. The siting decisions of four or five American hyperscalers, plus Oracle, now determine, in practice, how electricity generation capacity, transmission headroom, and fresh water get allocated across multiple countries, decided inside corporate real estate and infrastructure teams rather than through any public planning process. A grid operator in Ireland, Malaysia, or the American Southeast now has to plan years of capacity expansion around whether one company decides to build there, a degree of private control over public infrastructure that previous generations of industrial siting, a steel mill, a refinery, a car plant, rarely produced at this scale or speed.

Governments have responded by treating grid capacity and water rights as an industrial-policy lever in their own right, the way an earlier generation of states competed on port access or rail lines to win manufacturing plants. Countries and states now offer dedicated power allocations, streamlined water permits, and tax terms built specifically around data center investment, competing openly for a facility the way they once competed for a factory, because the facility increasingly is the factory of the AI economy, and losing the bid can mean losing both the jobs and the claim to be a serious host of the compute economy.

The same small set of companies building this plant also operate the AI answer engines that now decide which businesses and institutions get named when someone asks a question. The infrastructure question, who gets the power and the water to run the compute, and the visibility question, who gets named in the answer that compute produces, sit inside the same handful of corporate roofs. A government negotiating grid capacity with a hyperscaler and a business trying to be read by that hyperscaler's answer engine are, in a real sense, dealing with the same landlord.

What every dominant medium has done twice

Every medium this series has examined liberated something. The telegraph collapsed weeks of transit time between a decision and its execution. The printing press put text within reach of anyone who could read. Large-scale AI compute is doing a version of the same thing: it puts research, drafting, translation, and analysis capacity that once required a large institution within reach of a single founder with a laptop and a subscription, and it is why small businesses and researchers with no data center of their own can now run tools that, a decade ago, only a well-funded lab could run.

Every medium in this series also concentrated something, usually the same infrastructure that did the liberating. The telegraph's liberation of information ran through cables that a handful of powers owned and could read before anyone else. The AI buildout's concentration runs through the physical plant itself: a small number of companies, sited in a small number of places, controlling the power, the water, and therefore the pace at which the rest of the world gets to use the technology. A founder anywhere can now rent a slice of a GPU cluster; whether that cluster exists at all, and how fast it grows, is decided by decisions made in a handful of corporate campuses and a handful of national capitals.

For the towns and countries that win a facility, the trade is not obviously bad. A single campus can bring long-term tax revenue, construction jobs, and a credible claim to being a serious node in the AI economy, which is exactly why local and national governments compete for them rather than resist them by default. The tension is not that hosting a data center is a loss; it is that the terms of the trade, how much power, how much water, and for how long, are negotiated privately between a hyperscaler and a jurisdiction, with the broader public that shares the same grid and the same aquifer rarely at the table.

Which side of that ledger dominates by 2030 is not yet settled, and the numbers in this article should be read as the range of plausible outcomes they are, not as a forecast. What is already measurable is the size of the bet: a spending program approaching a trillion dollars a year, a resource draw large enough to show up in a country's own electricity and water statistics, and a siting decision that has quietly become as consequential to a region's economy as the arrival of a rail line once was.

The evidence

Key findings, with their sources

Global data centers consumed roughly 448 to 490 terawatt-hours of electricity in 2025, which would make them about the world's eleventh largest electricity consumer if counted as a single country.

established Industry and research data center energy analysis, summarized by the Brookings Institution (2025-2026).
AI workloads accounted for roughly 15 to 20 percent of total data center electricity demand by the end of 2024, with projections suggesting AI's share of world electricity could reach about 3 percent if its share of data center power rises to 40 percent by 2030.

emerging Industry and research AI data center energy analysis (2024-2025).
US data centers expanded for AI workloads were consuming nearly one trillion liters of water annually by 2025, with a typical 100 megawatt AI facility using roughly 1.5 million to 3.0 million cubic meters of water a year for evaporative cooling.

established Water consumption industry analysis (2025-2026).
A UN-affiliated report projected data centers could consume up to 9.3 trillion liters of water globally by 2030, driven substantially by AI-related cooling demand.

emerging UN-report-based analysis, reported by Earth.org (2025-2026).
Combined 2026 capital expenditure by Microsoft, Alphabet, Meta, and Amazon is projected to reach roughly $725 billion to $760 billion, up from about $413 billion in 2025, with individual 2026 guidance of roughly Amazon $200 billion, Google $175 billion to $185 billion, Meta $115 billion to $135 billion, and Microsoft roughly $190 billion.

established Statista and Tom's Hardware hyperscaler capex analysis (late 2025-2026).
When Oracle is included among the largest US cloud and AI infrastructure providers, the resulting "Big Five" collectively committed roughly $660 billion to $690 billion in 2026 capital expenditure, close to double 2025 levels.

established Industry hyperscaler capex analysis published by Introl (late 2025-2026).
Roughly 75 percent, about $450 billion, of projected 2026 hyperscaler capex is targeted specifically at AI infrastructure: GPU clusters, custom silicon, and the data center buildings and cooling systems to house them.

established Industry hyperscaler capex breakdown published by the Futurum Group (2025-2026).
The scale and grid-interconnection concentration of AI data centers has been identified by researchers as a source of regional power-system stress, since new gigawatt-scale AI campuses can rival the peak demand of a mid-sized city in a single facility.

emerging "Concentrated siting of AI data centers drives regional power-system stress under rising global compute demand," arXiv (2025-2026).

Calibration

What is proven, what is promising, what is unproven

Evidence tier	Tactics	What the evidence says
established	The overall scale of the buildout: global data center electricity consumption in the high 400 terawatt-hour range for 2025, and the multi-hundred-billion-dollar hyperscaler capital spending figures for 2026.	Corroborated across multiple industry analyses (Brookings, Statista, Introl, Futurum Group) that draw on published company guidance and independent energy tracking rather than a single modeled estimate.
emerging	The specific AI-attributable share of that consumption (15 to 20 percent of data center electricity, a projected path to 3 percent of world electricity by 2030), and the water figures (nearly one trillion liters a year in the US, up to 9.3 trillion liters globally by 2030).	Each figure rests on a single research report or a modeled forward projection rather than a settled, multi-source census, and should be read as a plausible range, not a fixed count.
contested	Whether the current pace of data center construction is a durable, structural claim on grid and water resources, or a buildout that will overshoot realized AI demand and leave stranded capacity behind.	The spending and consumption figures document the scale of the current commitment; they do not, on their own, settle whether that commitment is matched by lasting demand for the compute it buys.

Reference

Glossary

Hyperscaler: A company, such as Microsoft, Amazon, Google, or Meta, that operates cloud and AI computing infrastructure at a scale large enough to require dedicated power and water planning on the order of a national utility.
Capital expenditure (capex): Money a company spends building or acquiring long-lived physical assets, in this case the chips, buildings, and cooling systems of an AI data center, as distinct from its day-to-day operating costs.
Evaporative cooling: A cooling method that uses the evaporation of water to remove heat from data center equipment, the primary reason large AI facilities draw significant volumes of water.
Grid interconnection: The process by which a new electricity user, such as a data center, is connected to the power grid, including the engineering review and capacity planning that determines how much power it can draw and when.
Custom AI silicon: Computer chips a company designs specifically for its own AI workloads rather than buying off the shelf, such as Google's TPUs, Amazon's Trainium, Meta's MTIA, or Microsoft's Maia.

Straight answers

Frequently asked questions

How much electricity do data centers use worldwide?

Global data centers consumed roughly 448 to 490 terawatt-hours of electricity in 2025, according to industry and research analysis summarized by the Brookings Institution. Treated as a single country, that would rank data centers around the eleventh largest electricity consumer in the world. AI workloads made up roughly 15 to 20 percent of total data center electricity demand by the end of 2024, and that share is rising.

Why do AI data centers need so much water?

Most large facilities still rely on evaporative cooling to keep processors from overheating, which draws water directly rather than just electricity. US data centers expanded for AI were consuming nearly one trillion liters of water a year by 2025, and a typical 100 megawatt AI facility uses roughly 1.5 million to 3.0 million cubic meters annually. A UN-affiliated projection put global data center water use as high as 9.3 trillion liters by 2030, though that figure is a modeled scenario, not a measured result.

How much is Big Tech spending on AI infrastructure in 2026?

Microsoft, Alphabet, Meta, and Amazon are together projected to spend roughly $725 billion to $760 billion on capital expenditure in 2026, up from about $413 billion in 2025. Add Oracle and the combined figure for what analysts now call the Big Five reaches roughly $660 billion to $690 billion, close to double 2025 levels. Around 75 percent of that spending, on the order of $450 billion, is earmarked specifically for AI infrastructure.

Why does the location of a data center matter geopolitically?

A data center needs cheap power, cheap water, and a fast network route all at once, which narrows the list of viable sites sharply. That concentration hands the siting decisions of a handful of American companies real influence over electricity and water allocation in the countries that host them, and governments have started competing on grid capacity and water rights as an industrial-policy lever, the way they once competed over ports or rail access.

Is the AI data center buildout sustainable?

That is contested, and this article treats it as an open scenario rather than a settled fact. AI compute is putting research and analysis capacity within reach of far more people than before, which is a real form of access. At the same time, the physical plant behind it concentrates control over power, water, and siting in a small number of companies and places, and whether efficiency gains and new supply keep pace with the spending shown here is not yet resolved by the numbers.

Provenance

Sources

Brookings Institution, analysis of global data center energy demand within the AI regulatory landscape (2025-2026) (established)brookings.edu
Presenc.ai, research on AI data center energy consumption and projections to 2030 (2024-2025) (emerging)presenc.ai
WaterToday Magazine, analysis of AI data center water consumption and cooling architecture (2025-2026) (established)magazine.watertoday.org
Earth.org, reporting on a UN-affiliated projection of global data center water consumption by 2030 (2025-2026) (emerging)earth.org
Statista, hyperscaler capital expenditure chart for Meta, Alphabet, Amazon, and Microsoft (late 2025-2026) (established)statista.com
Introl, analysis of 2026 hyperscaler capex and AI infrastructure debt (January 2026) (established)introl.com
Futurum Group, breakdown of the 2026 AI infrastructure capex sprint (2025-2026) (established)futurumgroup.com
arXiv, "Concentrated siting of AI data centers drives regional power-system stress under rising global compute demand" (2025-2026) (emerging)arxiv.org

Every figure above is attributed to a real, dated source and tagged with its evidence tier. Where a claim could not be verified to a primary source, it is not stated as fact.