Electrical equipment is less than ten percent of what a data center costs to build, and roughly one hundred percent of the reason it does not get built.

That sentence is the whole lecture. The capital is available, the chips are shipping, the demand is contracted, and somewhere between a third and a half of the capacity announced for 2026 is not under construction, because a transformer ordered today may arrive in 2029 and there is nothing anybody can pay to change that.

Sightline Climate tracked twelve gigawatts of United States data center capacity announced across roughly a hundred and forty projects for 2026. About five gigawatts was actually under construction. Eleven gigawatts sat in the announced stage with no physical progress, against typical build times of twelve to eighteen months, and a quarter of those projects had disclosed no power strategy at all, which is a remarkable disclosure gap for a facility whose defining characteristic is electricity consumption. The semiconductor supply chain feeding those sites has been scrutinized far more closely than the electrical one, and the electrical one is what is binding.

Data center bottlenecks are therefore not a story about scarcity in the ordinary economic sense, where a price rise summons supply. They are a story about a synchronized demand surge meeting a manufacturing base that was sized, deliberately and rationally, for twenty years of flat load growth.

The distinction matters because it determines what relief looks like. A shortage caused by underinvestment resolves when capital arrives, and capital has arrived in quantities without modern precedent. A shortage caused by physical lead times resolves when the lead times elapse, and no amount of capital shortens the certification cycle for a transformer production line or the apprenticeship of an electrician. Data center bottlenecks are the second kind, which is why the most-funded projects in the world are slipping on the same schedule as everybody else.

What actually gates a build, and the order it binds in

The intuitive ordering is wrong, and correcting it is the useful part.

Capital is not the constraint. Alphabet, Amazon, Meta, and Microsoft collectively guided to more than $650 billion of AI infrastructure capital expenditure for 2026, and OpenAI’s Stargate carries a SoftBank loan commitment measured in tens of billions. Money is abundant and it does not build transformers faster.

Chips stopped being the constraint. GPU supply eased after 2023 primarily because TSMC expanded advanced packaging capacity, which had been the true bottleneck behind the shortages rather than wafer fabrication itself, doubling output in 2025 with a further doubling planned. Demand still outruns supply by an estimated factor of roughly one and a half through 2027, and hyperscalers have locked up much of the available allocation through multi-year commitments, which squeezes smaller providers and is a structural disadvantage for anybody without a decade-long vendor relationship. The minor metals, specialty gases, and substrate materials behind advanced packaging remain their own constraint, and hafnium and the other exotic inputs to the gate stack sit further upstream again. But a buyer with a signed power agreement and a finished data hall can now get accelerators on a materially shorter timeline than a grid connection.

What gates the build is the electrical chain: large power transformers, generator step-up units, medium-voltage switchgear, circuit breakers, protection equipment, and the substation work to install them. Under those sits the equipment to generate the power in the first place, and under that sits the labor to install and commission all of it.

Every one of those has a lead time now measured in years, and none of them responds to urgency in a way that helps a project scheduled for next year.

Worth naming the ordering explicitly, because it inverts the way these projects get discussed. Land is available. Capital is available. Chips are constrained but obtainable. Power contracts are negotiable. Interconnection is slow and at least has a queue position you can hold. Equipment is the wall, and labor is the wall behind the wall. A project plan that treats the accelerators as the long pole is a plan written in 2023.

Transformers, and the material underneath them

Large power transformers are the tightest constraint and the most instructive, because the shortage traces cleanly to a specific input.

Delivery windows for the largest units ran roughly twenty-four to thirty months before the surge. They now run three to five years, with average United States lead times near a hundred and twenty-eight weeks and the most constrained categories quoted into 2029 or 2030. Switchgear is effectively sold out through 2028 at some suppliers. Medium-voltage equipment and generator step-up units carry similar multi-year queues.

The manufacturing constraint underneath is grain-oriented electrical steel, the specialized silicon steel used in transformer cores. It is produced through a complex, capital-intensive process by a small number of global producers, and without it a transformer manufacturer cannot expand output regardless of order book. That upstream constraint is why the shortage cascades identically across large power transformers, generator step-up units, and distribution transformers, and why different sectors of the economy end up bidding against each other for the same steel.

The geographic concentration compounds it. China controls roughly sixty percent of global transformer production capacity, which places a critical input for American AI infrastructure inside a supply relationship that the export control regime has already disrupted in adjacent categories. Copper windings add a second commodity exposure, and the minor metals and specialty alloys in the rest of the electrical chain add several more, including the rare earth magnets in every motor and generator the buildout consumes.

Domestic expansion is proceeding and it is slow for a structural reason. Building transformer manufacturing capacity requires new production lines, certification, and a trained workforce, which is a multi-year industrial investment cycle. The lead time to fix a lead time problem is itself a lead time.

The physical characteristics explain why nothing about this compresses. A large power transformer weighs hundreds of tons, is built to a specification unique to its installation, requires precision core stacking and winding, and then has to pass a testing and commissioning cycle that cannot be shortened without accepting failure risk on a component whose failure takes a substation offline. It then has to be transported, which for the largest units means specialized rail cars or heavy-haul road moves that themselves require permits and route surveys. There is no version of this that arrives quickly.

Distribution transformers face a parallel squeeze from a different direction, since residential construction, electric vehicle charging, and grid replacement of aging stock all draw on the same production. A utility replacing storm-damaged equipment competes against a data center for the same factory slot, which is how a data center bottleneck becomes a visible local political grievance rather than an industry problem.

Turbines, and the backlog that reprices generation

If a project cannot get a grid connection, the workaround is self-generation, and that workaround has run into the same wall.

GE Vernova’s gas turbine backlog reached one hundred gigawatts in the first quarter of 2026, with the company expecting at least a hundred and ten gigawatts of orders and reserved slots, roughly three years of lead time, and about ten gigawatts of remaining availability across 2029 and 2030 combined. Wait times for the largest units have stretched as far as seven years, with the average around five, against a historical backlog of one to three.

The pricing response tells the story better than the queue does. Gas turbine prices are up roughly seventy-five percent for plants going online in 2030 or 2031. The estimated cost of a new gas-fired plant has more than tripled, from around eight hundred dollars per kilowatt in 2021 to something in the range of twenty-six hundred to twenty-eight hundred today.

That repricing has a consequence the discussion usually misses. Self-generation was the fast, cheap answer to the interconnection queue. It is now the slow, expensive answer to the interconnection queue, which narrows the gap between building your own power and waiting for the grid, and makes existing generation assets correspondingly more valuable.

The turbine manufacturers could reduce the backlog by adding capacity, and their reluctance is rational rather than obstructive. Between 2005 and 2020 United States power consumption grew by roughly a tenth of a percent annually. A manufacturer that builds factories for a demand curve that reverts to that baseline owns stranded plants. They are being asked to make a twenty-year bet on a four-year-old trend, by customers who cannot themselves agree on whether the trend holds.

That reluctance is the same calculation every supplier in this chain is making, and it is the reason the shortage persists against enormous demonstrated demand. A transformer manufacturer, a switchgear producer, a grain-oriented steel mill, and an electrical contractor all face the identical question: is this a permanent step change in load growth or a four-year surge that reverts. Each of them loses badly by guessing wrong in the expansionary direction, and each of them merely forgoes profit by guessing wrong in the conservative direction. Asymmetric downside produces conservative behaviour, which is exactly what the processing industries that consolidated through a long demand trough did for the same reason.

The labor constraint, which nobody can import

The bottleneck that gets least coverage and may matter most is that there are not enough electricians.

Microsoft’s president has publicly identified the shortage of electrical talent as the single largest obstacle to the company’s United States data center expansion, describing flying electricians in from seventy-five miles away or relocating them temporarily to keep projects moving. Oracle pushed completion timelines from 2027 to 2028 with labor shortages cited as a contributing factor.

The arithmetic is demographic and does not resolve quickly. Estimates put the additional electricians needed above three hundred thousand while roughly twenty thousand retire annually, and one commonly cited projection has one new entrant to the trade for every five retirements. Associated Builders and Contractors estimated construction needed roughly three hundred forty-nine thousand net new workers in 2026 alone, with other analyses putting the figure above four hundred thousand, and research on supply chain constraints curbing US development projecting a requirement rising toward 1.4 million by 2030.

The specificity is what makes it hard. A data center campus does not need generic construction labor. It needs licensed electricians with mission-critical experience, mechanical contractors familiar with high-density liquid cooling, controls engineers who can integrate power and thermal telemetry, commissioning agents who understand phased energization, and high-voltage substation crews. Construction costs reflect the scarcity, having risen from roughly $7.7 million per megawatt in 2020 to about $10.7 million in 2025, with further increases forecast, and a meaningful share of that escalation is labor rather than materials. Those are not interchangeable pools, and a market can post strong construction employment growth while lacking the specific teams a three-hundred-megawatt campus requires.

Unlike a transformer, this cannot be ordered early. An apprenticeship runs years, the instructors are drawn from the same shortage, and the wage premium the sector can pay pulls workers from residential and commercial construction rather than creating new ones. Data center projects with deeper budgets and harder deadlines are winning that competition, which is a redistribution rather than a solution.

The consolidation response has begun as well, with acquisitions in specialized electrical services aimed at securing crews rather than capabilities. When a firm buys another firm principally for its labor pool, the labor pool has become the scarce asset, and that is a reasonably reliable indicator of where a bottleneck actually sits.

Why the shortage was rational to create

The temptation is to read a five-year transformer queue as a planning failure, and the more accurate reading is that it is the correct response to twenty years of correct information.

United States electricity demand was essentially flat from roughly 2005 to 2020, growing around a tenth of a percent per year while efficiency gains offset economic growth. Every rational actor in the electrical supply chain sized capacity to that reality. Utilities bought just in time. Manufacturers ran lean. Nobody built speculative transformer capacity, because speculative transformer capacity would have been a decade of losses.

Then demand grew two percent in 2024 alone, and PJM’s load forecast attributed ninety-four percent of its projected thirty-two gigawatts of peak growth through 2030 to data centers, with data center growth outpacing new generation additions by roughly two to one. Dominion Energy in Virginia, the densest data center market in the world, has quoted waits beyond thirty-six months for new substation service and as long as seven years for a hundred-megawatt connection in its most constrained territory.

A supply chain optimized for flat demand cannot absorb that, and the optimization was not a mistake at the time it was made. This is the standard shape of an industrial bottleneck, and it recurs identically in critical minerals and in every processing industry that consolidated during a long demand trough: capacity gets rationalized during the lean period by firms behaving sensibly, and the rationalization is invisible until demand returns. The critical minerals sector spent two decades learning the same lesson and is still learning it.

The double-ordering problem

There is a measurement issue sitting underneath every number in this subject, and it is worth flagging because it cuts against the alarming version of the story.

Buyers facing multi-year lead times book capacity with several suppliers at once to guarantee delivery, then cancel the surplus. Utilities have begun placing speculative orders and paying to reserve factory slots before projects are approved, reversing decades of just-in-time procurement. Both behaviours inflate reported backlogs.

Which means the true depth of the shortage is genuinely hard to read. A hundred-gigawatt turbine backlog across ninety distinct customers in twenty-four countries is a real order book and it is not a reliable forecast of installed capacity, because some fraction of it is optionality rather than intent.

The same caution applies to the announced-versus-under-construction gap. Some of the eleven gigawatts sitting in announcement is genuinely blocked by equipment. Some is blocked by permits, some by financing, some by a customer contract that never closed, and some was never a real project. Announcements are cheap and they are a document with an author who wanted something.

The incentive to announce is real and worth stating. A gigawatt announcement supports a share price, secures a place in a state’s economic development pipeline, establishes a claim on utility attention, and costs nothing but a press release. There is no penalty for an announcement that quietly does not happen, and no register of the ones that did not. Any analysis built on announced capacity is therefore measuring intent plus marketing, and the same problem afflicts every forecast in a capital-intensive industry where announcing a project is orders of magnitude cheaper than building one.

Speed to power, and what it is worth

The scarcity has produced a repricing that is legible in real estate and worth naming, because it explains a great deal of transaction behaviour.

An asset with executed interconnection agreements, an energized substation, and documented electrical headroom commands a premium over an otherwise identical site without them. Proximity to substation corridors matters. Sites near transformer and switchgear supply nodes commission faster.

That premium is the market pricing time rather than capacity, and it explains why operators are buying powered land, retired industrial sites, and closed generating stations at valuations that make no sense on the real estate alone. What is being purchased is queue position, which is the same thing being purchased when an operator pays above market for an existing generating asset with an interconnection already attached.

It also explains the vertical integration. Crusoe has begun manufacturing its own switchgear to bypass conventional lead times. Operators are buying generation outright rather than contracting for output. Each of those removes a supplier from the critical path at considerable capital cost, and the calculation is straightforward when a sixty-megawatt facility delayed a month forgoes revenue in the millions.

The same logic drove twentieth-century industrial firms to buy their input suppliers rather than negotiate with them, and it produces the same result: capital intensity rises, flexibility falls, and the firm ends up owning assets it has no particular expertise in operating. It also concentrates risk, since a company that owns its switchgear factory has converted a supplier problem into an operational one, and the vertical integration that looked prudent during a shortage frequently looks like a distraction once the shortage clears.

The data center bottleneck cascade, and where it breaks

The bottlenecks interact, and the interaction determines which projects die.

A project needs land, a permit, an interconnection position, a transformer, switchgear, generation or a grid contract, cooling infrastructure, accelerators, and crews to install everything. Failing any one of those stops the project, and the failures are not independent, because the same electricians install the switchgear and commission the cooling, and the same substation work gates both the interconnection and the equipment installation.

That produces a specific failure mode. A project can hold a signed interconnection agreement, a completed study, financing, and hardware allocation, and still not exist, because roughly fifty-seven gigawatts of PJM projects have completed the study process and hold agreements while remaining stalled. Completing the queue turned out to be necessary and nowhere near sufficient.

It also produces a sequencing problem for anyone underwriting. The thermal design determines the electrical requirement, which determines the transformer specification, which has a three-to-five year lead time, which means the cooling architecture chosen in year one locks equipment orders that arrive in year four for hardware generations that do not exist yet.

And the workarounds have their own ceilings. Bypassing the grid entirely applies to only a small share of the current pipeline, on the order of a few percent. Behind-the-meter arrangements face regulatory constraints. Self-generation faces the turbine backlog. Every escape route is itself congested.

The cooling chain deserves a mention in the same breath, since the transition to liquid put a new set of components on the critical path. Coolant distribution units, cold plates, high-flow pumps, and the specialized manifolds that connect them are a young supply base scaling from small volumes, and the thermal architecture decisions that made those components necessary happened faster than the vendors could industrialize. Backup power adds another, with battery systems and uninterruptible supplies drawing on cell manufacturing capacity that the vehicle industry is also bidding for.

What breaks the bottleneck, and when

The relief timeline is knowable and it is not soon, which is the honest answer.

Manufacturing expansion is underway. GE Vernova, Siemens Energy, Eaton, Hitachi Energy, and others have announced capacity additions, and manufacturers of adjacent products have begun pivoting lines toward stationary power equipment. The consistent horizon for that relief is 2029 and beyond, because certification, workforce ramp, and line commissioning cannot be compressed.

Grain-oriented electrical steel capacity is the deeper fix and the slower one, involving the kind of heavy industrial investment that requires a decade-long demand thesis to justify.

The labor pipeline is the slowest of all, and it improves only through apprenticeship completions that started years ago, which means the 2029 electrician workforce is already substantially determined.

Which leaves demand-side adjustment as the fastest lever, and it is happening whether anybody likes it or not. Projects slip. Announcements quietly lapse. Capacity migrates to jurisdictions with available equipment and available crews, which is a siting logic driven by industrial inputs rather than by markets and which has historically produced durable industrial geographies. Design standards relax where redundancy can be traded for speed, with some operators accepting lower redundancy tiers or phased energization to get partial capacity earning while the rest waits. And a share of announced capacity turns out never to have been real, which resolves the shortage by reducing the demand rather than by increasing the supply.

That last mechanism is the one most likely to clear the data center bottleneck first, and it is the one nobody markets. Shortages of long-lead industrial equipment historically resolve through demand destruction more often than through capacity addition, because demand can fall in a quarter and capacity takes half a decade. Whether that happens here depends on questions being decided somewhere else entirely, in model architecture and inference efficiency, and on whether the contracted backlogs underwriting the buildout convert.

The claims that do not hold up

An audit, because this subject produces headline numbers with unstated denominators.

A five-year transformer backlog killed half of America’s 2026 data centers is the viral version and it overstates a real problem. Delays are real, the announced-to-construction gap is real, and attributing the entire gap to transformers ignores permitting, financing, and projects that were never funded.

GPUs are the bottleneck was true in 2023 and is not now. Advanced packaging capacity expanded, allocation is contracted, and the pacing item moved to the electrical chain.

The shortage is a planning failure misreads twenty years of rational behaviour under flat demand. The capacity that is missing was correctly not built.

Backlogs measure demand overstates order books inflated by double-ordering and speculative slot reservation.

Money will solve it is contradicted by the most-capitalized projects in the world slipping schedules. Stargate has essentially unlimited capital and cannot buy a transformer that has not been manufactured. As of April 2026 industry observers reported no significant physical progress on its data center buildouts, which is a data point about the limits of capital rather than about the project.

The bottleneck is temporary depends entirely on the timescale. Two years is not temporary for a five-year asset financed on three-year debt, and a delay of that length can consume the entire useful life advantage the project was underwritten on.

Onshoring will fix the supply chain understates how long building a transformer industry takes and ignores that the steel underneath it is the actual constraint. The same argument recurs in every critical materials reshoring proposal, and the answer is always that the upstream step is the one nobody wants to finance.

What the data center bottleneck is actually telling us

The reframing worth carrying is that the AI buildout is not a technology project encountering technology limits. It is a heavy industrial project encountering heavy industrial limits, and almost nobody involved in the software layer has any experience with those.

The constraint set is familiar to anyone who has built a refinery, a smelter, or a transmission line: long-lead custom equipment, a specialized labor pool that takes years to grow, an upstream material with a concentrated supply base, and a permitting environment with a political dimension. Anyone who has built a mine, a smelter, or a processing facility recognizes every item on that list. None of it is novel. What is novel is the speed at which capital arrived against it, and the inexperience of the people deploying it with the physical constraints they are deploying it into.

Which produces the specific asymmetry that defines the period. Software scales by copying, at near-zero marginal cost, on timescales of weeks. Everything the software now requires scales by manufacturing, at high marginal cost, on timescales of years. The buildout is the collision between those two clocks, and the slower one wins every time it is tested.

The ten-lecture briefing on how AI data centers work runs the sequence in the order the constraints actually bind, because that ordering is the analysis: the thermal density set the electrical requirement, the electrical requirement met a supply chain sized for a different world, and what got built is the intersection rather than the ambition.

The most sophisticated computing hardware ever manufactured is waiting on a steel mill. That relationship is the demand wall, and every data center bottleneck in this subject is a version of it.

A transformer is a steel core wrapped in copper wire, a technology essentially unchanged in principle since the nineteenth century, weighing several hundred tons and requiring a particular kind of silicon steel that a handful of mills produce. It is the least glamorous object in this entire industry, and in 2026 it is the thing that decides which announcements become buildings.

A county commissioner who votes to approve a data center in 2026 is making a career decision, and increasingly the wrong one.

That is a new fact. For most of the last two decades, approving a large industrial project with a tax base attached was politically routine and frequently popular. Then in April 2026, voters in Port Washington, Wisconsin approved by sixty-six percent a measure requiring voter approval before data center tax incentives can be granted. Voters in Missouri removed city council members who had approved a facility. In Michigan, candidates from both parties who campaigned on data center restrictions won their primaries. Polling has found roughly seven in ten Americans opposing construction in their own area, nearly half strongly, against about a quarter in favor.

Data center politics has become a domestic electoral variable in under two years, and it is one of three separate forces pressing on the same buildout from different directions. The second is geopolitical: export controls, sovereign compute programs, and a supply chain concentrated in a place with a strategic problem. The third is technological, and it is the one the other two cannot see coming, because it consists of the possibility that the hardware being fought over turns out not to have been necessary in the quantity ordered.

Each vector operates on a different timescale, each has its own mechanism, and the interesting question is not which one is most dangerous. It is that a project has to survive all three, and they do not arrive in a convenient sequence.

Vector one: data center politics turned local

The reversal happened fast and the reason is legible in the mechanism rather than in anybody’s rhetoric.

A data center brings a large assessed value and a small workforce. A multi-billion-dollar campus may employ a few dozen to a few hundred people in steady state, which is a capital-to-labor ratio no traditional manufacturing plant produces and which sits at the opposite end of the spectrum from the extractive and processing industries that historically anchored rural tax bases, and which means the standard economic development pitch is doing less work than it appears to.

Against that, the visible costs arrive quickly. Capacity prices rise across a whole zone when a large load tightens supply, which puts the cost on bills belonging to people who never saw a permit application. Water withdrawals are local and measurable, and the evaporative cooling that drives them is most efficient in exactly the arid climates where water is most contested. Construction traffic, generator testing, and the low-frequency noise of a large cooling plant are experienced by neighbours rather than by the county at large. And the tax abatements that secured the project in the first place mean the assessed value is not producing the revenue the pitch implied for the first decade.

The legislative response has been rapid and it runs in both directions, which is what makes it interesting rather than a simple wave. New Jersey enacted a Data Center Fair Share Act requiring facilities over fifty megawatts to commit to paying at least eighty-five percent of projected power costs for a decade, one entry in a running tally of state moratoria, tax changes, and ballot measures that has become the closest thing to a scoreboard for this fight. Virginia retained its sales tax exemption while adding an energy consumption tax of roughly a cent per kilowatt-hour capped at six hundred million dollars annually. Arizona and Illinois paused data center tax incentives. South Dakota passed a law allowing local governments to regulate or ban rather than imposing a statewide prohibition. New Hampshire moved the opposite way with an off-grid electricity providers law easing self-supplied power.

The restriction efforts also fail regularly, which the coverage tends to underreport. Maine’s governor vetoed a state moratorium. Ohio’s proposed constitutional amendment banning facilities over twenty-five megawatts gathered roughly seventy thousand of the four hundred thirteen thousand signatures required and missed the 2026 ballot. Pennsylvania’s statewide hyperscale moratorium was rejected in committee two to nine on constitutional and local-authority grounds. Sangamon County, Illinois voted down a six-month moratorium and approved a five-hundred-million-dollar campus the same evening.

So the accurate description is not that communities are blocking data centers. It is that approval has stopped being free, that the terms are being renegotiated in public, and that the venue has moved from a county board to a ballot. That shift matters procedurally as much as substantively, because a county board can be lobbied, briefed, and negotiated with, and an electorate cannot.

The information asymmetry that makes it bitter

One structural feature turns an ordinary siting dispute into a trust problem, and it is worth isolating because it is fixable and mostly has not been fixed.

Project water and power figures are frequently confidential during negotiation, sometimes under non-disclosure agreements signed by local officials. A community is asked to approve a withdrawal and a load whose magnitude it cannot verify, by representatives who have seen the numbers and cannot repeat them.

Pennsylvania considered legislation prohibiting local elected officials from signing non-disclosure agreements with developers. It had bipartisan sponsorship, passed a Senate committee, and stalled on a tie in House Judiciary after opposition from business groups arguing that confidentiality is necessary for economic development competition.

Both positions are defensible on their own terms. A developer negotiating simultaneously with several jurisdictions has a real commercial interest in not publishing its requirements. A resident asked to accept an unnamed quantity of groundwater withdrawal has a real interest in knowing the number. There is no clever resolution, and the absence of one is why the same argument recurs in every county.

The long history of large extractive facilities arriving in rural jurisdictions supplies the pattern, and the pattern is that the terms get renegotiated after the community discovers what it agreed to, usually from a weaker position. The company towns built around a single employer with a resource requirement are the long-run version, and the resemblance is close enough that several county boards have started citing it explicitly.

Affordability, which is the vector’s real fuel

Data center politics attached itself to an existing grievance rather than creating a new one, and that is why it moves votes.

Electricity prices rose materially in several regions, and capacity market outcomes are the transmission mechanism. When an auction clears short of its reliability target, prices rise for every customer in the zone. Data centers contributed substantially to the demand that tightened those auctions. They were not the only contributor, since generation retirements, transmission costs, fuel prices, and weather all matter, and the attribution question is genuinely contested.

The scale of the political reaction has been documented across primary season, with reporting on how the issue reshaped 2026 races finding candidates in both parties winning on data center restrictions and at least one state authority official losing a seat he had previously held comfortably after backing a project.

What is not contested is the political arithmetic. A voter receiving a higher bill and reading about a facility that consumes as much electricity as a small city does not require a causal econometric analysis to form a view, and no candidate has an incentive to supply one. That is not a criticism of voters. It is a description of what happens when a diffuse cost meets a visible cause, and it is the same dynamic that has attached to every large industrial facility with a legible footprint.

Which produces a policy proposal that keeps surfacing across party lines: require large new loads to bring their own new generation. The Nevada governor’s race featured a candidate promising legislation to that effect. PJM’s market monitor has argued for it. The logic is sound, since a load that adds its own supply does not tighten the market and therefore does not raise anybody’s price.

The objection is equally sound. No other class of customer faces that obligation, thresholds are arbitrary, and an aluminium smelter or a large industrial extraction operation with comparable draw would either face the same rule or receive an exemption somebody has to justify. That is the shape of the argument and it is not going to resolve cleanly.

Vector two, and the export control regime

The geopolitical vector operates on a different clock and through a completely different instrument, which is a licence.

The regime has become more granular rather than simply tighter. In July 2026 the Bureau of Industry and Security moved the United Arab Emirates into Country Group A:5, the most favourable export tier, and created a supplement naming entities permitted to receive advanced computing items without a licence at all. Saudi Arabia is not in that group and has no equivalent listing. What it has instead is a November 2025 authorisation permitting its national AI champion, alongside the UAE’s G42, to purchase up to the equivalent of thirty-five thousand GB300-class accelerators each, subject to security requirements, reporting, and compliance monitoring.

A rule and a licence are different instruments. A rule is durable and predictable and can be relied on for capital planning. A licence is discretionary, revocable, and creates a permanent relationship with the issuing agency. Two states with similar ambitions and similar capital are operating under structurally different regimes, and that difference is a policy instrument rather than an accident.

Enforcement has changed character as well. The Department of Justice’s national security enforcement policy has moved export violations out of the quiet civil settlement category, and individual prosecutions have followed. Compliance, audit, and legal review are becoming recurring operating costs rather than transaction expenses.

The consequence for anyone building is that jurisdiction is now an architecture decision. Which chips can be deployed where, which models can run on them, and which data can cross which border are questions with multi-year lock-in and refactoring costs that can exceed the original deployment budget. That is the same material and regulatory fragmentation that reshaped critical minerals, arriving in a different industry with the same mechanics.

Sovereign compute, and the queue behind the queue

Every major economy has now announced a national compute program, and the collective effect is competition for the same scarce hardware.

The European Union unveiled a two-hundred-billion-euro AI Continent Action Plan with thirteen AI Factories established across seventeen member states. Saudi Arabia announced over fifteen billion dollars in new investments including a partnership with Google Cloud and plans to deploy five hundred megawatts each of AMD and NVIDIA capacity. The UAE is developing a twenty-six square kilometre campus in Abu Dhabi with five gigawatts of planned capacity, with Stargate UAE targeting an initial two hundred megawatts. India, Japan, and France have their own programs.

The physical problem is that export-cleared, high-bandwidth-memory-heavy accelerators are supply constrained, and every sovereign program is bidding for the same units. A licence that is theoretically approvable still queues behind allocation, which means the binding constraint on national AI ambition is frequently manufacturing capacity rather than policy. High-bandwidth memory is the specific chokepoint, produced by a small number of suppliers, and it sits alongside the specialty gases, substrates, and minor metals that gate advanced packaging regardless of how many licences get issued.

Underneath that sits the concentration nobody has solved. Advanced logic manufacturing remains overwhelmingly concentrated in Taiwan, fab construction elsewhere is proceeding slowly, and the semiconductor supply chain has a single point of geographic failure that a decade of policy has not meaningfully diversified. The materials underneath it are concentrated separately and add their own chokepoints, which is the structural reason chip policy keeps discovering new dependencies after it has addressed the previous one. Any serious disruption there stops the buildout everywhere, on a timescale measured in quarters rather than weeks.

There is also a physical security dimension that the sovereignty debate largely skipped until recently. Concentrating enormous compute capacity in a region with an active missile threat is a different risk profile from concentrating it in Iowa, and the February 2026 strikes in the Gulf made that concrete rather than theoretical.

Vector three, and the efficiency problem

The third vector is the one that does not announce itself, and it is the reason the first two might not matter.

The buildout is premised on compute demand scaling with capability. If capability improvements come increasingly from algorithmic efficiency, architectural changes, and inference optimization rather than from raw floating-point operations, the relationship between capability and capital expenditure loosens, and a great deal of hardware gets bought for a curve that bent.

The demonstration event was DeepSeek’s R1 release in January 2025, which showed competitive model performance achieved on constrained hardware, though several of the cost claims circulated at the time were misleading and the comparison was less clean than the coverage suggested. The more consequential follow-on was a model released in April 2026 serving queries on Huawei’s Ascend architecture, which established that a frontier-class model can be built around non-American silicon.

The strategic implication for the export control regime is uncomfortable and has been argued at length: controls restrict computational resources without preventing capable models, because laboratories adapt around hardware constraints through efficiency work rather than being stopped by them. That is a criticism of the instrument rather than of the objective, and it is contested, since restricting compute plausibly slows an adversary even if it does not stop one. The historical record on export restriction as an instrument is mixed in exactly this way: controls reliably impose cost and reliably accelerate domestic substitution in the restricted country, and which effect dominates depends on how long the restriction holds and how substitutable the input is.

The implication for the buildout is more direct. Every efficiency gain reduces the compute required per unit of capability. Whether that reduces total compute demand depends entirely on whether cheaper inference expands usage faster than it reduces cost per query, which is the standard rebound question, and the honest answer is that nobody knows. The investors underwriting five-year debt against hardware are taking a position on that question whether or not they have articulated it.

Where the three vectors of data center politics interact

The vectors are usually analysed separately and their interactions are what determine outcomes.

Politics and geopolitics reinforce each other in one specific way: national security framing is the strongest available argument against local opposition. A project characterised as strategic infrastructure in a competition with China is politically harder to block than a project characterised as a warehouse full of computers owned by a company with a market capitalisation larger than the state’s budget. Federal preemption arguments and expedited permitting proposals both run through that framing.

They also work against each other. Export restrictions and sovereign programs push capacity offshore, and capacity built in the Gulf or Europe is capacity not built in a county that would have fought it. Domestic opposition is, in that sense, an offshoring pressure, which is an outcome nobody in the domestic argument is optimising for.

Efficiency interacts with both. If the compute requirement per unit of capability falls, the political fight gets smaller because the load gets smaller, and the geopolitical fight gets harder because controls become less effective. A world with abundant cheap inference is a world with less local opposition and less leverage over adversaries, which is not the combination either camp is planning for. The efficiency-versus-demand-growth question has the same structure in every materials-intensive industry, and the historical answer has usually been that cheaper inputs expand consumption rather than reducing it.

And the timing mismatch is the structural problem. Local political cycles run two to four years. Export control regimes change with administrations. Model architecture shifts on a horizon of months. Data center capital commitments run fifteen to thirty years on the building, five to twenty on the power contract, and five or six on the silicon inside, assuming the depreciation schedule holds. Every clock in the political and technological environment is faster than the clock on the asset.

What the operators are doing about it

The industry response is legible and worth cataloguing without endorsement, because it is the practical content of this vector.

Community benefit agreements have become standard, with commitments on water recycling, noise mitigation, workforce training, and direct payments to school districts. Some are substantive and some are public relations, and the distinction is usually visible in whether the commitment is enforceable.

Rate structures are being negotiated that shift cost risk onto the facility. New Jersey’s minimum-take requirement and the special large-load tariffs several utilities have constructed are the visible examples, covering contract duration, exit fees, and collateral, and they amount to an attempt to make a data center behave like a creditworthy long-term industrial customer rather than a load that might leave.

Bring-your-own-generation is the technical response to both the affordability argument and the queue, and it is also the response that produces on-site gas turbines, which is a different local fight with a different set of objections. Air permits, emissions monitoring, and turbine noise replace capacity prices and interconnection studies, and the firm generation alternatives that would avoid combustion are either slower, smaller, or geographically limited.

Geographic diversification is the response to concentration, and it is why capacity is moving toward jurisdictions with available power, permissive permitting, and cool climates, which is a search for the same combination of inputs that has always determined where energy-intensive industry locates, which is the same logic that governs any siting decision where the physical inputs dominate, and which has historically produced industrial geographies that outlast the reason they were created.

Vertical integration is the strategy nobody calls one. An operator that owns its generation, its water rights, and its interconnection has removed three negotiations from the schedule, which is expensive up front and is why the largest players are increasingly buying power assets outright rather than contracting for output. The same logic drove twentieth-century industrial firms to buy mines rather than sign supply agreements.

And silence is a strategy. Projects announced late, structured through shell entities, and negotiated under confidentiality reduce opposition until permits are issued, at a cost in trust that the next project in the same county pays.

What a project actually has to clear

Setting the three vectors against a single hypothetical project makes the compression concrete, and it explains why announced capacity and built capacity have diverged.

Land and zoning come first, and this is where the ballot measures bite. A conditional use permit, a rezoning, or an incentive package can now trigger a referendum in jurisdictions that allow citizen initiatives, and twenty-three states do. A project that would have been a routine county board item in 2020 can now become a ballot question, which adds an election cycle to the schedule regardless of the outcome.

Power comes second and is the longest pole. An interconnection position, a utility contract, and increasingly a special large-load tariff with minimum-take obligations and exit fees. Bring-your-own-generation shortens the schedule and substitutes an air permit fight for a rate case.

Water comes third and is jurisdiction-specific to a degree that surprises developers. A withdrawal permit in a prior-appropriation state is a different instrument from one in a riparian state, and in some places it means purchasing a senior right from an agricultural holder, which is legal, orderly, and locally understood as a data center buying a farm’s water. The aquifer drawdown question is separate again and turns on recharge rates rather than on permits.

Hardware comes fourth and is where the geopolitical vector applies, through export classification, end-user screening, and in some jurisdictions location verification requirements. The minor metals and specialty materials inside the equipment carry their own origin and content rules that most procurement teams discover late.

And the financing sits across all four, because a lender is underwriting the whole sequence rather than any single approval. A project with power and no permit is worth nothing, and so is a project with a permit and no interconnection date. The long record of infrastructure projects that failed at one stage of a multi-stage approval is the relevant base rate, and the base rate is not encouraging.

The claims that do not hold up

An audit, because this subject generates confident assertions from every direction.

Communities are blocking data centers overstates it. Approvals continue at scale, moratoria fail regularly, and the accurate description is that terms are being renegotiated rather than that projects are being stopped.

Data centers caused your electricity bill to rise is directionally supported and routinely overstated. They contributed to demand that tightened capacity markets, alongside retirements, transmission costs, fuel prices, and weather.

Export controls have failed is a stronger claim than the evidence supports. They demonstrably restricted computational resources available to constrained laboratories. Whether restricting resources achieves the strategic objective is a different question with a genuinely unresolved answer.

Export controls are working is equally overstated, for the mirror reason.

DeepSeek proved compute does not matter misreads the finding. It demonstrated that capability is not a pure function of compute, which is a much narrower and still significant claim.

Sovereign AI means technological independence overstates what any current program achieves, since essentially all of them remain dependent on a small number of American-designed accelerators and a single fabrication geography, and buying capacity is not the same as controlling a supply chain.

Data centers create jobs is true and small. Construction employment is substantial and temporary; operational employment is modest and permanent. The electrical and mechanical trades that build them are genuinely constrained, which is a separate labour story from the one in the economic development brochure.

Data center politics is simply NIMBYism dismisses the specific and checkable objections about water, capacity prices, and abatement terms by assigning a motive rather than answering the point.

The federal government will preempt local control assumes an outcome that has been proposed and not enacted, and that runs into state authority over siting and retail rates.

What the squeeze is actually telling us

Assemble the three vectors and what they share is that none of them is about data centers.

The political vector is about electricity affordability and about who captures the benefit of a tax abatement, which are old questions with a new object. The geopolitical vector is about whether technological leadership can be maintained through export restriction, which is a question with a long and mixed record predating semiconductors. The technological vector is about whether capability scales with capital, which is the oldest question in industrial investment.

Data center politics is where those three arguments happen to be colliding right now, which means the outcome will be determined substantially by how those arguments resolve rather than by anything specific to the facilities. Data center politics is a venue rather than a subject.

The practical consequence for anybody reading a project announcement is a checklist. Ask whether the load is bringing its own generation or drawing from a constrained market, because that determines the local political trajectory. Ask which jurisdiction the hardware sits in and under which instrument, because a rule and a licence carry different risk. Ask who signed a non-disclosure agreement and what it covers, because that predicts how the trust breaks. And ask what the capital commitment assumes about compute intensity per unit of capability, because that assumption is being tested continuously by people with no interest in whether the buildings get paid for.

The ten-lecture briefing on how AI data centers work runs the physics, the money, and the politics in sequence because they resolve in that order: the thermodynamics set the load, the load set the financing, and the financing arrives in a county that gets a vote.

A commissioner in Wisconsin, a licensing officer at Commerce, and a researcher publishing an efficiency result have nothing to do with each other and no knowledge of each other’s schedules. All three are, this year, in a position to determine whether a particular building gets built and whether it earns back what it cost.

The industry-average server rack draws about seven and a half kilowatts. An NVIDIA GB200 NVL72 rack draws one hundred and twenty to one hundred and forty. It weighs roughly 1.36 metric tons, it cannot be air cooled, and NVIDIA does not offer an air-cooled version, because there is not one to offer. The magnets, specialty alloys, and minor metals inside it are a supply story of their own; the heat coming off it is this one.

That is the whole of the cooling problem stated in three numbers, and everything else follows from it. Air has a specific heat capacity of about one kilojoule per kilogram per degree Celsius and a density of roughly 1.2 kilograms per cubic meter. Water carries about four times the heat per kilogram and is roughly eight hundred times denser. Moving one hundred and forty kilowatts of heat with air requires a volume of air, at a velocity, that a data hall cannot physically supply and a fan array cannot economically push. The transition to liquid was not a sustainability decision. It was a thermodynamics decision that arrived with a sustainability press release attached.

Which brings up the thing that most coverage of data center cooling gets structurally wrong. The public argument is about water, treated as a scandal to be eliminated. The engineering reality is that water and electricity are two currencies for the same purchase, the heat has to leave the building through one of them, and every configuration on the market is a decision about which meter to run it through.

Why air cooling stopped working

Rack density is the variable that broke everything, and the curve is steeper than the public conversation registers.

A conventional enterprise rack in 2015 ran five to ten kilowatts. The Uptime Institute’s 2025 industry average sits around seven and a half, which tells you most of the world’s installed base has not moved. An eight-GPU H100 server draws roughly ten kilowatts on its own, putting an H100 rack in the thirty-five to forty-five kilowatt range. The GB200 NVL72 puts seventy-two Blackwell GPUs in one rack at one hundred and twenty kilowatts nominal, with observed full-load figures in the low one-thirties. NVIDIA’s Rubin-generation roadmap points at two hundred and fifty to nine hundred kilowatts per rack, and the industry is developing eight-hundred-volt DC distribution architectures for megawatt racks targeted around 2027. Whether those figures materialize is a separate question from whether they are being planned around, and infrastructure gets committed on the roadmap rather than on the shipped product, which is a structural feature of this buildout worth watching and which the long history of infrastructure projects sized against projected demand suggests is where the expensive mistakes live.

Air cooling has a practical ceiling somewhere in the twenty to fifty kilowatt range depending on how much you are willing to spend on containment, fan power, and floor space. Past that, the required airflow produces noise, vibration, and pressure differentials that become their own engineering problems, and the fan energy starts consuming a meaningful fraction of the power you were trying to deliver to the chips.

The individual processors tell the same story. The chips themselves depend on a supply chain with its own constraints, from the hafnium in the gate stack to the noble gases the lithography consumes, and the thermal envelope is fixed at design time long before a building exists. An H100 has a thermal design power around seven hundred watts. A B200 runs one thousand to twelve hundred. The GB300 generation pushes higher. At those densities a cold plate sitting directly on the package is not an optimization, it is the only mechanism with enough thermal conductance to keep the die below its throttling threshold.

So the first thing to understand about data center cooling in the AI era is that the choice was made by physics rather than by procurement. The semiconductor roadmap that produced these parts did not consult the facilities engineers, and the buildings had to follow.

The stranding problem that creates is worth naming early, because it recurs at every layer of this subject. A data hall built to twenty kilowatts per rack cannot host a Blackwell rack, and not for one reason but for four: the electrical distribution is undersized, the floor loading will not carry 1.36 metric tons in a rack footprint, there is no coolant distribution infrastructure, and the heat rejection plant is sized for a load that no longer exists. Retrofitting is frequently more expensive than building new, which is why so much of the current buildout is greenfield and why data center cooling design has become a constraint on where capacity can go at all.

What liquid cooling actually is

The term covers several distinct technologies with different costs, and conflating them makes the tradeoffs invisible.

Rear-door heat exchangers are the gentlest step. A radiator mounted on the back of the rack captures hot exhaust air and transfers the heat to a liquid loop, which means the servers remain air cooled internally and only the room-level heat rejection changes. It handles up to roughly forty or fifty kilowatts per rack and it is not sufficient at Blackwell density.

Direct-to-chip liquid cooling is the current standard for AI racks. Cold plates sit directly on the GPU and CPU packages, coolant circulates through microchannels inside them, and the heat moves into the liquid without ever entering the room air. It captures the large majority of rack heat, typically seventy to ninety percent, with the remainder still handled by air for memory, power supplies, and networking, which means a direct-to-chip facility still needs an air-handling system and still has a data hall with fans in it. Hybrid rather than replacement is the accurate description. This is what the GB200 NVL72 requires.

Immersion cooling submerges entire servers in a dielectric fluid. Single-phase immersion circulates the fluid to a heat exchanger. Two-phase immersion uses a fluid that boils at low temperature, exploiting latent heat for much higher heat flux, and it ran into a supply problem rather than a technical one: the fluorochemical fluids involved fall within the PFAS category, and 3M announced exit from PFAS manufacturing by the end of 2025, which removed a principal supplier from a market that had been building around those chemistries. That is a supply-chain constraint determining a thermal architecture, and it belongs alongside the other cases where a specialty chemical or material quietly gates a technology roadmap.

The plumbing that makes any of this work is the coolant distribution unit, which sits between two loops and is the piece of the architecture that the entire water argument turns on.

Every one of these approaches introduces failure modes that air cooling did not have. Liquid near energized electronics means leak detection becomes a safety system rather than a convenience. Coolant chemistry has to be maintained against corrosion, biological growth, and galvanic incompatibility between the metals in the loop. Filtration matters because a microchannel cold plate is a small aperture and fouling it destroys the thermal path. The cold plates themselves are typically copper, which puts them in the same industrial metals demand stream as everything else in the electrification buildout, and the quantities at fleet scale are not trivial. And the whole assembly introduces a serviceability problem, since a technician can no longer simply pull a hot-swap component without addressing the plumbing attached to it.

The two loops of data center cooling, and where the water goes

This is the distinction most coverage misses, and missing it makes the numbers unintelligible.

A liquid-cooled facility runs at least two separate loops. The technical cooling system is the closed loop that carries treated coolant from the coolant distribution unit through the cold plates and back. That loop is sealed. It is filled once, monitored for leaks, and it does not consume water in any ongoing sense. When a vendor says the cooling is closed-loop, this is usually the loop being described.

The facility water system is the second loop, and it is where the heat actually leaves. The coolant distribution unit transfers heat from the technical loop into the facility loop through a heat exchanger, and the facility loop has to dump that heat somewhere. The options are a cooling tower, which evaporates water, or a dry cooler or chiller, which uses mechanical refrigeration and fans, which uses electricity.

That is the entire argument, structurally. Direct-to-chip cooling moves heat out of the chip more effectively than air, and it does not by itself determine whether the building consumes water. A direct-to-chip facility rejecting heat through an evaporative tower still evaporates water. A direct-to-chip facility rejecting heat through dry coolers consumes almost none and spends more power doing it.

Which means the widely repeated claim that liquid cooling saves water is true only when the heat rejection stage changes as well, and the reason it usually does change is that direct-to-chip lets the loop run hot enough that dry rejection becomes viable. That is the actual mechanism, and it is more interesting than the marketing version.

There is a third loop in many facilities worth mentioning, since it accounts for a share of consumption that gets attributed to cooling. Humidification systems add moisture to data hall air to keep static discharge within tolerance, and that water is consumed outright. Modern designs have widened acceptable humidity ranges considerably and reduced this substantially, and it remains a line item in any honest data center cooling water accounting.

The physics of evaporating water, and why it is so cheap

Evaporative cooling looks wasteful and is thermodynamically excellent, which is why the industry adopted it and why abandoning it costs something real.

Vaporizing water absorbs roughly 2,260 kilojoules per kilogram at atmospheric pressure. That is the latent heat of vaporization, and it is an enormous amount of energy to move per unit mass with no compressor involved. A cooling tower works by exposing warm water to an airstream, allowing a fraction to evaporate, and returning the rest at a lower temperature. The energy that leaves is carried by the vapor.

The critical advantage is that evaporation can cool water below the ambient air temperature, down toward the wet-bulb temperature, which in a dry climate can be twenty degrees Celsius below the dry-bulb reading. A dry cooler cannot do that. It can only approach ambient, which means on a hot day a dry system either fails to reject enough heat or has to fall back on mechanical refrigeration, and mechanical refrigeration is where the power goes. That relationship is why an evaporative facility in a humid climate performs worse than the same facility in a dry one: high wet-bulb temperature collapses the advantage, which is why Gulf Coast siting carries a cooling penalty that Phoenix does not.

That relationship is why the industry built in Phoenix and Mesa in the first place. Arid climates have low wet-bulb temperatures, which makes evaporative cooling maximally effective, which makes the facility maximally power-efficient. The same aridity that makes evaporative cooling attractive is what makes water contentious there, and the siting logic and the political problem have the same cause.

Water consumption for an evaporative facility runs on the order of 1.8 to 1.9 liters per kilowatt-hour at industry average, with the best conventional designs reaching 0.3 to 0.7. At a hundred megawatts of IT load running continuously, the arithmetic gets large quickly.

Cooling towers also have a second water loss that gets omitted from casual descriptions. Evaporation concentrates dissolved minerals in the remaining water, and past a threshold the concentrated water has to be discharged and replaced, which the industry calls blowdown. Cycles of concentration, meaning how many times the water is recirculated before discharge, is the operational lever, and pushing it higher requires chemical treatment that then creates a discharge quality problem under the Clean Water Act. Water efficiency and water quality compliance are in tension, and the engineering that improves one frequently complicates the other.

WUE, PUE, and the trade nobody states plainly

The industry has two metrics and they move in opposite directions, which is the fact that most reporting omits.

Power usage effectiveness is total facility energy divided by IT equipment energy. A PUE of 1.0 would mean every watt entering the building reaches a chip. Evaporative designs achieve low PUE because evaporation does the cooling work without a compressor.

Water usage effectiveness is annual water consumption divided by IT energy, in liters per kilowatt-hour. An air-cooled facility with no evaporation has a WUE near zero and a PUE that is considerably worse.

Set them side by side and the substitution is visible. Improving WUE by eliminating evaporation degrades PUE by requiring mechanical cooling. Improving PUE by using evaporation degrades WUE by consuming water. These are not independent sustainability metrics that a well-run facility optimizes simultaneously. They are two readings on one tradeoff, and a facility reporting an excellent number on one of them without reporting the other is showing you half a ledger.

The aggregate direction is worth knowing. Lawrence Berkeley National Laboratory put the aggregate site WUE for the United States fleet at 0.36 liters per kilowatt-hour through 2023 and projected a rise to between 0.45 and 0.48 by 2028, even as individual facilities using the newest designs approach zero. Fleet-average water intensity is going up while best-in-class goes down, because the fleet is growing faster than the new designs are deploying.

The measurement conventions deserve a caution of their own. WUE as normally reported covers cooling and humidification at the site, which means it excludes the water embedded in electricity generation, the water used in chip fabrication, which is substantial and considerably more contaminated, and the water in construction. A metric that draws its boundary at the fence line will always make the fence line look good, and every party publishing data center cooling figures chose that boundary knowing it. The same boundary trick appears wherever an industry reports its own footprint, from by-product metals whose environmental cost sits with the primary producer to recycling rates that count manufacturing scrap and omit end-of-life.

The zero-water design, and what it actually costs

In August 2024 Microsoft announced a data center design consuming no water for cooling during operation, and the announcement is worth reading closely because it contains its own rebuttal.

The design uses chip-level closed-loop cooling with the coolant filled once at construction and recirculated continuously. Microsoft put the avoided consumption at more than 125 million liters per year per facility, with pilots at Phoenix, Arizona and Mount Pleasant, Wisconsin. Reported fleet WUE improved from 0.49 liters per kilowatt-hour in 2021 to 0.30 in 2024 and 0.27 in 2025.

Then the sentence that matters. In the company’s own announcement of the zero-water evaporation design, the stated result is a nominal increase in annual energy usage compared with the evaporative designs across the global fleet. The infrastructure engineering lead’s framing in trade coverage of the announcement was that moving from evaporative to mechanical cooling is expected to increase PUE, mitigated by chip-level cooling permitting warmer coolant temperatures and therefore high-efficiency economizing at elevated water temperatures.

That is the whole trade, stated by the party with the least incentive to state it. Water goes to zero. Power goes up. The mitigation is real and it is a mitigation rather than an elimination.

The warmer-water point is the genuinely clever part and deserves credit. Air cooling requires supply air in the range of eighteen to twenty-seven degrees Celsius, which forces cold coolant. A cold plate sitting on a die that tolerates much higher junction temperatures can run coolant at thirty to forty-five degrees, and warm coolant can reject heat to ambient air across far more hours of the year without a compressor running. Direct-to-chip does not merely move heat better. It raises the temperature at which the heat is available, which is what makes dry rejection affordable. That is the single most important sentence in the engineering of modern data center cooling, and it is almost never the one that gets quoted.

The water nobody counts

Here is the accounting problem that makes most of the public numbers incomparable, and it is the single most useful thing to carry out of this subject.

Thermoelectric power generation consumes water. Coal, gas, and nuclear plants reject waste heat through cooling systems that evaporate water at rates well above what a data center consumes directly per unit of electricity delivered. A facility that eliminates onsite evaporation and draws grid power from a thermal fleet has not eliminated the water. It has moved the water upstream, to a plant that reports it in a different document, under a different regulator, in a different county.

Which produces the uncomfortable version of the zero-water claim. A closed-loop facility with a slightly worse PUE draws more electricity, and if that electricity comes from thermoelectric generation, the additional offsite water consumed by generating it can exceed the onsite water saved. The scope boundary is doing the work, not the engineering.

That is not an argument against closed-loop design, and it is an argument for insisting that any water figure specify its boundary. Onsite consumption, onsite withdrawal, and total water footprint including generation are three different numbers, and coverage moves between them without notice.

It also explains why the cooling question cannot be separated from the power procurement question. A facility on hydro or wind has genuinely low water intensity in its power supply. A facility on combined-cycle gas does not. The cleanest way to make a data center water-efficient is to change what is generating its electricity, which is a decision made several layers away from the cooling plant. That is why geothermal generation and hydro attract data center siting for reasons that have nothing to do with carbon accounting, and why the nuclear restarts now being negotiated carry their own water footprint that rarely appears in the announcement. Grid-scale storage changes the arithmetic again, since a facility firming intermittent supply with flow batteries or lithium packs is buying a different water profile than one firming with a gas peaker.

Where the water comes from

Source matters as much as volume, and the public argument frequently conflates them.

Potable municipal supply is the contested case, since it puts a data center in direct competition with residential and agricultural users through the same infrastructure. Reclaimed and non-potable water is the mitigation the industry has leaned on hardest, with Google reporting reclaimed water at facilities in Singapore, Georgia, and elsewhere accounting for roughly twenty-two percent of data center cooling volume in 2023. Reclaimed water carries its own engineering cost, since treated effluent has higher dissolved solids and biological load than municipal supply, which means more aggressive chemical treatment, lower achievable cycles of concentration, and more frequent heat exchanger maintenance. It is the right answer environmentally and it is not free, and the treatment chemistry required is its own small specialty-chemical supply question at scale.

Groundwater is where the lecture title comes from and where the durable damage lives. An aquifer recharges on a timescale set by geology and precipitation, and pumping above the recharge rate is mining rather than use. The distinction that matters is between a facility drawing from a river with a permitted allocation and a facility drawing from a confined aquifer whose drawdown is measured in feet per year and whose recovery, if it happens, happens over decades. The same extraction arithmetic that governs any finite subsurface resource applies, and the fact that the resource is water rather than ore does not change the mathematics. The recharge arithmetic is the part that separates a permit dispute from a permanent loss. Land subsidence is the irreversible version: pumping a confined aquifer compacts the sediment, the pore space collapses permanently, and the storage capacity is destroyed rather than merely drawn down. Parts of the Central Valley and the Houston area have subsided by meters, and no amount of subsequent recharge restores the volume.

The siting overlap is the aggravating factor. The climates that favor evaporative cooling are arid. The regions with cheap land, cheap power, and permissive permitting frequently overlap with those climates. As of mid-2026, a substantial majority of the continental United States has been under drought conditions, and a meaningful share of new AI capacity is being built inside those regions.

The politics of a permit

The third lens is where the physics arrives at a county board agenda, and the pattern has become consistent enough to predict.

In August 2025 the Tucson city council unanimously rejected a data center project on water grounds. In South Carolina, residents near an overdrawn aquifer have pushed for limits on groundwater withdrawal by data centers. Comparable disputes have run in Arizona, New Mexico, Georgia, and internationally in Chile, Uruguay, and Spain. The pattern is consistent enough to be predictive: a facility announced with a jobs number and no water number, a permit process that surfaces the water number late, and a community that discovers the magnitude after the tax abatement has been negotiated.

The structural feature that makes these fights bitter is asymmetric information. Project water figures are frequently confidential during negotiation, sometimes under non-disclosure agreements signed by local officials, and a community is asked to approve a withdrawal whose magnitude it cannot verify. The economic development case arrives with employment figures, and the employment figures for a hyperscale facility are small relative to the capital, because a data center is an unusually capital-intensive and labor-light industrial installation. A multi-billion-dollar campus may employ a few dozen to a few hundred people in steady state, which is a ratio no traditional manufacturing plant of comparable capital would produce, and which changes what a county is actually being offered. The critical minerals buildout runs the same negotiation with the same asymmetry, and the counties on the receiving end have generally learned the same lessons at the same speed.

Water rights add a legal layer that varies enormously by jurisdiction. Prior appropriation in the western United States assigns seniority by date of first use, which means a new entrant gets a junior right that is curtailed first in a shortage, unless it purchases a senior right from an existing holder, which is generally agricultural. That transaction is legal, orderly, and locally understood as a data center buying a farm’s water. Whether that is a good trade depends on what you think the water was producing, and reasonable people in the same county reach opposite answers, which is why these fights do not resolve on technical grounds.

None of this is unique to data centers. It is the standard pattern for any large industrial facility that arrives in a rural county with an extraction requirement, and the towns that were built around a single extractive employer are the long-run version of the same negotiation. The difference is duration. A mine works an orebody for decades. A data center’s silicon has a useful life closer to five years, and what happens to a building sited on a water permit when the hardware inside it becomes obsolete is a question nobody has had to answer yet.

What the numbers do and do not say

An audit, because this subject has produced more circulated figures with unstated units than almost any other in current technology coverage.

Withdrawal and consumption are different quantities and get used interchangeably. Withdrawal is water removed from a source; consumption is water not returned to it. A once-through cooling system withdraws enormously and consumes little. An evaporative tower withdraws less and consumes most of what it takes. Any comparison between facilities using different figures is meaningless.

The claim that an AI query consumes about half a liter of water is the most-shared figure in the field and it rests on a specific 2023 analysis with specific assumptions about model, hardware, region, and whether offsite generation water is included. It is a defensible estimate under its own assumptions and it is not a constant of nature, and it circulates without any of them.

Training GPT-3 evaporated 700,000 liters is from the same body of work and carries the same caveats, and it describes one training run on one generation of hardware in specific locations, on a model that has since been superseded by architectures with different efficiency characteristics.

A data center uses as much water as a town of X people is a comparison that requires knowing whether the town figure is withdrawal or consumption, and whether the data center figure includes generation.

Liquid cooling eliminates water use conflates the technical loop with the facility loop, as above.

Zero-water means zero water is true at the fence line and not true at the system boundary.

And the reverse error deserves equal treatment. Data centers are draining the aquifers overstates a share that, nationally, remains small against agriculture, which accounts for the overwhelming majority of United States freshwater consumption. The legitimate concern is local and concentrated rather than aggregate, which is a different argument and a stronger one. A facility consuming a rounding error of national freshwater can still be the largest single withdrawal in its county, and county is the unit at which the permit gets issued and the aquifer gets drawn down.

What data center cooling is actually telling us

Assemble the pieces and the cooling problem stops looking like an environmental controversy and starts looking like a thermodynamics ledger with a political interface.

The heat is not optional. A hundred megawatts of IT load produces a hundred megawatts of heat, essentially all of the electrical input, because computation does not store energy. Essentially every joule delivered to a chip leaves as heat, which makes a data center a resistive heater that happens to produce tokens on the way through, and which is why the materials and thermal engineering constraints that govern any high-flux industrial process apply here without modification. That heat leaves through evaporation, which spends water, or through mechanical refrigeration, which spends power, or through ambient rejection when the temperature differential permits it, which is free and unavailable on a hot afternoon. There is no fourth option, and every design on the market is a weighting of the first three. Data center cooling is therefore not a problem that gets solved. It is a ratio that gets set, and the setting is made by climate, by power price, by water price, and by which of the two the local politics will tolerate.

The engineering is genuinely improving and the improvement is real. Warmer coolant temperatures from direct-to-chip cooling extend the hours per year when free rejection works. Closed-loop designs eliminate onsite evaporation. Heat reuse into district heating networks captures value that would otherwise be rejected, which works where the district heating network exists and is a European result more than an American one so far. The physics constrains it: rejected data center heat arrives at thirty to forty-five degrees Celsius, which is useful for space heating and useless for industrial process heat, so the addressable demand is space heating in cold climates near the facility, which rules out the higher-value industrial process heat applications entirely. It is a genuine efficiency gain in Denmark and Finland and mostly unavailable in Arizona, and it is the closest this industry comes to the closed-loop recovery that works elsewhere only when the waste stream is concentrated and arrives at a known address.

Air-side economization is the other lever and it is climate-bound in the same way. A facility in a cold dry climate can reject heat to outside air directly for a large fraction of the year, which is why the Nordics and the upper Midwest attract capacity, and why siting decisions are as much a thermal calculation as a real estate one.

What is not improving is the aggregate, because capacity is growing faster than efficiency. Fleet WUE is projected upward even as best-in-class approaches zero, and that gap is the whole story of this buildout in miniature: the newest facility is always the most efficient one, the fleet average is set by everything already built, and the fleet is doubling. Efficiency per unit is a real achievement and it is not a trend line that intersects the aggregate, which is the same relationship that shows up in every materials system where demand growth outruns intensity improvement.

Which leaves the question the ten-lecture briefing on how AI data centers work exists to work through. The cooling stack determines how much power the building needs on top of the compute. The power determines what has to be built to supply it. And the thing being supplied is a load that did not exist five years ago, arriving at a grid that spent twenty years planning for flat demand.

The useful discipline, and the one that survives contact with every figure in this subject, is to demand three things of any water number: the boundary it was measured at, whether it is withdrawal or consumption, and what the corresponding power number did. A water figure without its power figure is not a measurement of environmental impact. It is a measurement of which meter somebody chose to read.

A cooling tower in Arizona and a combined-cycle plant two hundred miles away are consuming water for the same reason: something has to absorb the heat, and evaporation is the cheapest absorber available. Move the tower indoors and the absorber moves to the power plant. The bill does not disappear. It changes address.