# Facility settings, heat paths and detection methods ## Facility settings ### Conventional campus Data halls with their own power and cooling systems. The layout is illustrative. Equipment and buildings don’t reveal which workload runs inside. [1](https://epoch.ai/data/data-centers-documentation/methodology) [2](https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf) ### Mislabelled workload A shared data center running a workload other than the one it declared. Watching a facility from outside can’t tell training from inference or prove compliance. [3](https://arxiv.org/abs/2507.15916) [4](https://arxiv.org/abs/2606.19262) [5](https://arxiv.org/abs/2609.00309) ### Civilian cloud Data halls inside a commercial cloud that serves many customers. Physical signs don’t identify the customer or show that separate facilities run one workload. [6](https://arxiv.org/abs/2304.04123) [7](https://arxiv.org/abs/2403.08501) [8](https://ojs.aaai.org/index.php/AAAI/article/download/41127/45088) ### Distributed sites Compute spread across several sites, each with its own power and cooling. Research on distributed training doesn’t show that a frontier-scale deployment could go undetected. [9](https://arxiv.org/abs/2507.07765) [10](https://arxiv.org/abs/2605.29359) [11](https://arxiv.org/abs/2604.04712) ### Industrial host A data hall inside a larger industrial site. The host’s own processes also produce heat, power demand and emissions. Separating them from the compute needs more evidence. [12](https://arxiv.org/abs/2408.16074) [13](https://arxiv.org/abs/2607.22619) [14](https://arxiv.org/abs/2506.15867) ### Underground Data halls in a mine, cavern or bunker. Existing sites use outside power and nearby lake or fjord water; a hidden one could run on its own off-grid power. Existing examples don’t show that hiding works at larger scale. The surface buildings and equipment are still there. [15](https://datacentremagazine.com/data-centres/top-10-underground-data-centres) [16](https://arxiv.org/abs/2609.18824) ### Underwater / offshore Servers in sealed subsea vessels, fed by a power cable and cooled by seawater. One Chinese site is only partly submerged and tied to an offshore wind farm. Small demonstrations don’t prove a large hidden cluster would work. Detection studies so far are modelling only. [17](https://news.microsoft.com/innovation-stories/project-natick-underwater-datacenter/) [16](https://arxiv.org/abs/2609.18824) ### Restricted site A data center inside a compound with controlled access. The cited sources show no proven remote method to tell what workload runs at such a site. [14](https://arxiv.org/abs/2506.15867) [18](https://arxiv.org/abs/2608.05173) [19](https://arxiv.org/abs/2311.02651) ### Foreign hosted A data center running compute outside the operator’s home jurisdiction. A building’s location doesn’t reveal its customers, workload or legal status. Records and cooperation matter. [7](https://arxiv.org/abs/2403.08501) [20](https://www.iaps.ai/s/AI-chip-smuggling-into-China-final.pdf) [21](https://www.rand.org/pubs/working_papers/WRA3056-1.html) ## Heat paths ### Evaporative towers Water carries heat to cooling towers, which release it to the atmosphere mainly by evaporating part of the water. The visible plume is condensed water droplets. The heat itself is invisible. Weather and plume-abatement design decide whether you see droplets. [2](https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf) [22](https://eta-publications.lbl.gov/sites/default/files/lbnl-1005775_v2.pdf) [23](https://d-nb.info/1241453594/34) ### Dry cooling Fans push air across heat exchangers. Dry coolers and air-cooled chillers release heat without a wet tower. Dry coolers and air-cooled chillers use no water, so there’s no vapour plume to see. Their fan arrays still show from above, and the fan count gives a capacity estimate. [2](https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf) [24](https://newsletter.semianalysis.com/p/datacenter-anatomy-part-2-cooling-systems) [1](https://epoch.ai/data/data-centers-documentation/methodology) ### River / lake outfall Water passes through a heat exchanger and returns to a river or lake warmer than it came in. The surface temperature pattern depends on mixing, flow and discharge layout. Power-plant results don’t set a threshold for data centers. [2](https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf) [25](https://iopscience.iop.org/article/10.1088/1748-9326/8/3/035006) [26](https://www.sciencedirect.com/science/article/pii/S0034425723002584) ### Coastal outfall Seawater takes heat through an exchanger and returns through a coastal outfall. Tides, mixing and intake/outfall design decide what shows at the surface. Discharging at the surface differs from diffusing heat at depth. [27](https://www.nature.com/articles/s41545-021-00101-w) [26](https://www.sciencedirect.com/science/article/pii/S0034425723002584) [28](https://doi.org/10.1016/j.marenvres.2005.09.001) ### Deep / diffused water Deep-water cooling and subsea heat exchangers connect the system to water below the surface. The diffuser shown is a concept. The sources cover different systems. A deep intake doesn’t mean a deep outfall, and infrared sees only the water surface. [16](https://arxiv.org/abs/2609.18824) [17](https://news.microsoft.com/innovation-stories/project-natick-underwater-datacenter/) [29](https://fcs.cornell.edu/Lake_Source_Cooling_Anniversary) ### Ground exchange A buried loop or groundwater system moves heat into the ground. Seasonal systems later recover it or release it. The model shows closed-loop boreholes. Aquifer storage pumps groundwater between wells instead. Both need a long-term heat balance worked out for the specific site. [32](https://geocom.geonardo.com/assets/elearning/6.2.art4.pdf) [30](https://pmc.ncbi.nlm.nih.gov/articles/PMC13353728/) [31](https://doi.org/10.1016/j.apenergy.2025.125858) ### Sensible storage A fixed mass of water or another material warms as it absorbs heat, which delays the need to release it. Storage fills up. Varanto is a planned district-heating cavern. It doesn’t show data-center cooling or storing heat indefinitely. [33](https://www.vantaanenergia.fi/en/about-us/projects/varanto-the-cavern-thermal-energy-storage/) ### Ice / phase change Ice or another phase-change material absorbs heat as it melts. Refrigeration later restores its storage capacity. Evidence supports cooling storage and small prototypes. Recharging releases the stored heat plus the refrigeration energy. [34](https://doi.org/10.1016/j.rser.2026.117124) [35](https://doi.org/10.1016/j.applthermaleng.2023.121598) ### Transported coolant Mobile thermal storage would carry heat away to be released or reused somewhere else. The cited feasibility study looks at delivering industrial waste heat. It doesn’t show continuous data-center cooling by truck. [36](https://www.osti.gov/etdeweb/biblio/965445) ### LNG cold recovery A heat-exchanger network uses the cold released as liquefied natural gas warms and turns back into gas. The cited system is a concept design. Cooling depends on how much gas the terminal handles and on keeping the fluids separated by design. [37](https://doi.org/10.1016/j.rineng.2026.111031) ### District heating Heat pumps raise recovered heat to the temperature a district-heating network needs. Exporting heat needs customers and network capacity. Demand changes with the seasons, so storage or another way to reject heat is still needed. [38](https://www.ramboll.com/projects/energy/meta-surplus-heat-to-district-heating) [39](https://datacenters.atmeta.com/wp-content/uploads/2026/09/Metas-Odense-Data-Center.pdf) [40](https://eu-mayors.ec.europa.eu/sites/default/files/2023-10/2023_CoMo_CaseStudy_Stockholm_EN.pdf) [41](https://link.springer.com/article/10.1186/s40517-026-00383-8) ### Industrial reuse Recovered heat feeds suitable industrial or farming processes, sometimes through a heat pump. The receiving process needs a matching temperature and demand. Absorption cooling still needs a final way to reject heat. [42](https://doi.org/10.1016/j.rser.2023.113777) [43](https://doi.org/10.3390/en14092433) ### Radiative cooling Exposed surfaces radiate heat to the sky through the atmospheric window, adding to conventional cooling. Net cooling depends on humidity, clouds, sunlight and emitter area. The cited study doesn’t show it can be a data center’s only heat sink. [44](https://www.energy-proceedings.org/wp-content/uploads/icae2023/1708960121.pdf) ## Detection methods ### Optical features Optical images show building shape, cooling arrays, generators and construction changes. Clouds, blocked views and coarse resolution hide detail. Counting equipment estimates capacity. It says nothing about the workload. [45](https://fas.org/publication/tracking-hyperscale/) [1](https://epoch.ai/data/data-centers-documentation/methodology) [46] [47](https://docs.google.com/document/d/1hU_TF1A4p0WtYGhsuda6kcYSQa-COM9XiSEv9gabQaM/edit) ### Radar imagery Radar sees structures, terrain and vessels through clouds and at night. Published radar work on infrastructure and ships doesn’t add up to a data-center detector. Radar can’t see a buried hall. [48](https://developmentseed.org/ml-grid-detection/) [49](https://doi.org/10.1038/s41586-023-06825-8) [50](https://sentiwiki.copernicus.eu/web/s1-mission) ### Ground deformation Repeated radar passes can measure how the ground moves toward or away from the satellite. Needs stable radar reflections and repeated passes. It doesn’t work on open water. Ground moves for many reasons, so motion doesn’t identify compute. [51](https://doi.org/10.1016/j.jag.2022.102721) [50](https://sentiwiki.copernicus.eu/web/s1-mission) ### Power infrastructure Substations, high-voltage lines and on-site generators point to how much electrical capacity a site has. Many industries use the same substations and lines. They don’t show actual load or the workload. [45](https://fas.org/publication/tracking-hyperscale/) [19](https://arxiv.org/abs/2311.02651) [48](https://developmentseed.org/ml-grid-detection/) [46] ### Electrical signals Metered power systems can record large load swings and oscillations linked to computing. This evidence comes from inside the grid. Smoothing and other loads make it harder to tie a signal to compute. Classifying workloads remotely isn’t validated. [52](https://arxiv.org/abs/2508.14318) [53](https://arxiv.org/pdf/2407.21783) [54](https://www.nerc.com/globalassets/who-we-are/standing-committees/rstc/whitepaper-characteristics-and-risks-of-emerging-large-loads.pdf) [55](https://sites.ecse.rpi.edu/~vanfrl/documents/publications/pre-prints/20250218_CMLV_DataCenter.pdf) ### Land thermal IR Thermal infrared sensors measure the infrared given off by roofs, pavement, equipment and vegetation. Surface temperature also depends on weather, materials and sunlight. A thermal image can’t directly read server power. [56](https://scienceandglobalsecurity.org/archive/sgs08zhang.pdf) [57](https://arxiv.org/abs/2603.20897) [58](https://www.satellitevu.com/news/satvu-releases-first-of-its-kind-thermal-image-revealing-true-operational-activity-inside-major-u-s-data-centre) [59](https://www.usgs.gov/landsat-missions/landsat-9) [60](https://edcintl.cr.usgs.gov/downloads/sciweb1/shared/co/nli_pecora/pecora_22/presentations/TS_6-1_TIRS%202_MMontanaro.pdf) [61](https://www.eoportal.org/ftp/satellite-missions/iss/ISS-ECOStress_300721/ISS-ECOStress.html) ### Water thermal IR Thermal images of the surface have mapped warm-water outfalls at power plants. Infrared only sees the top of the water, so an underwater plume shows up only where it reaches the surface. Nobody has validated how well this finds data centers. [62](https://beyondparallel.csis.org/thermal-imagery-analysis-of-yongbyon/) [26](https://www.sciencedirect.com/science/article/pii/S0034425723002584) [16](https://arxiv.org/abs/2609.18824) ### Visible vapour Wet cooling may make a visible cloud of water droplets, as documented at power plants. Evaporation can be invisible. Weather and plume abatement decide whether a visible plume forms. [63](https://doi.org/10.3390/rs16071290) [23](https://d-nb.info/1241453594/34) ### Chemical plumes Atmospheric sensors can see emissions from large enough combustion or industrial sources. Needs on-site combustion or a gas release. A data hall, closed coolant loop or LNG terminal alone doesn’t mean there’s a plume. [64](https://amt.copernicus.org/articles/19/6099/2026/) [65](https://www.pnas.org/doi/10.1073/pnas.2317077121) [66](https://www.nature.com/articles/s41586-018-0747-1) [67](https://doi.org/10.3390/s24061881) ### Logistics Trucks, vessel movements and construction work give context about a site. Imagery can miss short visits, AIS vessel tracking is incomplete, and ordinary traffic isn’t specific to compute. [68](https://doi.org/10.3390/rs14071595) [49](https://doi.org/10.1038/s41586-023-06825-8) [16](https://arxiv.org/abs/2609.18824) ### Network activity Traffic analysis has been proposed as a clue. Known fibre routes also help find candidate sites. Fibre shows that a site is connected. Routing, encryption and isolated networks limit what its traffic reveals about the work. [13](https://arxiv.org/abs/2607.22619) [46] [11](https://arxiv.org/abs/2604.04712) ### Local physical sensing Nearby sound, vibration or electromagnetic sensors could add evidence that equipment is running. The cited data-center work is proposed or modelled. Other machinery and how signals travel make it hard to pin a signal on its source. [13](https://arxiv.org/abs/2607.22619) [16](https://arxiv.org/abs/2609.18824) ### Site inspection Inspectors can check the equipment and records they see against a declared facility design. The precedent is nuclear verification. It needs access and trustworthy records. A walkthrough alone can’t show what workloads ran in the past. [12](https://arxiv.org/abs/2408.16074) [6](https://arxiv.org/abs/2304.04123) [3](https://arxiv.org/abs/2507.15916) [Complete references](https://luiscosio.github.io/sieve/references.md)