I help organisations adopt AI, so I’m biased: I want more AI inside companies, not less.
Wider AI adoption depends on cheaper access to models, tokens, and tools, and cheaper access depends on more compute capacity.
More compute capacity means more data centers, not fewer.
From the UK, I’ve been reading the American data centre fight from the outside in. I see staggering figures, local anger, and public backlash. It doesn’t seem to have stopped construction much.
There’s also the opposing argument that splurging on these contentious behemoths might bring not just material benefits to society (the way we argue that nuclear power plants are good for society), but might in its own way herd us further along the Kardashev scale.
In any case, it appears that the relentless pace of technology won’t stop any time soon. So the practical question has moved past whether these facilities get built, to what a good project should look like, and what conditions would let people on both sides of the political divide accept it.
I don’t believe it has to be a zero-sum problem, and the public record already points to most of the answers.
The scale of the data center build-out
I’ll focus mainly on the US data centre market for this thought exercise, since it’s the largest data centre market worldwide, for which we have the most data, and which is experiencing the most resistance. The conclusions are applicable to any other country considering its own data centre rollouts.
The American data center market saw close to $300bn in investment in 2025, and McKinsey predicts global demand for data center capacity could almost triple by 2030, with roughly 70% of that demand coming from AI workloads.
Cleanview’s data center tracker lists 1,661 US data centre as of writing, with 950 more on the way. The US already holds about 45% of the world’s data centre servers (that’s 2024 data—likely much higher by now)—about 8 times that of the next country on the list, Germany.
Data centers are power hungry: current US operating capacity is 23 GW, with 332 GW planned, more than 15 times today’s usage. US data centers consumed 183 terawatt-hours (TWh) of electricity in 2024, according to IEA estimates via Pew—more than 4% of total US electricity consumption, and roughly equivalent to the annual electricity demand of Pakistan.
That number is expected to rise. The Department of Energy reports that data centre load has tripled over the past decade and could double or triple again by 2028. Goldman Sachs predicts a 165% rise in data centre power consumption through 2030.
This is private capital, committed at industrial scale, and is the backdrop to every local dispute.
Where the data center friction comes from
The opposition to data centers crosses party lines, and it rests on legitimate concerns. Four costs come up most often: water, electricity, jobs, and tax.
Water
US data centers consumed about 17 billion gallons of water directly through cooling in 2023, a figure that could double or quadruple by 2028, and in some places this reaches over 25% of the local water supply.
In Texas, one analysis found data centres would use 49 billion gallons in 2025 and as much as 399 billion gallons by 2030, enough to draw down Lake Mead by more than 16 feet in a year.
About 80% of the water a data centre draws evaporates, according to a June 2025 report from American Water. The American Society of Civil Engineers concluded that the water that returns carries higher levels of calcium, chloride, and silica, which can affect drinking-water taste, lower crop yields, and harm aquatic life.
Electricity and the grid
A single hyperscale facility can use as much electricity as 100,000 homes a year, according to the International Energy Agency. Gartner projects AI-optimised servers will draw 44% of all data center power by 2030 and account for 64% of new power demand.
The pressure’s already reaching household bills: in the PJM grid region, which serves 65 million people across 13 states, power supply costs jumped from $2.2bn to $14.7bn in a single year, with data centers behind nearly two-thirds of the rise, according to an analysis by the Brookings Institution. The S&P reported that residential electricity rates nationally rose about 32% between July 2020 and July 2025.
Jobs
The employment numbers are small next to the investment. By the end of 2024, as few as 23,000 people worked in the US data centre industry, and research published in late 2025 found no clear evidence that these facilities grow local tech employment.
In Virginia, data centres created one permanent job for every $54mn invested. One facility outside Reno estimated it would create 73 permanent jobs over a decade, against more than 4,000 temporary construction roles. The reason is automation: a hyperscale campus larger than 100 MW can run on as few as 20 to 30 staff.
Tax
Tax breaks present a visible public cost. Virginia has more data centers than anywhere in the world—its data center sales-tax exemption cost an estimated $1.6bn in fiscal year 2025.
In Illinois, $370mn to $465mn in tax breaks equalled the federal pandemic relief the state sent to its 1,200 smallest communities.
The reverse is also on record: in Loudoun County, Virginia, officials expected data centre tax revenue to approach $900mn in fiscal year 2025, close to the county’s entire operating budget, which shows how large the contribution can be when revenue isn’t abated to near zero.
| Cost area | Key finding | Figure |
| Water | Direct cooling use in 2023, set to double or quadruple by 2028 | 17 billion gallons |
| Electricity | National rise in residential rates, 2020 to 2025 | 32% |
| Jobs | Permanent staff to run a hyperscale campus over 100 MW | 20 to 30 |
| Tax | Virginia sales-tax exemption cost, FY2025 | $1.6bn |
The political response to data centers
State lawmakers have noticed. In 2025, they introduced more than 200 bills on data center issues across more than 40 states, according to reporting by lobbying firm Multistate.
In the first six weeks of 2026, lawmakers filed more than 300 bills across 30-plus states, a move away from incentives and toward oversight.
New York, South Dakota, and Oklahoma introduced moratorium bills, and at least 11 states have moved to restrict or ban data center development since 2025, with Maine the first state to ban construction outright.
At least 18 states have introduced bills creating special rate classes for large energy users, some requiring data centers to fund infrastructure upgrades and show benefits to ratepayers.
In Virginia, an amendment to Senate Bill 253 would put more of the power cost on data centers and less on households. State regulators project it would lower a typical residential bill by about $5.52 a month, a 3.4% cut, while raising the data centre rate by 15.8%.
What the perfect AI data centre would look like
With all of this impact (largely negative), backlash (largely warranted), and political wrangling, it’s clear something must give. So: what would the perfect AI data center project look like?
The short answer is it must minimise its harmful effects, maximise its intended benefits, and centre communities.
The design answers already exist, scattered across working facilities, state bills, and corporate commitments. Pulled together, they describe a building that adds capacity without pushing its costs onto the people nearby.
Privately funded infrastructure
A shared objection from both parties is the corporate giveaway: tax breaks plus free grid upgrades paid for by ratepayers.
A well-designed facility pays for its own grid connection and power upgrades, so the public doesn’t subsidise a profitable business.
Several of the $1.6bn-scale exemptions may be subsidising investment that would’ve happened anyway.
Closed-loop or air cooling
Water use is an engineering choice, not a fixed cost. Google’s air-cooled facility in Pflugerville, Texas, used about 10,000 gallons in a year, about what one Texas home uses in two months, while its water-cooled sites used far more.
Where cooling needs water, reclaimed or recycled water keeps the facility off the drinking supply, which is the point of designing out the 80% that otherwise evaporates.
On-site clean generation
Rather than only drawing from a strained grid, a facility can bring its own power. Microsoft and Clearloop agreed in February 2025 to deploy 100 MW of solar across underinvested communities in states including Arkansas and Louisiana, a model for adding generation instead of only consuming it.
Distance from homes
Data centers are loud. Cooling systems, diesel generators, and server fans produce a hum you can hear hundreds of feet away.
A large server warehouse can push noise to 96 decibels, and sounds above 60 or so decibels raise stress and blood pressure.
In Virginia, 29% of data centers are within 200 feet of residentially zoned properties. Causing poor sleep shows up in lost productivity and more sick days, which eats into state and corporate coffers.
Siting data centres as industrial facilities, well away from homes, addresses that at the source.
Ratepayer protection
The power bill can’t be pushed onto households after the ribbon-cutting photos are done.
Virginia’s State Corporation Commission has approved a GS-5 rate class for customers demanding 25 MW or more, with the class starting on January 1, 2027.
Its data-center measures also say these customers must pay at least 85% of their transmission and distribution costs, regardless of use.
If a facility needs the grid rebuilt around it, the facility should pay for it.
Real tax contribution
Loudoun County’s data-center tax base shows what a fair local deal can fund when the county doesn’t give the money away through subsidies and abatements.
For FY 2025, Loudoun projected roughly $900 million in combined data centre real estate and business personal property tax revenue, while the adopted budget sent $1.3 billion in local funding to public schools.
The county also says data-center revenue helps fund libraries, parks, recreation, community centers, child protection, mental health, and capital projects such as schools, roads, and bridges.
Which is the whole point: if the community hosts the burden, the community should see the return.
Local AI literacy and reinvestment
Since these buildings are highly automated and produce few permanent jobs, the local return has to come from somewhere other than headcount.
A facility that funds retraining and AI-literacy programmes gives a community a stake in the technology the building runs, rather than just the building itself.
| Community concern | Design response | Anchor |
| Subsidised by ratepayers | Privately funded grid connection and upgrades | $1.6bn in exemptions |
| Water draw on local supply | Air or closed-loop cooling; reclaimed water | 10,000 gallons at air-cooled site |
| Strained grid, higher bills | On-site clean generation; large-load rate class | 100 MW solar; 18 states |
| Few permanent jobs | Funded retraining and AI-literacy programmes | 20 to 30 staff per 100 MW |
| Lost public revenue | Full tax contribution, no deep abatement | $900mn in Loudoun County |
The record shows measurable costs, with several of them having known engineering and policy answers some states and facilities are already testing.
The legislatures pushing through those 300+ bills are setting the terms. A data center built to meet those terms will meet less of a resistance now and in the future.
