The megawatt and the degree

> 60 MW

electrical power of the Alixan site “once complete”

> 60 MWth

rated thermal power of the backup generators, “even higher”

On 10 July 2026, the judge hearing urgent applicationsUrgent-procedure judgeThe judge who rules urgently, for example to suspend a decision pending a ruling on the merits.See the glossary → at the administrative court of Grenoble suspended a building permit. The project: a “Computer Center” dedicated to artificial intelligence, €1.5 billion, led by the company Sesterce on the Rovaltain business park, next to the Valence TGV station. The permit had been signed by the mayor of Alixan on 18 December 2025.

What interests me is not the decision. It is the reasoning.

The judge does not talk about bottles of water per promptPromptThe written instructions given to the AI model to steer its answer.See the glossary →. He does not talk about tokensTokenTwo meanings. For a model: a piece of a word, the unit that measures processed text and therefore cost. In security: a temporary key proving an access right.See the glossary →. He notes two things.

From these two numbers he infers that an environmental impact assessmentImpact assessmentThe file that assesses a project’s effects on the environment (water, air, noise, biodiversity), mandatory above certain thresholds.See the glossary → was mandatory, and that it was missing. He adds a doubt about the local zoning plan, which does not allow this kind of classified installationClassified installationAn industrial facility that presents risks or nuisances and is subject to stricter environmental rules (ICPE in France).See the glossary → on this plot. And he suspends.

A power. A heat. Without saying so, the judge asked the only two questions that matter. Everything else — water, local opposition, the regulation coming from Brussels, Sacramento, Albany and Richmond, the geopolitics of a Virginia county — follows from those two.

Pulling that thread all the way, I reached a conclusion I did not expect. You hear everywhere that AI is moving too fast and that this is what makes it dangerous. I think that is the wrong framing. What moves too fast is the electrical power it calls for — and what is dangerous is not that speed in itself, but the gap between it and the speed at which a grid, a river basin and a local democracy can respond. Today that gap is filled with gas, by bypassing the public grid, and by sending the bill to households. That is the measurable risk. The restRESTThe most common style of web API: data is exchanged through addresses and simple verbs (read, create, update, delete).See the glossary → of this article is the demonstration.

I read the full text of what I quote whenever it was available, including the European delegated regulation of 21 September and two academic studies whose limits I detail. When I could not, I say so.

1. A unit that changes everything: the megawatt

When people talk about the impact of AI, they almost always talk in watt-hoursWatt-hour (Wh, kWh, TWh)A unit of energy: a power sustained over time. 1 kWh = 1,000 W for one hour; 1 TWh = a billion kWh.See the glossary →. A prompt is 0.24 Wh according to the measurement Google published in August 2025. That is energy: a quantity, accumulated over time, like the litres running through a meter.

A data center, however, is sized in megawattsMegawatt (MW)A unit of power: what must be supplied at every instant. 1 MW = 1,000 kW. Not to be confused with the megawatt-hour, which measures an amount of energy used over time.See the glossary →. That is power: what must be supplied at every instant, day and night, without interruption. The grid does not deliver kilowatt-hours “when it has some”; it must hold 60 MW to the second, at 3 a.m. in January as at 3 p.m. in August.

To picture 60 MW: France consumes about 456 TWh a year, an average power draw of just over 52 GW for 68 million people, all uses combined. That is roughly 0.76 kW per person. The Alixan site alone would permanently draw the equivalent of 75,000 to 80,000 people. Roughly Valence, the neighbouring town.

This is not an exception. It is the new norm.

Density has broken away. For forty years, a server cabinet — a rackRackThe standard cabinet that holds servers: 60 cm wide, about two metres tall. Its power is measured in kilowatts.See the glossary →, 60 cm wide, two metres tall, 42 slots — drew between 5 and 12 kW. A few electric heaters in a cupboard, and the industry knew very well how to cool that by blowing air. Average density rose from 16 kW per rack in 2025 to 27 kW in 2026. An Nvidia GB200 NVL72 rack, with its 72 GPUsGPUGraphics processing unit: the massively parallel chip used to train and run AI models. It is what draws and heats the most.See the glossary →, draws 120 to 140 kW, more than a gigawatt-hour a year: at full load, the consumption of about 240 French households, in one cabinet. The next platform, Vera Rubin, announces up to 246 kW per rack.

At this point a data center is no longer an office building with computers inside. It is an electro-intensive plant, in the sense we speak of an aluminium smelter.

The scale. The International Energy Agency counts 415 TWh consumed by data centers in 2024, 1.5% of global electricity, and projects about 950 TWh in 2030, or 3%. Consumption grew 17% in 2025, and that of AI-dedicated data centers 50%. For the European Union, the delegated regulation of 21 September 2026 puts forward its own figures: 68 TWh in 2024, 114 TWh expected in 2030, 3.2% of EU demand. France, with its 393 data centers, is small in this game: about 10 TWh, 2% of its consumption.

What stands out is not the total. It is the speed, more than four times that of the rest of electricity demand, and the concentration: the United States accounts for 45% of the world total, and a handful of counties hold most of it.

2. The degree: where all that electricity goes

It is worth recalling something we forget because it is too simple: the electricity that goes into a data center comes out entirely as heat. A 60 MW site is a 60 MW radiator. There is no other way out.

At chip level. An H100 GPU dissipates about 700 watts, a B200 between 1,000 and 1,200, an MI355X up to 1,400. Per unit of area, an H100 must remove 86 watts per square centimetre — more than an induction hob, on a surface the size of a stamp.

The limit of air. Air only carries heat away fast enough up to about 50 W/cm², roughly 35 kW per rack. Beyond that, chips throttle themselves so as not to burn out. That is the physical reason, not a marketing choice, why liquid coolingLiquid coolingRemoving chip heat with a liquid rather than air: cold plates on the chips (direct-to-chip) or servers immersed in a fluid (immersion).See the glossary → went from 3% of deployments in 2021 to 37% in 2026.

The metric to know: PUEPUEPower Usage Effectiveness: a data center’s total energy divided by the energy reaching the servers. 1.0 would be perfect; 1.5 means 50% extra energy for cooling and losses.See the glossary →. You divide the total energy consumed by the site by the energy that actually reaches the servers. A PUE of 1.5 means spending half a kilowatt-hour on cooling, lighting and losses for every kilowatt-hour of compute. Cooling accounts for 30 to 40% of the bill according to the Uptime Institute. In the text of 21 September, the European Commission puts the European average at 1.6 and notes that moving to 1.2, “achievable with the best technology available today”, would cut a data center’s electricity consumption by 25%. A quarter of the electricity, on the choice of thermal system alone.

In a liquid installation, two circuits cross without mixing, separated by a coolant distribution unit (CDUCDUCoolant Distribution Unit: the unit that separates and exchanges heat between the chips’ liquid loop and the building’s water loop.See the glossary →): a closed loop of ultra-pure fluid running to the chips, and a building water loop that goes up to the roof to reject the heat. Keep this picture in mind: when you are told a site “consumes” so many litres, the question is always which loop we are talking about, and whether that water evaporates or comes back.

The good technical news is that the temperature is rising. The hotter the liquid leaving the chips, the less refrigeration you need to cool it. Nvidia announced at CES in January 2026 that Vera Rubin accepts 45 °C liquid. At that temperature, simple dry coolersDry coolerA heat exchanger that cools a liquid with outside air, without evaporating water. It uses a bit more electricity than an evaporative tower.See the glossary → are enough, with no compressor and no water. On 21 September 2026 Schneider Electric published a white paper comparing four 100 MW architectures in the climates of Paris and Dallas: 45 °C liquid cuts water consumption by at least 50% compared with air cooling. One caveat to read alongside: Schneider bought 75% of Motivair, a liquid-cooling specialist, in 2024, and aims to buy the rest by 2028. It is an interested source. The order of magnitude is consistent with the rest of the literature, but I do not treat it as an independent measurement.

3. Water is not a third problem, it is the same one

Here is the point that defuses half the controversies.

Water and electricity are substitutes. Evaporative cooling uses little power and a lot of water. A dry cooler uses almost no water on site, but more power. And producing that power uses water elsewhere, at the power plant: zero litres per kilowatt-hour for wind and solar, 0.8 for gas, 2.2 for coal, 3.3 for nuclear, and up to 68 for hydro if you count evaporation from reservoirs.

A “waterless” data center therefore does not remove water. It moves it from its own site to the power plant’s, sometimes hundreds of kilometres away.

What the most precise measurement says, and what it does not. In June 2026, two researchers at the Harvard T.H. Chan School of Public Health, Guidi and Dominici, mapped 472 US hyperscaleHyperscaleThe very large data centers of cloud giants (AWS, Microsoft, Google, Meta), sized in tens or hundreds of megawatts.See the glossary → sites (20,041 MW, about 116 TWh a year). Their baseline result: about 300 billion litres of water a year, of which 74 (a quarter) is direct water on site and 226 (three quarters) indirect water, consumed to generate the electricity. Virginia, the world’s leading hub, is almost entirely in the second case: 68 billion litres indirect against 6.6 direct. In Virginia, the water problem is a power-plant problem.

I read the paper in full, and two things need saying that the abstract does not. First, it is not yet peer-reviewed — it is an arXiv preprintPreprintA scientific paper published before peer review. Its results should be read with caution.See the glossary → (2607.02531). Second, the authors themselves call the hydroelectric coefficient they use (8 litres per kilowatt-hour) “methodologically contested”: without it, indirect water drops by 43%, from 226 to 128 billion litres. The famous “indirect is three times direct” ratio thus rests largely on a disputed accounting convention. The direction holds — a data center’s water is mostly its power plant’s water — but the magnitude is less certain than announced.

Withdrawal, consumption, and the misleading word. Evaporated water is not destroyed. It falls again, nine days later on average, somewhere else. The USGS defines consumption as the share of withdrawn water that is evaporated, incorporated into a product, or returned to another river basin. The accounting is about a place and a season, not the molecule. Actual rates are known site by site from Google’s data, audited by EY: an evaporated share of 0.576 in Henderson, Nevada, and 0.826 in Douglas County, Georgia. A closed loop consumes 5 to 10% of what it withdraws; an evaporative system loses 70 to 80%. Quoting a withdrawal as “consumed” multiplies the figure by ten.

The metric that will matter: WUEWUEWater Usage Effectiveness: litres of water consumed per IT kilowatt-hour.See the glossary →. Litres of water per IT kilowatt-hour. The European regulation notes that going from 1.0 to 0.5 halves a site’s water consumption. WUE and PUE are the two numbers that will appear on the European label (part 8). They are the only two that translate the degree directly into litres and kilowatt-hours.

4. Climate decides the rest, and it is worsening faster than measured

Free coolingFree coolingCooling a data center with outside air rather than a refrigeration machine. Only possible when it is cool and dry enough.See the glossary → — cooling with outside air rather than with a machine — only works when it is cool enough often enough. That is what makes Nordic countries attractive, and what makes the climate question inseparable from the electricity question.

A study published in Scientific Reports (Karamperidou et al., universities of Hawaii and Maryland, NSF-funded, peer-reviewed) crossed the ERA5 climate reanalysis for 1940–2025 with the ASHRAE thresholds beyond which free cooling stops being possible: 27 °C with relative humidity above 70%, or a dew pointDew pointThe temperature at which moisture in the air condenses. The higher it is, the more humid the air and the harder it is to use for cooling.See the glossary → above 15 °C, with a conservative margin of 1.5 °C between outside air and the rack.

What I found most interesting is the limit the authors themselves acknowledge. Their model does not resolve urban heat islandsUrban heat islandA built-up area warmer than the surrounding countryside, because concrete and human activity store and release heat.See the glossary →, nor the heat data centers release into the surrounding air. Yet that air is precisely what they hope to cool themselves with. The study therefore probably underestimates the problem it documents. A cluster of data centers locally warms the air it needs cold: it is a loop, and it closes the wrong way.

5. Where AI’s megawatts are

Before talking about speed, you need to know where things stand. I mapped the sites of the three labs people talk about most — xAI, OpenAI, Anthropic — keeping only what is located, quantified and dated.

xAI, Memphis. Colossus 1 is a former Electrolux factory in south Memphis, converted in 122 days in summer 2024 to host 100,000 GPUs. Colossus 2, in the Whitehaven neighbourhood, received its first cluster in January 2026 after 91 days of work; it is approaching 350,000 GPUs. When xAI arrived, the local grid offered only 8 MW. The answer was mobile gas turbines, installed without Clean Air Act permits, then regularised by a 1.2 GW plant with 41 turbines across the Mississippi state line — more than half the power of the Hoover Dam, for a single customer. The neighbouring, mostly Black, Boxtown neighbourhood breathes the nitrogen oxides; the NAACP sued in April 2026; in June the Department of Justice intervened — on xAI’s side, in the name of “national, economic and energy security”.

OpenAI, Stargate. Seven US sites, hardware owned by Oracle or SoftBank, $500 billion announced, more than 9 GW targeted. In spring 2026, only one was running: Abilene, Texas, 0.3 GW. The three largest — Shackelford County in Texas (2 GW), Doña Ana County in New Mexico (2.2 GW), Abilene — are or will be powered by natural-gas microgrids, off the public grid. The only mostly renewable site is in Wisconsin. The Michigan one is already contested; the Ohio one will soon face a local ban on new data centers.

Anthropic, fifteen campuses, nothing owned. Anthropic owns no building, no substation and, most often, not the chips. It leases: AWS cloud (Project Rainier, Indiana, Mississippi, Pennsylvania), Google TPUs, an upcoming Nvidia campus in West Virginia, and above all a series of sites in Texas, Kentucky, Louisiana, New York and Indiana that share something I did not see coming: they are former bitcoin mining farms — TeraWulf, Riot, Hut 8, Cipher. These sites already have the substation and a grid connection of several hundred megawatts. Anthropic is not buying land, it is buying megawatts already plugged in, the scarce asset of 2026.

And the three cross paths. Since May 2026, Anthropic has leased most of Colossus 1 from xAI — 500 MW, $45 billion over three years, terminable at 90 days. Since June, Google has leased 110,000 GPUs there for $920 million a month. The three rivals share the Memphis turbines. And none of these sites is in Virginia: newcomers go where gas and land are left, not where the historic internet is.

Remember three things from this map. Gas is everywhere speed is needed. An existing grid connection is worth more than the land. And the gap between announced and real is huge — 0.3 GW out of 9 at OpenAI.

6. Why it is in Virginia, and why it is so hard to move

If the megawatt and the degree are the two variables, geography is what makes them political. And there is no better case than Data Center Alley, in Northern Virginia, 40 km from Washington.

The figures come from geography work by CNES and Géoconfluences by Laurent Carroué (July 2025), based on Pléiades satellite imagery. Virginia has 564 data centers, run by 83 companies, 155 of them by Amazon Web Services alone. The Ashburn hub, in November 2025: 154 centers, two million square metres, 2,830 MW installed. The power of this single area exceeds that of Dublin, London, Frankfurt, Amsterdam, Singapore and Sydney combined. On my scale from part 1, it is the average power drawn by 3.5 to 4 million French people, on a strip of land along an airport. It is also more than xAI, OpenAI and Anthropic combined actually have in service today.

Why there. Because everything piled up there. The Pentagon, the CIA, the NSA and the FBI in the same metropolitan area. ARPANET, the ancestor of the internet, born in Arlington in the late 1960s. MAE-East, one of the first major internet exchange points, in Ashburn from 1998; more than half of US internet traffic already passed through it in 2009. Long-cheap electricity supplied by Dominion Energy, water from the Potomac, land, and a tax exemption on equipment passed in 2009 and extended to at least 2035, worth $1.7 billion cumulatively between 2014 and 2023. It is not a market that chose a place; it is a place that made a market. Economists call that path dependence.

The megawatt, Virginia edition. Data centers account for 20 to 25% of Dominion Energy’s sales, and Dominion already buys 22% of its needs from outside, at a high price. Contracted capacity rose from 931 MW to 3,888 MW in 2025, and should reach 7,686 MW in 2033. The state estimates it will need 11,000 MW by 2035 — a fourfold increase in thirteen years. Loudoun County has more than 4,000 backup generatorsBackup generatorAn engine, often diesel, that produces emergency electricity when the grid fails. Its thermal power exceeds its electrical output.See the glossary →, tested every month. Remember the Grenoble judge and the generators’ “rated thermal power”: it is the same object, multiplied by four thousand. And the new 500 kV line meant to secure Ashburn is still at the route-selection stage.

The degree, Virginia edition. The grid powering US data centers is on average dirtier than the national average: 548 gCO2e per kilowatt-hour against 369, because sites settle where electricity is abundant and connectable, not where it is clean. Virginia is at 576. And its water, as we saw, is three-quarters its power plants’.

The political price. The Briarfield Estates and Hiddenwood subdivisions, opened in 2013 in a rural area, are now surrounded by digital warehouses. In July 2025, the county refused residents the rezoning to industrial use that would have let them sell and leave. The median home in Loudoun is worth $983,625. Data centers bring $9.1 billion to the state’s GDP and a quarter of the county’s tax revenue — but of 74,000 jobs, the vast majority are construction jobs; an operating site employs a few dozen people.

What has just changed: when Carroué wrote, Republican governor Glenn Youngkin was pushing development. His Democratic successor, Abigail Spanberger, in office since January 2026, presented a plan on 18 September that speaks exactly the language of this article — mandatory local approval above 25 MW, an end to fast-track permits for large sites, transparency and a ban on confidentiality clauses, restrictions on water-hungry cooling towers, incentives to replace diesel generators with batteries. The Register summed it up with a phrase the governor did not use but that says it all: data centers have become a “political cancer”. These are announcements, not laws. The Virginia General Assembly will decide.

And where we stand. 80% of European cloud spending — a €330 billion-a-year market according to Cigref — goes to the United States. 70% of French digital data is hosted across the Atlantic. On 20 October 2025, an AWS outage that started in a Virginia data center simultaneously took down Snapchat, Fortnite, Venmo, Lloyds bank, Airbnb, Reddit, Zoom, Perplexity and Netflix. Loudoun’s megawatts are, in part, ours.

7. The speed gap

This is the heart of the article. Everything before describes quantities; what makes the situation dangerous is rhythms.

The rhythm of AI is counted in days. A GPU is depreciated over two to four years, and the model race leaves nobody time to wait. The rhythm of the grid is counted in years. In the United States, at the end of 2025, 8,200 projects were waiting for a transmission connection: 1,312 GW of generation and 749 GW of storage, according to the Lawrence Berkeley National Laboratory. The typical delay between request and commissioning is about five years. Of the requests filed between 2000 and 2020, only 13% were completed; 75% were withdrawn.

Between these two rhythms there is a gap. And in an economy, a gap never stays empty.

First way to fill it: gas. The US queueInterconnection queueThe list of projects (power plants, storage, large consumers) waiting to be connected to the transmission grid.See the glossary → changed character in a year. Gas rose by 86% (253 GW) while solar fell 19%, storage 16% and wind 19%. Global Energy Monitor counts 252 GW of gas plants in development in the United States, a quarter of the world total, and more than a third of that capacity is meant to power data centers directly on site. In Texas alone, 40 GW out of 80. The IEA expects gas and coal to cover more than 40% of additional data-center demand by 2030, and notes that coal plant closures are being postponed for that reason. Order books confirm it: GE Vernova went from 100 to 116 GW of turbines in one quarter and targets 125 GW by the end of 2026, expanding its Greenville plant by 35%; Siemens Energy reports a record €162 billion backlog, and 60 to 65% of its gas turbine orders this year come from data centers. The installed price of a combined-cycleCombined cycleA gas plant that recovers heat from its turbines to produce more electricity. It is the most efficient gas technology.See the glossary → plant doubled in fifteen months, from $1,000 to over $2,000 per kilowatt.

Second way: bypass the grid. It is the fastest and least visible move. By mid-2026, Cleanview counts about 90 GW of “behind-the-meterBehind the meterElectricity generation installed on the consumer’s own site, off the public grid: turbines, engines, fuel cells.See the glossary →” generation announced for US data centers — more than a quarter of all planned data-center capacity in the country. 92% of these announcements are less than twenty months old. And 2 GW are in service. The rest is at the permit stage (36%) or announced (60%). The suppliers are those who deliver fast: Caterpillar for a third, Bloom Energy for 14%, aeroderivative turbines, engines, refurbished turbines. The top five states, led by Texas, account for 83%. Memphis is not an anomaly; it is the prototype.

Third way: send the bill to households

On PJMPJMThe largest grid operator and electricity market in the United States (Virginia and twelve other states). Its capacity auctions pay for power available at peak times.See the glossary →, the largest US electricity market (Virginia and twelve other states), the capacity price rose from $28.92 per megawatt-day (2024–25) to $329.17 (2026–27). Eleven times more. IEEFA attributes 63% of the increase in the 2025–26 auction to data centers: $9.3 billion passed on to all customers. Dominion zone (Virginia): $444. Baltimore zone: $466. On a residential bill in Ohio or western Maryland: +$16 to $18 a month right now, and an estimate — which I give as such — of +$70 a month in 2028.

That is why opposition has changed character. It is no longer only the Briarfield Estates neighbour; it is the Baltimore customer who has never seen a data center and pays for them. Seven Americans in ten oppose a site near their home. A hundred and twenty projects were blocked or delayed in the first half of 2026.

The counter-argument to take seriously: speed is also the speed of announcements. Sightline Climate counts 39 GW of operational data centers in North America, 35 under construction and more than 129 announced. Of the 16 GW supposed to be delivered in 2026, 5 were actually under construction in spring; 30 to 50% of this year’s pipeline will not happen. In 2025, 26% of expected capacity slipped. Sightline adds a sentence that sums it up: “a data center announcement is a request for electricity, not a commitment to build”. Joining the queue costs almost nothing. And Sightline names the real bottleneck of 2026: “neither capital nor chips — the electrical layer”, meaning high-voltage transformers and medium-voltage switchgear.

This counter-argument does not overturn the thesis, it sharpens it. If the pipeline deflates, the danger is financial — Oracle’s debt, the high-yield bonds of Meta and CoreWeave projects — before it is physical. But what has already been built was built the same way: fast, on gas, off the grid. And what has already been paid for is paid for by PJM households.

The danger is not that AI moves too fast. It is the gap between how fast it calls for power and how fast a grid, a basin and a town can respond. Today that gap is filled with gas, by bypassing the public grid, and by shifting costs onto people who asked for nothing. It is local, it is dated, it is measurable — and it is already being regulated.

Three caveats. This is not a planetary energy problem: 3% of world electricity in 2030, less than 1% of emissions. It is not specific to AI: electric vehicles and heat pumps also drive demand, but none at this speed or with this concentration. And it is not universal: China builds gas and the grid that goes with it, France is at 2% — the speed gap is to a large extent an American institutional problem, which AI reveals rather than creates.

8. Regulation is coming, and it speaks exactly these two languages

What made me write this article is a coincidence of dates. In five days, from 18 to 22 September 2026, five jurisdictions moved, and all chose the same variables: a power that triggers, an efficiency that ranks, and — the new part — a cost that changes hands.

1. Brussels21 September500 kW threshold

   The Commission adopted a delegated regulation (C(2026) 3472 final, signed by Ursula von der Leyen) establishing a common rating scheme for data centers. I read all fifteen pages, and it corrects what the press said the next day. It is not a label “voluntary for sites above 500 kW”. It is the opposite. Since 2024, every EU data center with at least 500 kW of IT power has already been required to report its energy and water performance each year to a European database, under article 12 of the Energy Efficiency Directive; two reporting rounds have taken place. The 21 September text automatically generates, from these mandatory reports, an electronic label rating each site on two scales: PUE and WUE. First label on 15 August 2027, then every year. Only sites below 500 kW and those not yet in service take part on a voluntary basis. And article 6 sets a review on 31 December 2028, with the explicit possibility of introducing an aggregate indicator, a certification or an audit of the reported data. The text says so itself: this label is a precursor to minimum standards.

   The memorandum also contains the most telling figure in the file: reusing half the waste heatWaste heatHeat produced by an activity and released without being used. Data-center heat can warm homes.See the glossary → of European data centers would cover the heating of nearly four million homes. The degree is not only a problem; it is a resource being thrown away.

2. Sacramento21 September7 laws

   Gavin Newsom signed seven laws, a year after vetoing one for fear of slowing AI down. Three shift onto operators the electrical infrastructure costs previously borne by households (SB 1168, SB 886, AB 2383) — the direct answer to the third way of filling the gap. Three require disclosure of water and other resource use (AB 2469, AB 1577, AB 2619). One removes the automatic exemption from environmental review, with a fast track for efficient sites (SB 887).

3. Albany21 September50 MW threshold

   Since July 2026 New York has had a one-year moratorium on any new data center above 50 MW. Governor Hochul added continuous reporting obligations from 1 January 2027.

4. Austin18 September

   The governor of Texas ordered penalties for operators who do not report their consumption.

5. Richmond18 September25 MW threshold

   The Spanberger plan, and bill SB 253 to shift capacity and distribution costs onto data centers.

6. Grenoble10 July~50 MW threshold

   The urgent-procedure judge: an impact-assessment threshold that Reporterre puts at 50 MW, the power of the generators, the local zoning plan.

Look at what these texts have in common. The trigger is a power everywhere: 500 kW to report in Brussels, 25 MW for local approval in Virginia, 50 MW for the New York moratorium and for the French impact assessment. The metric is everywhere the ratio between incoming electricity and useful compute. And the rising question is everywhere the same: who pays for the grid connection. Nobody regulates prompts. Everybody regulates megawatts, cooling, and the bill.

9. What it changes for you, and what it does not

Let’s go back to the prompt, since that is what we feel guilty about.

Google measures 0.24 Wh and 0.26 mL of water — five drops — for a median Gemini prompt. Mistral, in a life-cycle analysis audited by Carbone 4, counts 45 mL for a 400-token answer. A factor of 170 between two serious players, which does not come from a lie but from everything above: models, scopes, sites, electricity mix, carbon accounting method. There is no common measurement standard. That is exactly what the European label is starting to build.

For text and images, your usage is not a lever. For video, it starts to be one. The only individual decision that matters is the choice of modality.

The rest is not decided at your keyboard. It is decided in the choice of site, cooling system, electricity mix and who pays for the grid connection — that is, in building permits, capacity auctions and labels. These are collective decisions, and they are being made right now.

One last trap: Google announces a 33-fold reduction in energy per prompt in twelve months. It is measured, and it is real. Over those twelve months, total data-center consumption rose by 17%. That is the Jevons paradoxJevons paradoxWhen a resource becomes more efficient to use, people use more of it, and total consumption can rise.See the glossary →, formalised for AI by Luccioni, Strubell and Crawford (FAccT 2025): what becomes cheaper gets used more. Efficiency per request guarantees nothing about the total. That is why regulators chose to rate sites and cap power, not prompts.

10. What I don’t know

An article that does not list its gaps is not reliable. Here are mine.

• The “three times more indirect water than direct” ratio depends on a contested hydroelectric convention; without it, it is closer to 1.7.
• Nobody has done for France the mapping Guidi and Dominici did for the United States.
• The exact thresholds of the PUE and WUE classes of the European label are in annexes I could not read.
• Cleanview’s 90 GW “behind the meter” are 98% announcements, and I have shown what announcements are worth. The Memphis water figures come from opponents, not the operator.
• The estimate of $70 a month in 2028 on PJM comes from a secondary source.
• Three studies I cite second-hand (Barnett-Itzhaki, Usman and Zakir, Hankendi et al.) are behind paywalls. The Schneider white paper is from a liquid-cooling vendor. And “political cancer” is a newspaper headline, not a quote.

Above all: this thesis is written to be falsified. If in 2027 connection delays fall below three years, if the PJM auction drops back below $100, if gas and coal cover less than a quarter of additional demand, it is wrong. I am keeping the list.

11. The loop, on one page

AI makes rack density explode. Air reaches its physical limit, and that limit gets closer as the climate — and the heat of data centers themselves — reduces free-cooling hours. We switch to liquid. The trade-off between water and electricity is not one: it is the same water, moved to the power plant. Sites concentrate where megawatts can be connected, not where they are clean or where water is plentiful. Then the rhythm of AI — days — meets the rhythm of the grid — years — and the gap fills with what is fast: gas, mobile turbines, private power plants, and a bill sent to people who asked for nothing. Opposition stops being local. And regulation, in five days of September 2026, converges on three variables: a power that triggers, a thermal efficiency that ranks, a cost that changes hands.

The Grenoble judge did not need to know how much water a prompt uses. He asked for the site’s electrical power, and the thermal power of its backup generators. It was the right question. It is not AI that is moving too fast. It is us who have no grid at its speed — and who, in the meantime, burn gas.