In 2010, Virginia introduced a tax exemption for data centers and got a dense concentration of facilities. Now the same state has placed a consumption tax on data centers' electricity use: $0.011 per kilowatt-hour starting July 1, 2026, and expiring in 2028. Virginia has not turned anti-tech, and the exemption remains in place. But the state's own oversight agency, JLARC, calculated in December 2024 that every dollar of revenue forgone through the exemption returned an estimated 48 cents to the state treasury. My conclusion is that someone has redone the math.
The benefit comes in the construction phase, the cost in operations. In 2016, Good Jobs First estimated that an operating data center needs 30 to 50 permanent employees, and found that eleven American megadeals had cost an average of $1.95 million in subsidies per permanent job.
Gallup found in March 2026 that 71 percent of Americans oppose an AI data center in their local area, 48 percent strongly. The reasons are concrete: the electricity price, the water, the noise, what the local government got in return. Not every story holds up. The water withdrawal in Georgia that is usually described as unauthorized went through two unmetered connections, just over 29 million gallons, and was billed afterward without penalty. The objections are not weakened by the worst version being wrong.
Above the local level, another conversation is under way, and it is conducted honestly. Demis Hassabis, then CEO of Google DeepMind, sat down with Scott Pelley on 60 Minutes in April 2025. On the benefit: “I think one day maybe we can cure all disease with the help of AI.” The end of disease, Pelley asked. Within reach, Hassabis replied, maybe within about a decade, he didn't see why not. In the same interview he named two worries. One is that users repurpose the systems for harmful ends. The other is the systems themselves as they become more autonomous and more powerful: whether we can keep control, whether they are aligned with our values and stay within the guardrails. Is the race for AI dominance a race to the bottom on safety, Pelley asked. One of his biggest worries, he answered. The resources and the pace are good for progress, but they can lead actors to cut corners, and safety and responsibility are one of the corners that can be cut.
That is unusual. He does not shy away from either end: civilizational benefit against civilizational risk, both at once. But that axis has nothing to say to the neighbor wondering what the electricity bill will be. The neighbor's question does not sit between benefit and risk. It sits between whose benefit and whose cost. That question has its advocates, but they sit in town halls, and no one carries the argument up to the level where he speaks. No one is being dishonest here. He answers his question well. The other one has an advocate in Sirdal.
Sirdal's mayor, Jonny Liland, said yes to a data center. In an interview with E24 he describes the situation: the site sits next to Statnett's transformer station, where the Nordlink cable runs to Germany. He argues that Sirdal was a poor farming municipality and has become one of Norway's richest on the power it produces, that around 95 percent of it is sent out, and that some of the value ought to stay. The zoning plan was changed in a simplified process without an impact assessment. Statsforvalteren i Agder calls the decision legally questionable and has halted the work while the appeals are processed, so the case is unresolved. The argument about whose benefit comes from the man who said yes, not from those protesting.
So it is not one half of the conversation that is missing. It is two. Coverage swings between hype about what the models will be able to do and alarm about what they might cause. The middle is barely reported: what AI has actually done, with names and results. The consequence is local. A municipality hears the developer say jobs and the opponents say electricity and water. No one says what the computing power will be used for. Then every hall looks like pure extraction, something that takes from the community and gives somewhere else.
The middle exists. The 2024 Nobel Prize in Chemistry went half to David Baker for computational protein design and half jointly to Demis Hassabis and John M. Jumper “for protein structure prediction.” AlphaFold's open protein database opened on July 22, 2021 with just over 360,000 structure predictions, including the entire human proteome. It has grown to more than 200 million structure predictions, and EMBL-EBI reports more than 3.4 million users in 190 countries. And one result from it: in 2024, an AlphaFold search pointed to the protein Tmem81 as part of the complex on the sperm surface that binds to the egg, together with Izumo1 and Spaca6, and experiments in zebrafish and mice showed the protein to be necessary for male fertility. Published in Cell in December 2024. That is not a promise. It is a prize, a database and a single result, and none of it appears in the material a local politician gets on the table.
Data centers come to the Nordic countries for the power and the cold, and it shows in the queue. Svenska kraftnät's queue table for September 2, 2026 showed data center applications totaling 6,984 megawatts in the transmission grid, of which 5,169 megawatts in bidding zone SE3. The queue is not a forecast, Svenska kraftnät points out. But it is the queue that creates the pressure.
Norway has required data center operators to register since January 1, 2025, with Nkom as the supervisory authority. In Stortinget, a proposal for a licensing requirement was voted down in March 2026, but Stortinget at the same time asked the government to examine the societal benefit of data center establishments. In March 2026, Energinet in Denmark paused new connection agreements for large loads — data centers, batteries and Power-to-X — after requests reached around 60 gigawatts against a peak load of about seven. Sweden abolished the reduced energy tax for data centers on July 1, 2023, after Riksrevisionen in 2022 found clear shortcomings in the decision basis and that no evaluation had been done. We have already made Virginia's journey, without a debate.
Hässleholm is the case. Everything that follows comes from the municipality's own questions-and-answers page, last updated November 2, 2026. The municipal executive board approved the letter of intent with Panattoni on June 10, 2026. It is not legally binding. It covers 300 megawatts and roughly 20 hectares in Hässleholm Nord, an area of roughly 200 hectares. The municipality itself writes what 300 megawatts means: roughly five to six times the current power level of the town of Hässleholm. The facility is described as hyperscale in two phases: E.ON has given an initial positive connection indication of 150 megawatts by around 2033, and Panattoni will therefore apply for a further 150 megawatts. Whether it will be one phase or two is not yet settled. Panattoni has started a grid study with E.ON that is to provide a cost indication and a preliminary timetable, estimated at between 6 and 18 months. After that, a feasibility study of between 3 and 9 months. The agreement was signed first. The cost indication arrives between six and eighteen months later, and the whole chain of studies is complete after nine months at the earliest, twenty-seven at the latest. On jobs, the municipality writes that the number cannot be stated, since it is not yet known which companies will use the facility, but that in operation it “can create a significant number of permanent jobs.” The exact activity will depend on future dialogue between Panattoni and interested tenants. The municipality, in other words, cannot itself say what the capacity will be used for. That is not an accusation; that is how a leasing model works. But then the question of benefit has no addressee either.
On the same page, E.ON answers that the electricity price is set day by day on the power exchange, driven by weather, temperature and available generation, and that a single large consumer of this kind does not affect it. Asked whether a new large connection affects other companies' ability to establish themselves or grow, E.ON answers that the entire grid is affected, that reinforcements must be procured and carried out, and that this affects when new customers can get their connection. There is the answer to the local question. The cost does not come as a higher electricity bill. It comes as someone else having to wait.
How the connection and tariff costs are divided between the data center and other customers cannot be read from open sources. A municipality that says yes in June and gets the number much later has not been deceived. It decides without the number. I have been in enough siting processes to know that the order is rarely the one that looks reasonable in hindsight: the letter of intent first, the land next, the question of who pays for the power up to the property line last. That is neither corruption nor stupidity. Whoever is in a hurry gains the most from the agreement being signed before the numbers exist.
Five questions should be asked before the decision. Is the connection guaranteed or merely indicated? Which of the jobs are permanent, and who is the employer? Who pays for the grid reinforcement if phase two is never built? What does it cost to say yes to phase one if phase two never comes? And what will the computing power be used for?
This is a scenario with stated assumptions: model capability keeps rising, the cost per unit of capability falls, demand for computing power grows. The load-bearing assumption is that the volume of applications in itself forces a political response, regardless of how much gets built. Denmark is my only case: Energinet paused on application volume, not connected load. One case makes the mechanism possible, not probable. The tipping point is whether volume triggers decisions without connected load, in two stages: new Swedish, Norwegian or Danish parliamentary decisions on connection queues, taxes or licensing between 2027 and 2029, and municipal planning decisions for data centers reviewed or overturned by county governors and county administrative boards, counted by year. The cost claim is tested separately, against household grid fees per grid company and municipality in the Nils Holgersson report and at Energimarknadsinspektionen, against the general price trend. The text does not claim that data centers drive grid costs; Svenska kraftnät's roughly 215 billion kronor over the decade is electrification, offshore wind and reinvestment. In the counter-scenario, the rules change in time and the cost lands on the data centers themselves or on projects that never happen.
That is what the decision basis looks like. The cost indication is up to eighteen months away, the full chain of studies up to twenty-seven. The benefit depends on tenants who do not yet exist. Two columns, both empty. The decision has been made.
Source notes
Where I go beyond what can be documented, I mark it in the running text. Two items deserve a particular reservation: I have not read the minutes of Hässleholm's municipal executive board, § 126, only the municipality's own account of the decision, and the information on Sirdal rests on E24's interview and on Norwegian news reporting, not on the municipality's documents or the county governor's decision.
Sources. Virginia: Joint Legislative Audit and Review Commission, report on the data center industry, December 2024; the state's consumption tax on data centers' electricity use, in force July 1, 2026 · Good Jobs First, Money Lost to the Cloud, 2016 · Gallup, survey March 2–18, 2026 · Fayette County, Georgia, on unmetered water withdrawal, 2026 · Demis Hassabis on 60 Minutes, CBS, Scott Pelley, first aired April 20, 2025 · The Royal Swedish Academy of Sciences, the Nobel Prize in Chemistry 2024, announced October 9, 2024 · EMBL-EBI, AlphaFold Protein Structure Database · Deneke et al., “A conserved fertilization complex bridges sperm and egg in vertebrates”, Cell, December 2024 · Svenska kraftnät, transmission grid queue table as of September 2, 2026, and investment plan for the transmission grid · Nkom, registration requirement for data center operators from January 1, 2025 · Energinet, pause for new connection agreements, March 2026 · Riksrevisionen, Statliga insatser för att stimulera investeringar i datacenter, RiR 2022:18 · Hässleholm Municipality, questions and answers on the planned data center at Hässleholm Nord, updated November 2, 2026, including E.ON's answers · Ada Lea, “Datasenteropprør i Sirdal: – Mange søvnløse netter”, E24, November 7, 2026
My own reasoning. That Virginia's consumption tax is the result of a recalculation is my conclusion, not the state's stated reasoning. That the two conversations sit on different axes — benefit against risk, and whose benefit against whose cost — is my description, not something any of those quoted says. That the cost shows up as someone else having to wait rather than as a higher electricity bill is my reading of E.ON's two answers placed side by side. The five questions are mine, drawn from my own experience of industrial siting processes, and they are not taken from any source.