Irish data centres use 23% of the nation’s electricity, more than all urban homes, yet a typical site employs about 50. What a server farm costs a city.
The Server Farm Next Door
What a data centre takes from a city centre, and what it gives back
In January 2026 a 14-storey, 7.4 MW data centre began operating about one kilometre from Osaka Station ✓ Established [1]. It is a small building by industry standards. Across Dublin, Frankfurt, Amsterdam and Northern Virginia, buildings like it now draw between a fifth and two fifths of local electricity, employ a few dozen people each, and have pushed four governments into pauses, caps or rewritten zoning codes [2] [3] [4] [5].
The short answer to what a server farm does to a city centre is that it converts the scarcest urban resource of the 2020s, spare grid capacity, into a digital service sold far beyond the city, while leaving most of the local costs in the neighbourhood and very few of the jobs. Ireland is the clearest measured case. Data centres used 23% of the country’s metered electricity in 2025, against 5% in 2015, and in 2024 they already used more than every urban household in Ireland combined ✓ Established [2] [6]. In Frankfurt they account for about 40% of the city’s power demand ◈ Strong Evidence [3]. In Loudoun County, Virginia, they supply 31% of local tax revenue, a fiscal dependence no other industry matches ✓ Established [7]. Yet Virginia’s own legislative auditor found that a typical 250,000 square foot facility employs about 50 full-time workers, roughly half of them contractors [7]. The pattern is consistent across jurisdictions. A data centre is a power plant in reverse, an industrial load the size of a town, housed in a building that looks like an office and is often permitted as one [8].
Osaka is a useful place to test that pattern because it sits at the early end of it. The queue item that prompted this report described a data centre under construction in Osaka’s central districts. That description is out of date and imprecise. The clearest downtown case, Optage’s Sonezaki data centre known as OC1, is already finished: the building in Kita-ku’s Higashi-Umeda district was completed on 30 September 2025 and opened for customers on 29 January 2026 ✓ Established [9] [1]. It has 14 floors above ground, seven and a half of them server rooms, space for up to 1,200 racks and a ceiling of 7.4 MW of power [1]. Its developer was a joint vehicle of Kansai Television and The Sankei Building, on a plot of roughly 2,240 square metres, which is the footprint of a mid-sized office tower rather than a campus [9]. The only announced future data centre in central Osaka is NTT West’s 8 MW Osaka Minami site in Chuo-ku, due in fiscal 2031; no evidence was found that construction has begun ✓ Established [10].
The large builds now under way inside Osaka’s city limits are on the bay front, not downtown. SC Zeus broke ground in December 2025 on a site in Nanko Kita, Suminoe-ku, that will take 100 MW from Kansai Electric and run 70 MW of computing, with phase one due in 2028 ✓ Established [11] [12]. ESR’s Cosmosquare campus nearby is planned for up to 98 MW across three buildings [13]. Further out, NTT Global Data Centers is building a 36 MW campus in Ibaraki, 20 km north of the city, and operators have filled the former Sharp plant in Sakai and the Keihanna science city on the Kyoto border [14]. The Osaka market is therefore splitting into two forms that the rest of the world already knows well: small, dense, vertical buildings close to the fibre exchanges of Umeda, Dojima and Shinsaibashi, and 100 MW sheds on industrial land at the edge [1] [15]. Each form imposes a different set of costs on the people around it.
Japan’s concentration makes the question urgent. The Ministry of Internal Affairs and Communications states that about 90% of the country’s data centres are concentrated in urban areas such as the Tokyo and Osaka regions, and that this is expected to continue ✓ Established [16]. JLL estimates that Greater Osaka’s development pipeline will grow by 170% in critical capacity, against 80% for Greater Tokyo, largely because power can be delivered in Osaka in three to five years while Tokyo’s planned supply is allocated into the 2030s and new connections take eight to ten years ◈ Strong Evidence [15]. Japan’s grid coordinator, OCCTO, forecasts that new and expanded data centres will add about 6.6 GW of national peak demand by fiscal 2035 [17]. A government liaison conference on data-centre siting, convened by the Cabinet Secretariat for the first time on 10 September 2026, put the increase at about 6 GW over fiscal 2026 and listed the concerns residents raise: waste heat, noise, traffic, landscape, the fire risk of stored heavy oil, water use and smoke from emergency generators ✓ Established [8].
None of those concerns has yet produced a documented planning fight inside the city of Osaka. The conflicts recorded so far in Japan sit around Tokyo, in Akishima, Hino, Kodaira, Inzai, Shiroi and Nagareyama, where residents have petitioned, sued and in one case forced a developer to cut building heights by a fifth [18] [19] [20]. The one Osaka Prefecture case found is a residents’ petition of August 2026 asking that data centres be made subject to environmental impact assessment, which they currently are not [21]. That gap between a fast-growing market and an untested local rulebook is the position Dublin, Amsterdam and Loudoun County were in a decade ago. What follows sets out what their experience shows a downtown data centre actually draws, emits and returns, and which responses changed the outcome. The comparison is not exact, since Japanese grids, land prices and zoning differ, but the mechanics of load, noise, fuel and heat are the same in every city that hosts these buildings, and so are the questions regulators eventually had to answer [7] [8].
Why Operators Want the City
Latency, fibre and, above all, a grid connection that already exists
The industry’s public case for urban sites rests on speed of light. The binding constraint is now speed of power: JLL’s 2026 outlook names it the primary site-selection criterion, ahead of community support and latency ✓ Established [22]. City centres win because their substations and fibre are already built.
Physics sets the floor of the latency argument. Light in standard optical fibre adds roughly 5 microseconds of one-way delay per kilometre, or about one millisecond of round-trip time for every 100 km ✓ Established [23]. For most consumer services, from video streaming to email, a few milliseconds are invisible. They matter for a narrower set of workloads: synchronous replication between data halls, which engineers typically restrict to metro or regional distances, financial trading, and interactive services that chain many calls between machines [23]. The rise of AI inference has widened that set. JLL argues that inference demand requires geographic distribution to reduce latency and serve users, and CBRE reports that inference workloads often need to be near end users, which is reshaping site strategy ◈ Strong Evidence [22] [24]. That is the strongest version of the case for downtown sites. It explains why operators want to be in or near a metropolitan area. It does not, on its own, explain why they want to be in the centre of one.
The fuller explanation is interconnection. Data centres are valuable partly because of the other data centres and networks next to them. Optage markets OC1 on its position within three kilometres of the Dojima and Shinsaibashi areas, where the major cloud providers’ access points sit, and on its connections to several internet exchanges [1]. NTT West’s planned Chuo-ku site will be linked by direct fibre to the existing Dojima and Sonezaki facilities [10]. In Amsterdam, the director of the Dutch Data Center Association told Data Center Knowledge in 2020 that the most desirable location for a new site is next to the existing ones, and that otherwise the industry looks within a 30 km circle [25]. Northern Virginia, which holds 13% of all reported operational data-centre capacity in the world, grew from the same logic: a strong fibre network, cheap reliable power, available land, proximity to customers and, from 2010, a state sales-tax exemption ✓ Established [7]. Clustering is self-reinforcing. Each new hall makes the location more valuable for the next.
JLL’s 2026 Global Data Center Outlook ranks speed to power first among site-selection criteria, followed by community support, latency and proximity to customers, and reports that the average wait for a grid connection in primary markets exceeds four years [22]. CBRE found primary-market vacancy in North America at a record low of 1.4% at the end of 2025, with 5,994 MW under construction [24]. In Japan, JLL attributes Osaka’s faster pipeline growth to a power-delivery time of three to five years, against eight to ten in Tokyo [15].
That is why the grid has become the real prize. A plot in a city centre that already has a high-voltage connection, or sits beside a substation with headroom, is worth more to a data-centre developer than almost any other attribute. The Ministry of Internal Affairs and Communications notes that data centres tend to cluster locally and that a new site can require large grid works such as new or expanded substations ✓ Established [16]. In Frankfurt, the grid operator Mainova receives five to ten qualified data-centre connection requests a year, typically for 50 to 100 MW each, and has said that new high-performance connections can only be expected again in the mid-2030s [3]. Operators therefore buy land banks near substations and reserve capacity years ahead. Japan’s Agency for Natural Resources and Energy now names this practice, capacity reserved but not used, as one cause of long connection queues, and from October 2026 grid operators can cancel stalled applications ✓ Established [8].
- Latency: about 1 ms of round-trip delay per 100 km of fibre, decisive for replication, trading and some AI inference, marginal for most consumer traffic [23] [22].
- Interconnection: proximity to exchanges and cloud access points, the reason OC1 is sited within 3 km of Dojima and Shinsaibashi [1].
- Grid headroom: an existing connection or substation capacity, now the first criterion in site selection [22].
- Disaster backup: after the 2011 earthquake, firms began using Osaka to back up Tokyo, a role the government still cites [26] [16].
- Land and zoning: Japan classes data centres as offices or warehouses, so they are allowed outright in commercial and industrial zones [8].
Osaka adds a fifth motive that Dublin and Frankfurt lack: disaster resilience. After the Great East Japan Earthquake of March 2011, firms began using Osaka to back up Tokyo data centres, and in 2014 Osaka already held about 26% of Japan’s data-centre demand against 68% in Tokyo ◈ Strong Evidence [26]. The government still frames concentration as a national risk. Its 2026 white paper warns that a major earthquake hitting the Tokyo or Osaka regions could disrupt communications nationwide, and argues for dispersed siting and new submarine cables on the Sea of Japan coast [16]. The irony is that the resilience argument favours building in Osaka rather than Tokyo, but not necessarily in central Osaka. A bay-front campus in Suminoe-ku provides the same geographic diversity as a tower in Umeda, at lower land cost and with more room for cooling, generators and fuel [12] [13]. The government’s own fund for regional data centres, running since fiscal 2021, points the same way: resilience is served by distance from Tokyo, not by proximity to Osaka Station [16].
One hundred kilometres of fibre costs about one millisecond of round-trip time [23]. For the large majority of traffic, a site at the edge of a metropolitan region performs the same as one in its centre. What a downtown plot really offers is an existing grid connection and adjacency to exchanges, which are commercial advantages to the operator, not technical necessities for the user. Planning authorities that accept latency as a reason for central siting are often accepting a land-and-power argument under another name.
Japanese zoning makes the downtown option easy. Under the Building Standards Act, data centres are classified as offices or warehouses according to how they are used, the Ministry of Land, Infrastructure, Transport and Tourism told the September 2026 liaison conference ✓ Established [8]. Large facilities are therefore restricted in exclusive residential zones but allowed outright in neighbourhood-commercial, commercial, quasi-industrial and industrial zones, and in some residential zones up to size limits [8]. There is no separate land-use category, no environmental assessment trigger and no special hearing. SC Zeus’s 100 MW site, for example, sits in a quasi-industrial zone where it needs no special consent [11]. In Inzai, east of Tokyo, ten residents sued in March 2026 over precisely this point, arguing that a 52 m data centre should not have been certified as an ordinary structure, and their lawyer told AFP that Japanese law is not catching up with the situation [27] [19]. Whatever the merits of the case, it shows where the pressure will fall as Osaka’s pipeline grows.
A Town-Sized Load in an Office Block
What an urban data centre draws from the grid and the water supply
The International Energy Agency estimates that a typical AI-focused data centre consumes as much electricity as 100,000 households ◈ Strong Evidence [28]. Where such loads cluster, they compete directly with housing for the same substations, and in Dublin, west London and Amsterdam housing has already lost time [29] [30].
Ireland offers the cleanest data on what concentration does. Its Central Statistics Office has measured data-centre electricity directly since 2015. Consumption rose from 6,335 GWh in 2023 to 6,969 GWh in 2024, which was 22% of all metered electricity, while urban households used 18% and rural households 10% ✓ Established [6]. In 2025 data centres used 7,663 GWh, or 23%, against 28% for every household in the country combined ✓ Established [2]. Quarterly consumption rose 584% between the first quarter of 2015 and the last quarter of 2025, and more than tripled between 2019 and 2025 [2]. In 2022 Ireland had 70 operational data centres, 65 of them in the Dublin area [31]. The Prime Time programme on RTÉ reported in August 2025 that a single data centre can use as much electricity as the city of Kilkenny [32]. Concentration on that scale is not a national statistic diluted across a country. It is a load sitting on a handful of transmission corridors around one city.
The Central Statistics Office measured data-centre consumption at 22% of metered electricity in 2024, against 18% for urban households and 10% for rural households [6]. The share reached 23% in 2025, up from 5% in 2015, and consumption has risen every year since measurement began [2]. EirGrid responded in late 2021 by signalling that no new data-centre connections would be offered in the Dublin region until at least 2028 [31].
Frankfurt shows the same arithmetic in a single city. As early as 2017, data centres in Frankfurt consumed 1,037 GWh, more than all of the city’s households combined ✓ Established [33]. AlgorithmWatch reported in November 2025 that data centres now account for about 40% of Frankfurt’s power demand, and that the grid operators Mainova and Syna say no new data centre above 10 MW can connect before 2030 ◈ Strong Evidence [3]. In the Netherlands, Statistics Netherlands counted 5,100 GWh of data-centre consumption in 2024, 4.6% of national electricity and about as much as two million households, with most of the country’s roughly 200 sites located around Amsterdam [34]. Inside Amsterdam the share is higher still. Three 85 m towers that Microsoft is building in the city’s western port area will draw, according to Binnenlands Bestuur, as much power as every household in Haarlem combined [30]. A single campus can therefore carry the electricity demand of a mid-sized city while occupying a few hectares of port land, which is why grid operators, not planners, were the first to raise the alarm in all three countries [31] [3] [4].
The cost of that load is paid partly in housing. In July 2022 the Greater London Authority warned that new homes in west London could be held up because the grid in Hillingdon, Hounslow and Ealing had reached capacity, in part because of data-centre connections [29]. The London Assembly’s Planning and Regeneration Committee reported in December 2025 that housing schemes were temporarily halted, that more than 12,000 homes were eventually connected through short-term fixes, and that data-centre demand in London is forecast to grow by 200% to 600% ✓ Established [29]. Its chair, James Small-Edwards, said London was at a critical moment, with energy capacity becoming a real constraint on both housing delivery and economic growth [29]. In Amsterdam, a council letter of December 2025 recorded 591 Liander customers waiting for a connection and put the construction of 30,000 homes at risk from congestion expected to last at least ten years [30]. Industry rejects the causal claim: Kao Data argued in 2023 that network operators released more than 200 MVA of unused capacity once they looked, and that the housing shortage pre-dated data centres ⚖ Contested [35].
Water is the second draw, and its urban profile is different from the rural one. Lawrence Berkeley National Laboratory estimates that US data centres used 66 billion litres of water directly in 2023, 84% of it at hyperscale and colocation sites, and that the indirect footprint through electricity generation is far larger ✓ Established [36]. Averages hide the variation. Virginia’s legislative auditor found that most data centres use about as much water as a large office building, around 6.7 million gallons a year, but that in 2023 eleven buildings each used more than 50 million gallons and one used 243 million, a tenth of the industry’s total ✓ Established [7]. Where water is scarce, the conflict is direct. Microsoft’s campus in Hollands Kroon used 84 million litres of drinking water in the 2021 drought year [37]. In Chile, an environmental court partially reversed the permit for Google’s Cerrillos project, which would have used 7.6 million litres of potable water a day for cooling [38].
Dense urban sites tend to be air-cooled or use closed loops, partly because water rights and cooling towers are hard to secure in a city centre, so their water footprint is usually smaller and their electricity footprint larger. The new generation of AI halls changes that balance. SC Zeus designs its Osaka site for up to 130 kW per rack with direct-to-chip liquid cooling, and NTT West plans liquid cooling on some floors of its Chuo-ku building [11] [10]. Liquid cooling is more efficient than air at those densities, but it concentrates heat rejection on the roof or at the plot boundary. The Japanese liaison conference listed water consumption alongside waste heat among the concerns residents raise, and Japan has no published data on how much water its urban data centres use [8]. Osaka is therefore starting the AI build-out without the baseline measurement that Ireland and the Netherlands now publish every year [2] [34].
The national forecasts make the scale clear. OCCTO’s fiscal 2026 outlook expects data-centre peak demand to rise by about 6.6 GW by fiscal 2035 and data-centre consumption to reach about 49 TWh, against national peak demand of 159.6 GW in fiscal 2026 ◈ Strong Evidence [17]. The regional split for the Kansai area was not published in the material reviewed for this report. What is known is that data centres tend to arrive in clusters and that each cluster can require new or expanded substations [16]. In Amsterdam, stopping new large data centres in June 2025 allowed Liander to avoid building three substations, a saving of about 125 million euros, and TenneT to build two fewer stations ◈ Strong Evidence [4]. The cost of hosting a data-centre cluster is therefore not only the energy it uses but the grid capital it absorbs, which is paid for by every connected customer unless regulators decide otherwise [39]. Japan’s grid reforms acknowledge part of the problem: the government now names capacity reserved but unused as one cause of long connection queues [8].
Hum, Smoke and Heavy Oil
The neighbourhood impacts that permits were not written to measure
Virginia’s legislative auditor found that 29% of operational data-centre properties in the state are within 200 feet of residentially zoned land ✓ Established [7]. The complaints that follow are less about loudness than about constancy, and less about daily emissions than about thousands of diesel engines held in reserve [7] [40].
Noise is the most persistent complaint and the hardest to regulate. The Joint Legislative Audit and Review Commission estimated that the data-centre noise that has prompted complaints in Virginia ranges from 40 to 59 decibels, typically below the 55 or 60 decibel limits that Loudoun, Prince William and Fairfax counties allow ✓ Established [7]. Its conclusion was that the problem is not volume but constancy: residents describe a drone or hum that never stops [7]. Loudoun County acknowledges that its A-weighted decibel standard has limitations in capturing low-frequency and tonal noise, calls noise complaints among the most persistent issues brought to county officials, and expects amendments on tonal and low-frequency noise in 2027 ✓ Established [41]. The Environmental and Energy Study Institute reports that cooling towers produce about 70 dBA within 400 feet and that industrial diesel generators reach about 105 dB [42]. A planning regime built around a single decibel ceiling misses most of what neighbours actually hear.
The industrial scale of data centers makes them largely incompatible with residential uses.
— Joint Legislative Audit and Review Commission, Data Centers in Virginia, December 2024 [7]The case of Great Oak, a 291-home subdivision in Prince William County, shows both the harm and the fix. Residents measured noise from a neighbouring Amazon facility at up to 65 dB at night, against a county night limit of 55 dB ◈ Strong Evidence [43]. After a sustained campaign, Amazon replaced 424 rooftop exhaust fans with taller units and slowed the fan controls, work that took four months and two large cranes, and the level fell to about 50 dB [43]. The lesson cuts both ways. The noise was engineerable all along, which means it was a design choice rather than an unavoidable cost. Chandler, Arizona, drew that conclusion in December 2022, when its council unanimously restricted data centres to planned-development zones and required an annual noise study at peak operation for five years after construction [44]. Fairfax County followed in September 2024 with minimum distances of 200 feet between buildings and residential districts and 300 feet for generators, plus noise studies before and after construction [45].
Backup generators are the second hazard, and the least visible. Each data centre keeps enough on-site generation to run without the grid. JLARC found a median of 35 trailer-sized generators per Virginia site, typically one to three megawatts each, and about 8,000 permitted generators statewide ✓ Established [7]. Virginia Mercury reported in January 2026 that Northern Virginia alone has more than 4,000 diesel generators with over 11 GW of capacity, and that one Amazon permit allows up to 10 million gallons of diesel a year for 173 generators [40]. JLARC found that since 2015 nitrogen oxide emissions from data-centre diesel generators have more than doubled, carbon monoxide has tripled and particulate matter has grown fivefold, although they remain below 4% of regional nitrogen oxides [7]. The disagreement is over running hours. When Virginia’s environmental regulator proposed in late 2025 to let older-standard generators run during planned grid outages lasting days or weeks, the Piedmont Environmental Council objected and the proposal drew more than 500 comments, overwhelmingly opposed [46] [40].
Researchers at UC Riverside and Caltech estimate that backup generators in Northern Virginia already impose 190 to 260 million dollars a year in public-health costs on the region, rising to 1.9 to 2.6 billion dollars a year if they emitted at their permitted maximums [47]. Nationally they project up to 1,300 premature deaths a year by 2030 from data-centre air pollution. Shaolei Ren, an author of the study, noted that corporate sustainability reports focus on carbon and sometimes water, with no mention of unhealthful air pollutants [47]. The people who bear that cost live beside the generators, not where the computing is consumed.
Japan treats these hazards lightly by comparison. Under the Noise Regulation Act, data-centre blowers with motors of 7.5 kW or more are specified facilities only within designated areas, and while generators count as soot-emitting facilities under the Air Pollution Control Act, emergency units are exempt from the emission standards, the Ministry of the Environment told the September 2026 liaison conference ✓ Established [8]. The Japan Data Centre Council’s coexistence guideline of May 2026 asks operators to contact residents at the siting stage and lists the fire risk of stored heavy oil and smoke from emergency generators among the concerns to address [8]. Those concerns are concrete. A planned data centre in Kodaira, western Tokyo, on a former bread-factory site of 21,176 square metres next to homes and near a daycare, would store 1.1 million litres of heavy oil underground [48]. In Hino, residents told AFP that their greatest fear was a fire in the large battery banks that AI-era buildings now carry [19].
Japanese regulators are moving in the opposite direction on batteries. The Fire and Disaster Management Agency told the liaison conference that next-generation AI data centres are expected to hold large lithium-ion batteries in addition to underground oil tanks, and that rules may be eased under the July 2026 regulatory reform plan, for example by aligning with international battery fire tests and accepting sprinkler suppression ◈ Strong Evidence [8]. Virginia moved the other way on fuel: its 2026 General Assembly required generators meeting the strictest federal emission standard for air permits submitted after July 2026, a change Loudoun County confirms [41]. The contrast matters for Osaka. A 14-storey tower in Umeda stores its fuel and batteries in the middle of a commercial district, a bay-front shed in Suminoe-ku stores them on industrial land, and Japanese rules currently treat the two in much the same way [1] [12] [8]. The national coexistence guideline asks operators to explain these risks to residents, but it sets no binding distance between fuel stores and homes of the kind Fairfax now requires for generators [8] [45].
Landscape and trees are the last category, and in Japan they have proved the most mobilising. In Akishima, more than 220 residents petitioned in 2024 for an audit of the planning procedure for a GLP logistics and data-centre complex that would fell about 3,000 of 4,800 trees in a town that relies on groundwater [18]. In Nagareyama, residents stopped a data-centre plan in December 2023 [18]. In Hino, Mitsui Fudosan cut two of three buildings by a fifth, to 63.5 m, and added a green buffer up to 78 m deep after opposition ✓ Established [19]. Each of these cases began with a building whose bulk, plant and fuel were out of scale with its neighbours and whose permits had been granted under rules written for offices and warehouses [8]. JLARC’s residents used the same language in Virginia: one described a data centre as a giant monolith in the wrong place [7].
What the City Gets Back
Heat, jobs and tax revenue, measured against the promises
A finished data centre returns three things to its host city: waste heat, a small permanent workforce and property tax. Only the last has been large anywhere. Loudoun County draws 31% of its local tax revenue from data centres ✓ Established [7], while most heat-reuse schemes depend on a district-heating network that few cities have.
Heat is the benefit most often promised and least often delivered. Almost all the electricity a data centre consumes leaves the building as low-grade heat, typically 25 to 40 degrees Celsius. Where a city already runs a district-heating network, that heat can be lifted with heat pumps and sold. The largest scheme is in the Helsinki region, where Fortum is building a heat-pump plant at Microsoft’s Kirkkonummi data-centre site that is expected to supply about 40% of the district heat for Espoo, Kauniainen and Kirkkonummi and avoid about 400,000 tonnes of carbon dioxide a year ✓ Established [49]. In Odense, Meta’s data centre recovers about 100,000 MWh a year through ammonia heat pumps into the existing municipal network, enough for about 7,000 homes according to the equipment supplier [50]. In Saint-Denis, north of Paris, Equinix’s PA10 recovers heat at 28 degrees, lifts it to 65 degrees and supplies about 10,000 MWh a year free of charge for 15 years to the Olympic Aquatic Centre and a new development zone [51].
Stockholm has made heat recovery a city policy. Stockholm Data Parks, a partnership of the City of Stockholm, the heat utility Stockholm Exergi, the grid company Ellevio and the fibre network Stokab, aims to meet 10% of the city’s heating demand from data centres by 2035 ◈ Strong Evidence [52]. These schemes share three preconditions: an existing heat network close to the data centre, a utility willing to buy, and heat pumps paid for by someone. Where any one is missing, the heat goes to the sky. Frankfurt’s planners have found that waste heat of 30 to 40 degrees is too cool for their district network and that boosting it is costly at German electricity prices, and the city’s 2022 master plan committed only to a city-wide heat concept [53]. German researchers estimate that the country’s data centres currently reuse about 2.3% of their waste heat ◈ Strong Evidence [54]. Where no network exists, the heat cannot be sold at all, which is why the German government chose to exempt such sites rather than enforce the mandate against them [55].
Every large scheme identified for this report, Espoo, Odense, Stockholm and Saint-Denis, connects to an existing heat network through utility-owned heat pumps [49] [50] [52] [51]. Germany wrote a mandate into law in 2023, requiring new data centres to reuse at least 10% of their energy, rising to 20% for those opening from July 2028 [56]. In June 2026 the federal cabinet approved an amendment that counts on-site heat use and exempts operators where no technically and economically reasonable connection exists [55].
Germany’s retreat is instructive because it was the most ambitious attempt to force the issue. The Energy Efficiency Act of 2023 set a power-usage-effectiveness ceiling of 1.2 for new data centres operating from July 2026 and required existing ones to reach 1.5 by 2027 and 1.3 by 2030 ✓ Established [56]. The cabinet’s amendment of 24 June 2026 loosened the caps for existing sites to 1.6 and 1.4, doubled the compliance period for new sites from two years to four, and added the heat-network exemption [55]. Environmental groups put the value of the heat given up at about 5.76 billion euros [54]. The industry’s case for the change was that a mandate cannot conjure a heat network where none exists. Both points are true, and together they explain why heat reuse is a benefit cities can plan for but not one they can demand from a single operator. Efficiency gains have also stalled: the Uptime Institute reported in 2025 that average power-usage effectiveness had shown little change for the sixth consecutive year ✓ Established [57].
Jobs are the weakest part of the local bargain. JLARC found that a typical 250,000 square foot data centre may have about 50 full-time workers, one employee per 5,000 square feet against one per 650 for some distribution centres, and that about half are contractors ✓ Established [7]. Construction is a different matter: about 1,500 workers can be on site at peak during the 12 to 18 months a building takes [7]. The commission credited the industry with 74,000 Virginia jobs and 9.1 billion dollars of state GDP, but noted that most of these economic benefits derive from the construction phase rather than ongoing operations [7]. Ireland’s Department of Finance reached a similar view in June 2026, observing that much of the capital in data-centre investment is likely to have been imported and that its direct impact on employment is expected to be relatively limited [58]. Google’s newly approved extension at Grange Castle in Dublin is expected to bring up to 50 jobs [59].
Tax incentives make the arithmetic worse for the public purse. Good Jobs First counts 36 US states that exempt data centres from sales and use tax, of which only 11 disclose the recipients, and estimates that taxpayers typically pay at least one million dollars for each permanent data-centre job created ◈ Strong Evidence [60]. Virginia’s exemption was worth 928.6 million dollars to the industry in fiscal 2023, by far the state’s largest economic development incentive [7], and rose to 1.6 billion dollars in fiscal 2025, up 118% on the year before [61]. A later state evaluation found that the exemption returns 6.1 dollars of labour income per dollar exempted, but that the state recovers only 48 cents in revenue for each dollar it gives up ⚖ Contested [62]. Whether that is a good deal depends on whether the data centres would have come anyway, which is exactly what incentive studies struggle to establish. Ireland’s Department of Finance doubts the direct employment case [58], while the Virginia administration argues that the industry generates billions of dollars in local revenue that fund education, public safety and other local services [61].
Property tax is the real prize for a host jurisdiction, and Loudoun is its showcase. JLARC found that data-centre revenue accounted for 31% of total local tax revenue in Loudoun and 7% in Prince William ✓ Established [7]. That revenue funds county services that residents would otherwise pay for, which is why the county tolerated growth for two decades. It also creates dependence, and Loudoun has since ended by-right data-centre development and, in September 2026, paused legislative applications altogether [63] [5]. In Japan, local fixed-asset tax applies to data-centre buildings and equipment, but no ward-level figure for Osaka was found in public sources. For a 7.4 MW tower on a 2,240 square metre plot in Umeda the fiscal gain is real but modest; the larger question is whether the grid capacity it occupies could have served something else [9] [15]. Loudoun’s experience adds a warning about timing: a later pause cannot reach projects already in the pipeline, and 17 cases covering more than 11 million square feet were grandfathered when the county acted [5].
Pauses, Caps and Rewritten Zoning
How Amsterdam, Dublin, Frankfurt, Singapore and Northern Virginia responded
Every major market that hit its grid limit paused first and regulated second. Amsterdam and Haarlemmermeer froze permits in June 2019, Singapore paused from 2019 to 2022, Dublin effectively stopped new connections from 2021 to December 2025, and Loudoun County paused applications in September 2026 ✓ Established [64] [65] [66] [5].
Amsterdam was the first city to treat data centres as a planning problem rather than an investment opportunity. After the 2019 freeze, its 2020 policy limited additional capacity to 670 MVA until 2030, directed growth into designated clusters and required a power-usage effectiveness below 1.2 and the reuse of residual heat ✓ Established [64]. Neighbouring Haarlemmermeer restricted growth to designated business parks and declared that no new data-centre space would be available after 2030 [67]. The national government added an interim restriction on hyperscale sites, defined as more than 10 hectares and 70 MW, in February 2022 [64]. In June 2025 Amsterdam stopped new large data centres altogether, citing a lack of both space and grid capacity [4]. Rules written in megawatts invite workarounds: Microsoft is building three 85 m towers in the city’s port area whose applications were filed tower by tower, so the province did not classify the project as hyperscale ✓ Established [30].
Dublin’s pause came from the grid operator rather than the city. In late 2021 EirGrid signalled that new data-centre connections in the Dublin region would not be offered until at least 2028, and the regulator set criteria including on-site generation and the ability to reduce demand [31]. The de facto moratorium ended with the Commission for Regulation of Utilities’ decision on large energy users of 12 December 2025 ✓ Established [66]. New data centres must now provide dispatchable generation or storage on or near the site to cover their demand, and match 80% of their annual consumption with new Irish renewables within six years [66]. Planning refusals proved less durable. South Dublin County Council refused Google’s Grange Castle extension in August 2024, citing insufficient grid capacity and the lack of significant on-site renewable energy, and An Coimisiún Pleanála overturned that refusal in September 2026 subject to power purchase agreements and renewable fuel for the backup generators [59].
Frankfurt chose zoning. Its data-centre master plan, approved by the city council on 9 June 2022, divides the city into suitable, restricted and excluded areas, steering new sites into districts such as Sossenheim, Rödelheim, Griesheim, Gallus, Ostend, Fechenheim and Seckbach and estimating that about 75 hectares will be needed by 2030 ✓ Established [53]. The city also aims for Blue Angel environmental certification of new data centres [53]. Zoning has not solved the grid problem: Mainova expects new high-performance connections only in the mid-2030s [3]. Singapore chose allocation. After its 2019 pause, the government selected four proposals for about 80 MW in July 2023 from more than 20 bids, requiring top-grade green building certification [68]. Its 2024 Green Data Centre Roadmap promised at least 300 MW more, and a December 2025 call offered at least 200 MW to builds with an effectiveness below 1.3 and at least 50% green energy ✓ Established [65] [69]. The pause pushed growth across the causeway: Johor’s boom has been described as a direct consequence of Singapore’s moratorium [70].
Northern Virginia, the largest market in the world, has moved furthest and fastest in the past two years. Loudoun County ended by-right data-centre development on 18 March 2025, so that most new sites need a special exception, and it prohibits data centres in its rural policy area, about two thirds of the county [63]. On 15 September 2026 its board voted 7-1-1 to pause consideration of legislative applications for data centres and substations while it rewrites its standards, although at least 17 grandfathered cases covering more than 11 million square feet are not affected ✓ Established [5]. In Prince William County, a circuit judge voided in August 2025 the rezoning for the Digital Gateway, more than 2,100 acres beside Manassas National Battlefield intended for 37 data centres, because the county’s advertised notice did not comply with state or county code, and the Court of Appeals upheld the ruling on 31 March 2026 [72] [71]. Outside Virginia, Tucson’s council voted 7-0 in August 2025 to end work on the 290-acre Project Blue, and Data Center Watch counted 64 billion dollars of US projects blocked or delayed by local opposition between May 2024 and March 2025 [73] [74].
- Moratoria: Amsterdam 2019 and 2025, Singapore 2019 to 2022, Dublin 2021 to 2025, Loudoun 2026, each followed by conditional reopening or new standards [64] [65] [66] [5].
- Zoning and setbacks: Frankfurt’s suitable and excluded areas, Fairfax’s 200 and 300 foot buffers, Chandler’s planned-development zones [53] [45] [44].
- Capacity caps: Amsterdam’s 670 MVA ceiling to 2030 and Singapore’s allocation calls [64] [69].
- Bring your own power: Ireland’s requirement for on-site dispatchable generation and new renewables [66].
- Efficiency and heat mandates: Germany’s 2023 law, softened in 2026 [56] [55].
Japan is at the guideline stage. The Tokyo Metropolitan Government issued guidelines for data centres in harmony with the city in March 2026, collecting points for smooth local dialogue and describing data centres as core infrastructure for the digital city [75]. The Japan Data Centre Council published its own coexistence guideline in May 2026, and the Ministry of Internal Affairs and Communications will score subsidy applications higher from fiscal 2026 if they comply with it ✓ Established [8]. The land ministry has circulated the guidelines to local building departments and will survey district-plan tools that municipalities could use [8]. None of this is binding. Osaka has no data-centre zoning category, no noise rule specific to the sector and no environmental assessment trigger, which is why the residents’ petition to the prefecture in August 2026 asked for all three in substance [21]. The petition sought environmental assessment, weight for residents’ consent in development permits and consensus-building from the siting stage; the prefecture’s written reply was not reviewed for this report [21].
| Approach | Risk of failure | Assessment |
|---|---|---|
| Blanket moratorium without new rules | Displaces growth rather than shaping it: Singapore’s pause fed Johor, and Dublin’s share kept rising as contracted projects connected [70] [2]. | |
| Megawatt caps on single projects | Invite project splitting, as Microsoft’s three separately filed Amsterdam towers show [30]. | |
| Voluntary coexistence guidelines | Japan’s current approach; useful for dialogue but unenforceable, and residents are already turning to the courts [8] [20]. | |
| Heat-reuse mandates without heat networks | Germany’s mandate was softened within three years because many sites had no viable connection [55]. | |
| Zoning, setbacks and own-power rules | The most durable tools so far, as in Frankfurt, Fairfax and Ireland, though planning appeals can still override refusals [53] [45] [59]. |
Who Pays for the Grid
Household bills, cost allocation and the arguments on both sides
PJM’s independent market monitor attributes 29.4 billion dollars, 46% of the charges in the grid operator’s last four capacity auctions, to data-centre demand ◈ Strong Evidence [76]. Other researchers find that fast load growth has so far tended to hold retail prices down ⚖ Contested [77]. Both can be true at once.
The cost question is the most contested part of the debate, and it is the one that reaches residents who never see a data centre. In the PJM market, which serves Virginia and twelve other states, the independent market monitor found that data centres added 9.4 billion dollars to capacity costs in the 2025/26 auction alone, and wrote that current market conditions are almost entirely the result of large load additions from data centres, both actual and forecast ◈ Strong Evidence [78]. In July 2026 the monitor put data-centre-driven charges across the last four base capacity auctions at 29.4 billion dollars, 46% of 63.6 billion [76]. Capacity charges flow through to every customer’s bill. In November 2025 Virginia’s regulator approved a Dominion Energy increase of 13.60 dollars a month over two years for a typical residential customer, and created a new rate class for very large loads from January 2027 to ring-fence their costs [79].
Forecasts sharpen the point. JLARC estimated that, if data-centre demand grows without constraint, generation and transmission costs for a typical Dominion residential customer could rise by about 37 dollars a month by 2040 in constant dollars, or about 14 dollars under a scenario with half that demand ✓ Established [7]. In Ireland, RTÉ reported in August 2025 that the regulator had estimated domestic bills rising by 8% to 21% over five years, while large energy users including data centres could expect a 14% decrease, even though data centres and technology companies are responsible for 89% of EirGrid’s projected demand growth over the decade [32]. Harvard’s Electricity Law Initiative reviewed nearly 50 state regulatory proceedings in ten states and concluded that utilities are using special contracts, often secret, to give data centres discounted rates while shifting costs to other ratepayers ◈ Strong Evidence [39]. Dominion, for its part, said it recognised the impact on customers and was offering more energy-saving programmes and assistance options than ever before [79].
Allowing in another cohort of big new data centres would require yet-further network upgrades to accommodate them. All that upgrade work needs to be paid for, and this will propagate through to electricity bills. There is no free lunch.
— Paul Cuffe, Associate Professor, University College Dublin, to RTÉ Prime Time, August 2025 [32]The counter-evidence is serious. Researchers at Lawrence Berkeley National Laboratory, studying state retail prices between 2019 and 2024, found that load growth at the state level has tended to depress retail electricity prices, because fixed network costs are spread over more sales ⚖ Contested [77]. On that view, a large new customer that pays its own connection costs lowers everyone else’s bills. Loudoun County is the fiscal version of the same argument: data centres supply 31% of its local tax revenue, income that would otherwise fall on residents [7]. Industry adds that data centres are now treated as critical infrastructure in the United Kingdom, that the west London housing delays had other causes, and that limiting development in Ireland could put 94,000 jobs at risk, according to a KPMG study for the Department of Enterprise [35] [80]. The disagreement turns on timing and allocation. Past growth arrived when grids had spare capacity; future growth requires new generation and wires, and who pays for those is a regulatory choice, not a law of nature [39] [66].
The PJM market monitor attributes 9.4 billion dollars of one auction’s capacity costs to data centres [78], and Ireland’s regulator expects domestic bills to rise while large users’ bills fall [32]. Berkeley Lab finds that past load growth tended to lower retail prices [77]. Regulators who have acted, in Virginia with a new large-load rate class and in Ireland with own-generation rules, have done so by moving grid costs onto the data centres themselves [79] [66].
The second contested question is whether cities are the wrong place for data centres at all. The case against urban sites is that they compete with housing for grid capacity, as in west London and Amsterdam, and put generators, fuel and plant beside homes [29] [30] [7]. The case for them is that rural alternatives have their own costs: water conflicts in Chile and Arizona, trees and groundwater in Akishima, a battlefield landscape in Prince William County, and wasted heat wherever there is no district network to receive it [38] [73] [18] [72] [49]. Heat reuse, the strongest argument for siting close to people, only works where a city has built the network to use it ◈ Strong Evidence [50] [55]. The third question is whether pauses work. Singapore and Ireland used them to extract conditions, but both saw growth displaced or delayed rather than reduced [69] [70] [2]. Amsterdam’s stop, by contrast, spared its grid operator about 125 million euros of substations, a measurable gain even if some demand moved to Almere and Zeewolde [4] [25].
The Case for Downtown Data Centres
Exchanges and cloud access points sit in city centres; OC1 is within 3 km of Dojima and Shinsaibashi [1].
Espoo, Odense and Saint-Denis show waste heat can warm homes and pools where networks exist [49] [50] [51].
Data centres supply 31% of Loudoun’s local tax revenue [7].
State-level load growth has tended to depress retail prices since 2019 [77].
The Case Against
West London housing was halted from 2022 and 30,000 Amsterdam homes are at risk from congestion [29] [30].
About 50 full-time workers per large building, at over 1 million dollars of tax breaks per permanent job [7] [60].
Northern Virginia generators impose an estimated 190 to 260 million dollars a year in health costs [47].
Data centres account for 46% of four PJM capacity auctions’ charges [76].
Ireland shows how the cost debate widens beyond bills. RTÉ reported in August 2025 that about 16 large data centres rely on gas, and Hannah Daly, a professor at University College Cork, warned that Ireland was simply funding renewables to meet huge new sources of demand rather than cutting fossil fuels [32]. Friends of the Earth Ireland estimated that households paid 360 euros extra between 2015 and 2023 because of data-centre demand, and its campaigner Rosi Leonard argued that data centres are fuelling the country’s dependence on fossil gas as well as wholesale prices [80]. Against that, the KPMG study for the Department of Enterprise warned that limiting development could put 94,000 jobs at risk ⚖ Contested [80], a figure that sits uneasily beside the Department of Finance’s own view that the direct employment effect is relatively limited [58]. The Irish regulator’s answer, requiring new sites to bring dispatchable generation and match 80% of their demand with new Irish renewables, is an attempt to settle the cost question by making the data centre carry it [66]. Its effect on bills will not be measurable for several years.
Whether a data centre raises its neighbours’ bills, keeps them awake or warms their homes is decided less by the technology than by the rules around it. Amazon cut its Great Oak noise by about 15 dB once pressed [43], Ireland made new sites bring their own power [66], and Espoo built the heat pumps [49]. Cities that leave these terms to the operator get the default, which is the cheapest building for the operator.
What the Evidence Tells Us
Lessons for Osaka before its pipeline arrives
A downtown data centre is an industrial load that cities have been permitting as an office. The cities that managed it best set its terms early: where it may go, what power it must bring, how quiet it must be and who takes its heat ◈ Strong Evidence [53] [66] [45] [49].
The first lesson is that concentration, not any single building, is what changes a city. OC1 at 7.4 MW is a modest neighbour; Dublin’s 1,150 MW cluster and Frankfurt’s 40% of city power are what forced regulators to act [1] [66] [3]. Osaka is earlier on that curve. Greater Osaka’s development pipeline is projected to grow by 170% in critical capacity, its bay front is gaining two campuses of roughly 100 MW each, and the government expects data centres to add about 6 GW of national peak demand by 2035 ◈ Strong Evidence [15] [11] [13] [8]. Those numbers do not yet put Osaka where Dublin was in 2021, but they move it in that direction faster than Japanese planning law is moving. The second lesson is that the grid is the scarce asset. Once connection capacity is allocated to data centres, it is not available for housing, transit electrification or industry, and the queue is first come, first served unless the rules say otherwise [29] [8].
The third lesson is that neighbourhood harms are real but tractable, and they are tractable only when they are measured. Noise complaints in Virginia arose at 40 to 59 decibels because the standard measured the wrong thing; Loudoun is now rewriting it for low-frequency and tonal sound [7] [41]. Generator emissions were small in regional inventories but concentrated in particular neighbourhoods, and the health costs were invisible until researchers modelled them [7] [47]. Japan exempts emergency generators from emission standards and covers data-centre blowers only in designated noise areas [8]. It does not publish data-centre electricity or water use by city, as Ireland and the Netherlands do [2] [34]. Osaka will not be able to manage what it does not count. The fourth lesson is that the local return is mostly fiscal. Jobs are few, heat reuse needs infrastructure the city must build, and tax revenue is significant only where the cluster is very large [7] [49].
The likely shape of the next decade is already visible. Industry forecasters expect the sector to grow by about 97 GW worldwide between 2025 and 2030, and both JLL and CBRE argue that AI inference will pull more computing closer to users ◈ Strong Evidence [22] [24]. That points to a split model: small, dense halls in city centres for interconnection and inference, and very large campuses on industrial or rural land for training and storage. Osaka already has both forms, a 7.4 MW tower in Umeda and 100 MW-class sites at Suminoe-ku and Ibaraki, with an 8 MW Chuo-ku building to follow [1] [11] [14] [10]. The new buildings are denser than the old: SC Zeus designs for up to 130 kW per rack with liquid cooling, and NTT West plans liquid cooling for parts of its downtown site [11] [10]. Density raises the stakes for every local issue in this report, because more power, more heat and more backup fuel are packed into the same plot of urban land. The rules Osaka writes for small downtown towers will therefore shape the larger ones that follow.
- Count first: publish data-centre electricity and water use by prefecture or city, as Ireland’s statistics office does every year [2].
- Zone explicitly: create a data-centre use class with setbacks and plant screening, as Frankfurt and Fairfax have done and the London Assembly recommends [53] [45] [29].
- Measure the hum: add low-frequency and tonal criteria to noise rules and require post-construction studies, as Chandler does [41] [44].
- Price the grid: make large loads pay for the capacity they reserve, as Ireland and Virginia now do [66] [79].
- Plan for heat: tie urban approvals to heat offtake where a network exists or is planned [49] [55].
Japan has started to build the tools. From October 2026 grid operators can cancel stalled connection applications, and from April 2027 they will be able to release capacity that has been reserved but not used, with cost recovery ✓ Established [8]. Every grid operator now publishes a map of zones where power can be supplied quickly, which steers large loads towards places the grid can serve [8]. The government’s 2026 white paper restates the goal of dispersing data centres away from the Tokyo and Osaka regions for resilience [16]. These measures work at national scale. What is missing is the municipal layer that Frankfurt, Fairfax and Chandler supplied: a planning category, binding setbacks, a noise standard and a duty to consult that does not depend on an operator’s goodwill [53] [45] [44]. The residents who petitioned Osaka Prefecture in August 2026 asked for exactly that, framed as environmental assessment [21]. Tokyo’s 2026 guidelines show that the metropolitan government is aware of the gap, but guidance on dialogue is not the same as a rule an operator must obey [75].
Most of the debate about urban data centres is conducted in terms of jobs, noise and heat. The underlying exchange is simpler. A city grants a private operator a long-term claim on its scarcest infrastructure, the spare capacity of its substations and transmission lines, in return for property tax and a small workforce [7] [22]. Every jurisdiction that later regretted the terms, from Dublin to Amsterdam, had made that grant without pricing it [66] [4].
For Osaka the timing is favourable. Its first downtown tower opened in 2026 without controversy, its large builds are on industrial land at the bay front, and no planning conflict has been documented inside the city [1] [12] [21]. That is the position Loudoun County was in before its residents measured the hum, Dublin before its data centres outgrew its households, and Amsterdam before it ran out of substations [7] [6] [4]. The evidence does not say that data centres should be kept out of cities. It says that a city which wants them should decide in advance what share of its grid it will give them, what they must bring in return and how their neighbours will be protected, because after the connection is granted the city has very little leverage left ◈ Strong Evidence [66] [5] [59]. Osaka still has that leverage, and the decisions that use it will have to be taken before the next downtown tower is designed rather than after it is connected [10] [8].