AI Trends

Data centers face protests overseas as China ties rack utilization rates to expansion conditions

36Kr · AIUpdated Nov 17 22:05

AI summary

More than 100 protesters gathered at a data center construction site in Amsterdam planned for Microsoft, demanding a pause on large-scale construction until the government sets clearer rules. In March 2025, Loudoun County, Virginia, reclassified data centers as a use requiring a special permit, while AEP Ohio requires new large data centers to bear minimum costs and forgo compensation. As of the end of July this year, China's intelligent computing capacity reached 2.45 million PFLOPS, and it stipulates that cities with a rack utilization rate below 50% will in principle no longer plan new data center clusters or large and extra-large projects.

Body · AI translation

On November 6, more than 100 protesters gathered at a data center construction site in Amsterdam planned for Microsoft's use; some entered the site and climbed onto cranes. Reuters reported that the coalition organizing the action demanded a pause on large-scale construction until the government sets clearer rules. The project owner said it was willing to communicate with local parties, while Microsoft said it respects the right to peaceful expression but does not endorse illegal acts.

From cloud computing to artificial intelligence, data centers are taking on more and more demand, yet in some places they are encountering increasingly direct resistance when it comes to siting. For local governments seeking investment, they can bring tax revenue and construction orders; for people living nearby, the first changes may be continuously running equipment, land expansion, and worries about future electricity bills. Customers of technology services are widely distributed, but the land, power, and cooling conditions that support the services must be resolved in one specific place.

These disputes are changing the conditions for building machine rooms. Local governments are re-examining where data centers can be built, and power companies are requiring project developers to bear payment responsibility for future demand over longer periods. China is still rapidly building computing facilities, and has also incorporated existing utilization, energy supply, and task layout into expansion arrangements. What new investment can bring increasingly depends on what resources the local area can provide, when customers actually begin to use it, and who bears the cost if demand falls short.

Machine rooms built right at residents' doorsteps

Local governments welcome data centers for practical fiscal reasons. A 2024 study by JLARC, Virginia's legislative audit agency, found that in five places with relatively mature data center markets, data centers' contribution to local fiscal revenue ranged from less than 1% to 31% of total local revenue. This money can enter public budgets and also makes local governments willing to continue attracting projects. The benefits brought by data centers do not belong only to technology companies.

However, the construction phase and long-term operation leave different things for the local area. The same study, citing company feedback, said that a typical 250,000-square-foot data center may have about 1,500 workers on site at peak construction; after it goes into operation, there are about 50 full-time staff, about half of them contract workers. Machine room investment is large, but the manpower needed for continuous operation is relatively limited. A local area getting a large investment does not mean nearby residents will all get a long-term job.

Changes to the living environment are concentrated in a smaller area. JLARC found that the continuous low-frequency noise emitted by some machine rooms does not necessarily reach the penalty threshold under local noise ordinances, yet still draws complaints from nearby residents. For such households, expanding the local tax base and improving the environment in front of their own homes are two separate matters that need to be resolved.

Some machine rooms are close to housing as a result of earlier planning. In 2000, Loudoun County, Virginia, had no separate data center land-use category, so data centers were managed as office buildings. As long as a parcel allowed office buildings, data centers could usually enter without separately obtaining project approval from the county board. At the time, many office parks were laid out mixed with housing, and this path brought some machine rooms to residents' doorsteps. Continuously running cooling equipment is different from office buildings, and the noise complaints that later emerged have therefore been difficult to resolve solely through the original use classification.

Loudoun County has already changed its approval path. In March 2025, the county board changed data centers in the relevant industrial zoning district to a use requiring a special permit; some applications already accepted earlier retained transitional arrangements. Projects subject to the new rules must go through a public hearing, and then the county board decides whether to approve or reject them. Land allowing industrial development is no longer enough for a data center to directly obtain the original construction entitlement, and the project's impact on surrounding areas has also entered item-by-item discussion and voting.

In the period from 2022 to May this year, spanning before and after the rule change, 32 of Loudoun County's 39 data center applications requiring county board review were approved. Special permits still retain a channel for case-by-case approval, while also placing the surrounding impacts of projects subject to the new rules into public review. What local areas need to decide is which parcels are suitable for continuing to host machine rooms, and what conditions projects should meet.

Water allocation likewise has to return to the local level. JLARC judged in its 2024 study that Virginia data centers' water use at the time was generally sustainable, but among the six water utilities it reviewed, after excluding reclaimed water, data centers' share of water use ranged from 2% to 21%. Different cooling methods, machine room sizes, and computing densities create very different water withdrawal needs. When a local area approves a new park, besides how much water can be connected, it must also consider whether water sources can meet the future needs of residents and other industries.

Electricity, meanwhile, carries impacts beyond the machine room. A working paper revised by the Dallas Fed in October this year used a power dispatch model to estimate that, compared with a scenario without data center load, U.S. wholesale electricity prices in 2025 were on average about 6.9% higher, and Northern Virginia about 15% higher. This compares wholesale markets; how much residents ultimately pay still depends on retail contracts, rates, and cost-recovery arrangements. The same new demand will produce different bills in different places.

For the grid, large customers also have value. When existing power sources and lines have surplus capacity, steadily increasing electricity use can spread fixed costs; when capacity is tight, new machine rooms may require additional power sources, substations, and transmission lines. Research by the International Energy Agency this year repeatedly discussed this difference: what affects electricity bills is not only how much electricity a project uses, but also where it connects, how much more the grid must spend for it, and how that money is ultimately recovered.

Canceling a data center project may also require compensating grid expansion costs

The Virginia study produced a noteworthy result: the rates at the time were already able to reasonably allocate the cost of existing services, including the portion that data centers should bear. But this did not eliminate the risk of the next round of expansion. To meet new demand, the power system still needs to add investment, and once a line and a substation are built, their costs must be recovered through charges over many years.

Power companies cannot wait until all the servers are powered on before starting to build substations and transmission lines. They must first arrange investment based on the demand declared by the project developer and set aside capacity. But the completion of the machine room, the arrival of servers, and the start of customer use often do not happen on the same day. An International Energy Agency report in October this year pointed out that in some regions, the queue of data centers seeking grid connection includes duplicate applications and exploratory projects; even if a project breaks ground, electricity use will gradually increase along with equipment and business. If a project scales down or leaves midway, the facilities built for it still require repayment of construction costs.

Ohio has already begun writing this risk into contracts. Last July, the state Public Utilities Commission approved a dedicated rate for AEP Ohio aimed at large new data centers. Customers can increase load in stages, with a ramp-up period of up to four years, after which they must fulfill an eight-year contract. Even if actual use decreases, they must bear the minimum charges under the contract. The company looks at how much power supply capacity it has reserved for the customer, not just how many kilowatt-hours have passed through the meter.

Before a project is connected to the grid, obligations have already begun: under the current processing procedures, if a customer cancels a project before the scheduled connection date, or delays the project by more than 12 months, it must compensate the power company for all construction costs invested in the project. Customers whose financial strength does not meet the specified conditions must also provide a guarantee or collateral when signing the contract. Applying for a large block of power capacity is therefore no longer just submitting a future expansion plan; it also requires demonstrating the financial ability to perform.

AEP Ohio had suspended signing new data center agreements in central Ohio during the review of the new rules. In an announcement on July 9, 2025, the company explained that this arrangement was to protect the grid and other customers before regulators made a decision; the plan approved that day also clearly set out the procedure for lifting the suspension. Without determined power supply obligations, new projects could not continue to be signed in the original way.

Minimum charges, cancellation compensation, and guarantee requirements return part of the risk of demand failing to materialize to the project developer. When a developer applies for more power supply capacity, it also means assuming payment obligations earlier. Land use approval determines whether the project can be located locally, while the power supply contract determines how grid expansion investment will be recovered; the two procedures respectively affect construction eligibility and subsequent obligations.

China writes machine room utilization rates into expansion conditions

China has formed a construction layout focused on national computing hubs. As of the end of July this year, the country's intelligent computing scale reached 2.45 million PFLOPS, using the FP16 measure; the built intelligent computing scale of the eight national computing hubs and three computing-power and electricity coordination development zones accounted for more than 85% of the national total. As computing facilities cluster in specific regions, power sources, networks, and customer service must also be connected in these places. The construction tasks undertaken by local governments then extend to how completed machine rooms can continue to be put into use.

Customer demand has not remained on planning documents. GDS Holdings disclosed that as of the end of June this year, the company's utilized machine room area increased by 13.2% year on year, and utilized area accounted for 79.2% of area in service, higher than 77.5% in the same period last year.

There is still a time lag between customer commitments and actual use. At the same point, the commitment rate for GDS Holdings' machine rooms in service was 92.4%, higher than the actual utilization rate; the denominator for both indicators is area in service. The company explained that customers will activate committed services in stages; once a machine room is put into service, depreciation and amortization begin to be included in costs, and new machine rooms usually take several years to reach higher utilization and profitability levels. The pace of expansion therefore must also match the progress of existing customers coming online.

While utilized area increases, costs are also squeezing gross margin. GDS Holdings' second-quarter revenue increased by 6.5% year on year, but gross margin fell from 23.8% in the same period last year to 21.5%. The company's explanation was that utility costs such as electricity rose as a proportion of revenue.

The special action plan for green and low-carbon development of data centers issued in 2024 has already written existing utilization into conditions for new construction: if a city already has data centers in use for more than one year, and the overall utilization rate in the city area is below 50%, in principle no new data center clusters and large or ultra-large projects will be planned. For underused facilities, local governments must first clarify whether what is lacking is customers or the capacity to take on new tasks, and then decide whether to continue attracting investment, renovate, or expand.

The scale of electricity use makes power supply arrangements another construction condition. Data from the National Energy Administration show that in 2025, national data center electricity consumption was about 170 billion kilowatt-hours, accounting for about 1.6% of total social electricity consumption. Specific project sites must implement local substation, line, and power source conditions, and the scale of power supply must also align with customer activation progress. Computing-power and electricity coordination considers these conditions together: where computing tasks are suitable to be located, how much load local energy can support, and which supporting facilities need to be built simultaneously.

Direct green power connection is a power supply arrangement. The National Energy Administration disclosed in October that as of the end of September, localities had approved 12 direct green power connection projects for computing facilities. Such projects use dedicated lines to supply green power to computing loads nearby, directly linking power sources and electricity use projects in construction and trading, and also adding local electricity customers for new energy.

After the dedicated line is connected, changes in wind and solar output still need to match the data center load. How storage, regulating power sources, and the grid work together must be arranged according to customer business requirements, and the related investment will also enter project costs. Only by planning power sources, lines, and the data center together can it be clearer earlier how much investment a place's energy advantages require in supporting facilities and which computing businesses they can support.

Anhui tries letting computing power follow green electricity

In Anhui, some have already tried changing electricity use arrangements from the computing task side. On April 3 this year, the State Grid Anhui Electric Power Marketing Service Center and China Telecom Anhui conducted a valley-filling dispatch test. After the power side issued a load regulation demand, Telecom moved the computing tasks running at the Hefei Nangang campus to the Huaibei Telecom Cloud Computing Center. About three minutes later, Huaibei took over and continued inference, and the monitored single-machine electricity load rose by 50%.

Huaibei has surplus green electricity. Once the tasks are moved there, computing demand becomes part of the load that the locality can absorb. This test adjusted the location of computing tasks: the enterprise rearranged work among existing computing nodes so that demand suitable for migration could use the electricity conditions elsewhere. For projects still under construction, this also adds an option—first confirm which tasks can be scheduled, then determine where computing resources need to be increased.

Continuous scheduling also involves authority between customers and operators. The International Energy Agency points out that many data centers serve third-party customers and have limited understanding and control of internal workloads. After a data center receives a grid regulation demand, how much computing it can move depends on which tasks customers hand over to it and what adjustments they allow. Available remote resources, system adaptation, and task completion deadlines all need to enter the service arrangements between the two parties; businesses requiring immediate response still need priority delivery guarantees.

In some places overseas, residents opposing data centers near their homes have already pushed local authorities to rearrange approval procedures; power companies, through contracts, require new projects to bear clearer responsibility for grid expansion needs. As China continues to build computing facilities, existing rack utilization rate requirements constrain new projects, and it is also exploring arranging computing tasks according to energy conditions. How much value a new data center brings depends both on customer demand and on how the locality where it is built allocates land and energy, as well as how long all parties are willing to bear responsibility for their commitments.

This article is from the WeChat public account "Feng Wei Feng Da Bai", author: Feng Da Bai, published by 36Kr with authorization.

Source

36Kr · AI