ZanKore AI Factory Targets 1 GW, but Central Java Faces a Water Test
ZanKore has proposed a 1-gigawatt AI factory in Central Java, despite unresolved questions about the water needed to cool its servers. The first phase is expected to provide roughly 200 megawatts of capacity in Batang during the first half of 2027.
That scale would give Indonesia a major domestic source of computing capacity for training and operating artificial intelligence models. It would also place a continuously running industrial facility in a province that distributed more than 75 million liters of emergency water during the 2026 dry season.
The conflict is not simply between technology and the environment. It is between a national infrastructure ambition and the local systems expected to support it. NVIDIA hardware, international financing, and regional demand can accelerate construction. None of them guarantees a reliable or equitable water supply.
The ZanKore AI factory has disclosed its computing targets, hardware direction, partners, and prospective customers. It has not publicly provided the same level of detail about cooling technology, expected water consumption, water sources, or drought procedures.
That information gap now matters as much as the project’s headline capacity. Without it, residents and policymakers cannot determine whether ZanKore will become an efficient anchor for Indonesia’s AI economy or another industrial user competing for a strained resource.
ZanKore’s AI Factory Is Moving From Ambition to Infrastructure
ZanKore is no longer presenting a general AI strategy. It is assembling a specific, energy-intensive computing platform with a defined location and deployment schedule.
Indosat Ooredoo Hutchison launched ZanKore with Ooredoo Group, Nokia, and NVIDIA on August 6, 2026. The partners described a platform designed to serve hyperscalers, enterprises, governments, developers, and AI companies across the Asia-Pacific region.
Its long-term target is 1 GW of NVIDIA DSX AI Factory capacity. DSX is NVIDIA’s reference approach for combining accelerated computing, networking, power, cooling, and facility operations into one system.
The partners initially said ZanKore would deliver approximately 200 MW of capacity in the first half of 2027. A subsequent announcement described an initial 100 MW of NVIDIA infrastructure, indicating that the early deployment will proceed in stages.
The project intends to use NVIDIA GB300 NVL72 systems. These integrate large numbers of processors, networking components, and liquid-cooled server racks for high-density AI workloads.
ZanKore also says it will use NVIDIA DSX MaxLPS. That software coordinates power consumption across a GPU fleet, shifting capacity between workloads to make fuller use of the facility’s available electrical supply.
Those choices explain why the project is being called an AI factory rather than a conventional data center. The facility will be designed around dense GPU clusters that deliver computing capacity for model training, fine-tuning, and inference.
According to the project’s infrastructure announcement, the platform will also connect with Sahabat-AI, an open model supporting Bahasa Indonesia and regional languages. The government has described a 70-billion-parameter version covering Javanese, Sundanese, Balinese, and Batak.
Local computing capacity can reduce dependence on facilities in other countries. It can also give public agencies and regulated industries more control over where sensitive information is processed.
That sovereign-computing argument is central to ZanKore’s pitch. Indonesia has a large digital market, but much of the hardware supporting advanced AI services sits elsewhere.
Domestic infrastructure gives Indonesian developers lower-latency access to processors and creates a local base for AI services. ZanKore also expects to serve regional customers, making the facility an exporter of computing capacity rather than a purely domestic utility.
Yet the location turns this digital strategy into a physical planning problem. The Batang site sits in Central Java, where power, land, connectivity, and industrial infrastructure are being assembled for large manufacturers.
Computing can be sold across national borders, but its environmental costs remain tied to the host community. Electricity must reach the racks, heat must leave them, and any evaporated water disappears from the local supply.
The project therefore creates a clear bargain. Indonesia receives computing capacity and investment, while Batang assumes the operational demands of one of the region’s largest proposed AI facilities.
Whether that bargain is reasonable depends on details that ZanKore and public authorities have not yet disclosed.
Why the ZanKore AI Factory Water Estimate Is So Wide
The most alarming water estimates are scenarios, not measured consumption, because ZanKore has not published its final cooling design.
AI processors convert nearly all the electricity they use into heat. A facility must remove that heat continuously to keep servers within safe operating temperatures.
Evaporative cooling rejects heat by turning water into vapor. It can reduce electricity demand in suitable conditions, but the evaporated portion cannot immediately return to the local water system.
Direct liquid cooling works differently. A liquid carries heat away from processors through a closed internal loop, which can improve heat transfer inside dense racks.
However, liquid-cooled servers do not automatically create a waterless facility. The building must still reject the collected heat through cooling towers, dry coolers, chillers, or another external system.
This distinction is crucial. ZanKore’s reference to liquid-cooled infrastructure describes how heat moves away from the chips, not necessarily how the facility releases that heat outdoors.
Water Usage Effectiveness, or WUE, measures on-site water consumption relative to the energy used by computing equipment. It is usually expressed as liters per kilowatt-hour.
A facility relying heavily on evaporative cooling can consume far more water than one using dry heat rejection. Local temperature, humidity, workload, system controls, and water quality also affect the result.
Research summarized in a 2026 water footprint review notes that air-cooled facilities can approach zero operational WUE. Water-dependent designs can exceed three liters per kilowatt-hour under some conditions.
Applying an assumed efficiency rate to ZanKore’s announced electrical capacity produces a large theoretical number. One published analysis used an Indonesian projection of approximately 1.9 liters per kilowatt-hour.
At that rate, a continuously loaded 200 MW installation would consume about 9.1 million liters daily. Lower utilization or a more efficient system would reduce that figure.
The same analysis gave a daily range of approximately 4.8 million to 9.1 million liters for the first phase. At the full 1 GW target, a simple linear estimate would be five times larger.
These figures should not be presented as ZanKore’s confirmed demand. They depend on assumptions about server load, operating hours, cooling architecture, weather, and the share of heat rejected through evaporation.
They also exclude or blur other categories of water use. Electricity generation can consume water away from the site, while semiconductor fabrication carries an embedded water footprint from earlier in the supply chain.
Workload-level research shows why a single universal figure can mislead. The water used for a computing task can vary by more than four orders of magnitude depending on the hardware, facility, location, and time of operation.
A published estimate is therefore best treated as a pressure test. It shows what ZanKore’s water demand might look like under a water-intensive design, not what the facility will inevitably consume.
That uncertainty does not weaken the case for scrutiny. It makes disclosure more urgent.
ZanKore can narrow the estimate by publishing its design WUE, annual water budget, peak daily requirement, projected utilization, and cooling mode for each season. It should also separate potable water, raw water, recycled water, and water consumed indirectly through electricity.
Annual averages alone would be insufficient. A cooling system’s demand can peak on the same hot, dry days when households, farms, and public services face the greatest pressure.
The company should explain whether its closed loops require periodic draining, how it will manage concentrated minerals, and where discharged water will go. These operational details determine whether water is reused, evaporated, or returned in a condition suitable for another purpose.
Until those figures appear, both categorical reassurance and categorical condemnation go beyond the public evidence. The responsible conclusion is narrower: the disclosed computing scale creates a material water question that the current project description does not answer.
Central Java’s Drought Record Changes the Risk Calculation
Millions of liters per day mean something different in Central Java after more than 2.3 million residents needed drought-related water assistance.
From June 5 through September 28, 2026, government agencies and supporting organizations distributed 75,463,000 liters of clean water across Central Java.
The aid reached 2,341,548 people in 784 villages or urban wards, according to provincial drought data. The deliveries covered 33 regencies and cities.
That total was not a permanent reserve stored for one location. It was emergency water distributed across a large province over several months.
Comparing it directly with a data center’s theoretical daily consumption can therefore distort the underlying situation. The comparison remains useful as a scale indicator, but it does not prove that ZanKore would take water from emergency trucks.
The more relevant issue is competition within connected water systems during dry periods. Industrial users, households, agriculture, and environmental flows can depend on the same rivers, reservoirs, treatment assets, or aquifers.
If the first ZanKore phase consumed between 4.8 million and 9.1 million liters daily, it would equal the province’s entire drought distribution volume every eight to 16 days.
Again, that is an illustration based on an assumed cooling model. It is not a forecast confirmed by ZanKore.
The comparison shows why a facility-level water plan needs to address drought conditions rather than only normal annual supply. A data center cannot easily suspend critical customer workloads whenever a watershed tightens.
The northern Central Java coast also faces land subsidence, which occurs when the ground surface sinks. Researchers and government agencies have recorded rates of roughly 6 to 20 centimeters annually in parts of the Pekalongan-Batang area.
Groundwater extraction is not the only cause of subsidence, but excessive pumping can worsen it. Sinking land increases exposure to tidal flooding, damages infrastructure, and allows saltwater to move farther inland.
That makes the project’s water source essential information. Surface water, treated wastewater, seawater, and groundwater carry very different consequences.
A promise to avoid potable water would answer only one question. Raw surface water can still have competing agricultural or ecological uses, while seawater cooling requires controls for intake impacts, corrosion, and heated discharge.
Local authorities have been developing water infrastructure around the Batang industrial area. One commercial project describes a long-term plan for 250,000 cubic meters of daily demineralized water capacity, with an intermediate phase expected in 2027.
Capacity on paper does not automatically establish availability for ZanKore. It remains unclear which source, allocation, delivery schedule, or contractual arrangement would support the AI facility.
The Batang government has separately described plans to improve clean-water access for 5,000 homes near the industrial area by 2028. Officials said that supply would prioritize residents rather than large factories.
That timeline highlights a distribution question. A major computing installation is expected to begin its first phase before some nearby residential water improvements arrive.
A sound project assessment must therefore go beyond whether enough water can technically be moved into the industrial estate. It must ask who receives reliable service first, what happens during a shortfall, and which users bear the cost of new infrastructure.
Central Java does not need to reject industrial growth to ask those questions. It needs enforceable allocation rules before large users establish permanent demand.
The Real Tradeoff Is Water, Energy, and Compute
ZanKore can reduce freshwater consumption, but every cooling alternative shifts costs into electricity use, capital equipment, or another environmental system.
Dry cooling transfers heat into the air without relying on continuous evaporation. It can sharply reduce on-site water consumption, especially when paired with closed-loop liquid cooling inside the server hall.
The tradeoff is energy. Air-based heat rejection becomes less efficient during hot weather, which can increase electricity demand and reduce the computing capacity available within a fixed power envelope.
That matters for a project built around maximizing AI output from every megawatt. A cooling design that saves water can work against the platform’s stated efficiency goal.
Hybrid systems offer another route. They can use dry cooling during favorable conditions and introduce evaporation only when temperatures or workloads rise.
A hybrid design can reduce annual water consumption without accepting the full energy penalty of an entirely dry system. Its performance still depends on controls, weather, and the amount of time spent in water-intensive mode.
Recycled wastewater could reduce demand for freshwater. It would require treatment infrastructure that consistently delivers water at the quality needed by cooling equipment.
Water impurities can cause scaling, corrosion, and biological growth. Cooling towers also concentrate dissolved material as water evaporates, forcing operators to discharge some fluid and replace it.
Seawater is another possibility because Batang lies on Java’s northern coast. It could reduce pressure on rivers and aquifers if the facility and its supporting infrastructure were designed for it.
Seawater is not impact-free. Intake systems can affect marine organisms, while concentrated or heated discharge can alter nearby coastal conditions.
These options demonstrate why broad environmental labels are inadequate. A system described as water-efficient may use more electricity, and a system described as low-carbon may still depend on freshwater evaporation.
ZanKore’s use of NVIDIA’s power-management system adds another layer. Better workload scheduling can recover electrical capacity that would otherwise remain unused.
If that recovered capacity supports more active processors, total heat output can rise even while computing per megawatt improves. Efficiency at the server level does not guarantee lower resource use at the facility level.
This is a familiar rebound effect. Lower resource use per unit can encourage operators to produce more units, leaving total consumption unchanged or higher.
The 1 GW target makes total impact more important than efficiency ratios alone. A highly efficient gigawatt facility can still consume more electricity and water than a smaller, less efficient installation.
ZanKore’s strongest case is that Indonesia needs local computing infrastructure and should not surrender the economic value of AI to foreign facilities. Its partners also argue that an integrated platform can support national languages, regulated workloads, and regional developers.
The skeptical response is not that these benefits are imaginary. It is that they do not determine how the site should be cooled or who should carry its environmental risk.
There is also a business uncertainty. A 1 GW platform assumes sustained demand from Indonesia and neighboring markets.
Building in phases can limit exposure. The first deployments will reveal whether regional customers will commit workloads at the scale needed to justify later construction.
Phased deployment can also support environmental learning, but only if each phase has measurable conditions. Authorities can require verified water and energy performance before approving expansion.
Without such gates, early infrastructure may create momentum for later phases regardless of observed local impacts. The initial water plan must therefore account for the full buildout, even if permits are issued incrementally.
Public Disclosure Is the Missing Piece
The central problem is not that ZanKore has already proven unsustainable. It is that outsiders cannot test the project’s sustainability claims with the information available.
ZanKore has disclosed unusually specific computing ambitions. The company has named its hardware generation, reference architecture, initial capacity, long-term target, intended customers, and deployment period.
Its environmental information remains much less specific. Public materials do not identify a design WUE, annual consumption target, maximum daily draw, or cooling method for the Batang facility.
They do not state whether the project will use groundwater, surface water, municipal supply, reclaimed wastewater, seawater, or a combination.
They also do not explain what happens when water availability falls below normal. That omission is important because AI factories are sold on continuous service.
A credible drought plan would establish priorities before a crisis. It would describe whether ZanKore will reduce workloads, switch cooling modes, import water, or rely on reserved industrial supply.
The plan should also identify who has authority to order reductions. Voluntary commitments offer limited protection when a facility has contracts requiring high availability.
Indonesia’s communications minister has presented ZanKore as infrastructure that can improve competitiveness and attract valuable investment. The government says the project has received support since it was proposed two years earlier.
Officials also expect the facility to support local AI development through Sahabat-AI and computing services. Those goals give the government a direct interest in the project’s success.
That relationship makes independent oversight particularly important. The same public institutions promoting AI capacity must also protect water access and assess environmental permits.
The project’s government launch account says construction in Batang is scheduled for the first half of 2027. Central Java’s governor said in October 2026 that the proposal remained under study.
Those statements are not necessarily contradictory. A national project can have an announced commercial timetable while local reviews remain unfinished.
However, the difference creates a practical deadline. Authorities need enough time to assess water demand before procurement and construction make the cooling design difficult to change.
At minimum, the environmental review should publish several items:
Water consumption under average and peak weather conditions
Annual and daily limits for each project phase
The source and legal allocation of every water stream
The expected use of potable, raw, recycled, and seawater
Cooling modes and their estimated operating hours
Groundwater extraction limits, including indirect suppliers
Discharge volumes, temperature, and water quality
Drought triggers and mandatory response actions
Independent monitoring results after operations begin
Expansion conditions tied to measured performance
Public reporting should use water consumption and water withdrawal as separate measures. Withdrawal counts water taken from a source, while consumption tracks the portion not promptly returned.
A facility can withdraw a large volume but return much of it. It can also withdraw less and consume a high share through evaporation.
Monthly or seasonal figures would be more useful than a single annual total. They would show whether consumption rises when local supplies are most stressed.
The review should also evaluate cumulative demand from the entire industrial zone. Assessing ZanKore in isolation could miss competition from manufacturers, residential growth, power generation, and future data centers.
Community participation matters because aggregate capacity figures do not capture household reliability. Residents can experience low pressure, salinity, longer collection times, or higher costs even when a regional balance sheet shows enough water.
These requirements would not single out ZanKore unfairly. They would establish a template for future Indonesian data centers, including other proposed AI facilities.
The same standard should apply to domestic and international investors. A consistent rule would give operators clearer expectations while protecting communities from case-by-case bargaining.
Three Signals Will Decide Whether Central Java Can Carry the Load
ZanKore’s announced capacity is not the next decisive milestone. Cooling disclosure, enforceable water allocation, and measured first-phase performance are.
The first signal is a detailed environmental assessment that identifies the cooling architecture and water source.
If ZanKore commits to low-water heat rejection, reclaimed supply, and strict groundwater limits, the upper-end consumption scenarios will become less relevant. That would strengthen the case that Batang can support the facility without exposing residents to the highest modeled demand.
If the assessment depends primarily on freshwater evaporation, the scrutiny should intensify. Authorities would need to demonstrate how the supply performs during dry conditions, not merely on an average day.
The second signal is an enforceable allocation agreement between ZanKore, water providers, and public authorities.
The agreement should establish daily limits, drought thresholds, residential protections, and reporting duties. It should also specify which party finances treatment, pipelines, monitoring, and contingency capacity.
A vague assurance that sufficient industrial water exists would weaken confidence. Large headline capacity can mask incomplete pipes, delayed treatment plants, competing contracts, or seasonal limitations.
The third signal is independently verified operating data from the first phase. Design estimates cannot capture every interaction among weather, utilization, hardware, and cooling controls.
Once the initial GPU clusters begin operating, monthly WUE and total consumption will reveal how the facility behaves. The data should distinguish testing periods, partial utilization, and mature operations.
Early low consumption would mean little if most servers remained idle. Reporting water use alongside computing load and IT energy would make performance comparisons more meaningful.
Those results should govern later expansion. Meeting a transparent efficiency limit would support additional phases, while exceeding it should trigger redesign or a pause.
The project’s phased schedule gives Indonesia an opportunity to build that accountability into its approvals. It does not need to predict every consequence of a 1 GW facility before the first rack arrives.
It does need rules that prevent the final scale from becoming inevitable regardless of evidence.
The larger question extends beyond one industrial site. Countries seeking sovereign AI capacity must decide what resources that sovereignty requires and where its costs belong.
Compute independence means little if it creates a new dependency on scarce local water. Environmental restraint also becomes difficult to defend if it leaves a country permanently dependent on foreign infrastructure.
The answer lies in measurable design choices rather than slogans. ZanKore can publish its assumptions, accept binding water limits, fund the infrastructure it requires, and demonstrate performance before expanding.
Central Java can support digital development while treating household water security as a non-negotiable constraint. Those goals only conflict when resource planning remains hidden or arrives after construction decisions.
The ZanKore AI factory now has a public timetable, major partners, and a defined computing target. Its water strategy deserves the same precision.
Before the first 200 MW comes online, readers, developers, customers, and residents should ask one direct question: will ZanKore publish a phase-by-phase water budget that holds up during drought?
That disclosure would not settle every concern. It would turn a debate driven by theoretical estimates into one based on enforceable limits and measured results. Until then, Central Java is being asked to trust the project’s infrastructure promise without seeing the full resource bill.
