Imagery works on any place on Earth. It needs no equipment on the ground and no permission. You’ve probably seen the words "satellite imagery" in news articles: it refers to a combination of IMINT (what is in the frame) and GEOINT (where the image is placed in space and time).
Oftentimes, it’s much easier to find the evidence of a datacenter being built than the datacenter itself; think the transformer yard taking shape beside it, new transmission lines running toward the site, cooling equipment staged for installation, and more. The evidence is GEOINT, while the building itself is IMINT.
Imagery becomes geospatial when the object is not the building but what must reach it.
Facilities with this sort of power draw require specialized electricity infrastructure, such as large transmission lines. This electricity infrastructure can easily be seen via satellite; even publicly available satellite imagery reveals transmission lines for facilities like aluminum smelters and desalination plants which have comparable power requirements. It is possible to bury transmission lines, but this construction effort would take a long time and would itself be visible to satellites.
Scher, Abecassis, Barnett, and Abeyta | MIRI (2025)
A city-scale draw needs a city-scale connection, and hiding the connection is itself a slow construction project visible from orbit. This is the argument for GEOINT over IMINT in one sentence: you can disguise a box, but you cannot disguise the fact that a gigawatt arrived somewhere.
What the Analyst Looks For
A large datacenter can look much like a warehouse: a long, windowless building with loading docks. Analysts therefore pay close attention to the equipment around it. A transformer yard indicates a substantial power connection, while rows of generators and fuel tanks suggest provision for backup power. Fences, checkpoints and a wide setback provide further clues, though other industrial sites have these features too.
Krawec, Tracking Hyperscale AI Data Center Growth with Satellite Imagery, Federation of American Scientists, 2026.
Cooling equipment is particularly useful for estimating capacity. By identifying and counting chillers, dry coolers or cooling towers, an analyst can estimate how much heat the facility is equipped to remove. Epoch AI uses this approach to estimate IT power capacity, then combines it with information or assumptions about the chips to estimate compute. The result is an estimate of capacity, not a measurement of how much power the site is actually using.
A Prover could make these features harder to interpret by reusing an existing industrial power connection, sharing backup infrastructure, or enclosing parts of the cooling system. Some cooling designs already leave little equipment visible outside. Enclosing the machinery does not eliminate the need to reject heat, but it can make the cooling arrangement harder to identify.
Epoch AI, AI Data Centers Documentation: Methodology, 2025.
Images taken over time can show when construction began, when cooling equipment appeared and when buildings were added. They can also support a rough estimate of the facility’s capacity. Identifying its use is harder: an AI training cluster may look much like a conventional cloud datacenter, and imagery cannot reveal the workload running inside. Section 2.3.3 explains how to convert power estimates into compute estimates.
Krawec, Tracking Hyperscale AI Data Center Growth with Satellite Imagery, Federation of American Scientists, 2026.
GeoGuessr
A second entrance to this section: not a satellite, but a photograph an employee took. You might have heard of GeoGuessr. Players use the same methodology as Bellingcat, the open-source investigative collective, when they are trying to locate a place of a warcrime or FBI when they are trying to figure where the hostage is being held.
It is also being automated. In 2023 a Stanford team trained PIGEON on Street View panoramas; it put 40% of guesses within 25 km globally, ranked in the top 0.01% of GeoGuessr players, and beat one of the game's best professionals six games out of six.
General-purpose vision-language models are now tested on the same task: Bellingcat ran 500 trials in June 2025 and again in August, and found the leading models competitive with Google Lens on holiday photos with signs and architecture, weaker on remote landscapes, and prone to confident wrong answers. Academic benchmarks now exist for the task at country, city and street scale (GeoBench-style suites; "Where on Earth?", 2025; "From Pixels to Places", 2025).
How the IAEA Came to Use Satellite Imagery
Before 1991, IAEA safeguards focused heavily on declared nuclear material and facilities. The Agency had special-inspection authority, but its routine verification work relied largely on states’ declarations. After the Gulf War, inspections under UN Security Council Resolution 687 uncovered Iraq’s undeclared programme, including electromagnetic enrichment sites at Tarmiya and Ash Sharqat and weapons work at Al Atheer. These activities had gone undetected while declared facilities at Tuwaitha remained under safeguards.
The discovery prompted reforms intended to improve the detection of undeclared activities. These included the Model Additional Protocol, approved in 1997, which provided for broader reporting requirements and inspection access in states that accepted it. Satellite imagery became part of the wider effort to strengthen safeguards, although the Agency did not need an Additional Protocol to use it.
In 2001, the IAEA established its Satellite Imagery Analysis Unit. Imagery helped analysts check declarations, follow changes at known facilities and identify possible undeclared activities, including at sites inspectors could not visit. It supplied evidence to guide further investigation rather than settling a site’s purpose on its own. The Al Kibar case below examines the limits of this approach.
Pabian, Commercial Satellite Imagery as an Evolving Open-Source Verification Technology, Publications Office of the European Union, 2015. Quevenco, Completing the Picture: Using Satellite Imagery to Enhance IAEA Safeguards Capabilities, IAEA Bulletin, 2016.
What Imagery Can and Cannot Settle
Two cases from nuclear verification mark the limits.
Al Kibar, Syria, 2007. After the building was bombed, ISIS located it in commercial imagery and identified it as a reactor. The IAEA's own imagery experts, reading the same images, judged that unlikely. Every signature from the section above had been concealed: a remote canyon, a building partly underground, a false roof and walls that turned a Yongbyon-shaped structure into a plain box, no visible cooling, no visible power lines, no housing. Imagery did not settle the question. Photographs of the inside and outside of the building, obtained by US intelligence and reportedly from Israel, did; the IAEA concluded in 2011, years after the strike, that the building had very likely been a reactor. The lesson is not that imagery failed. It is that a Prover who knows the checklist can build against it, and that two competent analysts can read the same frame in opposite ways.
Turquzabad, Iran, 2018. A rival state's seized archive supplied coordinates. Commercial imagery then showed the site being emptied, container by container, over months. Environmental samples settled what had been there. Imagery did not find the site and did not identify the material; it established the timeline of the cover-up once someone else had said where to look.
The pattern in both: imagery is decisive about sequence and location, weak about identity, and it works best when another layer has provided the target. Read 2.3.1's four roles again with that in mind. Whether commercial imagery is a net gain for verification or a liability (shutter control, physical limits, AI-generated imagery) is one of the debates in 2.3.7.
The Al Kibar Reactor: Extraordinary Camouflage, Troubling Implications
Read the introduction (pp. 1–3) and "Summary and Lessons" (pp. 28–30); the sections between them are annotated images and you can skim them.
Albright and Brannan | Institute for Science and International Security (2008) | 12 min
Introducing the Frontier Data Centers Hub
Epoch's description of its own method: satellite images (cooling equipment counted and measured on the roof), building permits, and public documents, combined into power, compute, and construction timelines for thirteen US datacenters, about 15% of the world's delivered compute as of late 2025. Note the stated uncertainty: the cooling-based estimates can be off by a factor of two in either direction. The full method is on the hub's methodology page.
Epoch AI (2025) | 5 min