A cluster turns electricity into heat at close to 100% efficiency, and both ends of that conversion are measurable from outside: the power has to come from somewhere, and the heat has to go somewhere. Measurement and signature intelligence (MASINT) reads those two quantities and estimates how much compute is running here.
A watt measures how quickly energy is used: one joule per second. A 2 kW kettle converts 2,000 joules of electrical energy into heat every second. A computing cluster drawing 130 MW consumes 130 million joules per second. Watts describe a rate, not a total amount of energy.
A FLOP is one arithmetic operation, such as an addition or multiplication, on floating-point numbers—the computer’s way of representing values such as 0.25 or −3.7 across a wide range of magnitudes. FLOP count the arithmetic performed; FLOP per second measure how fast it happens.
Think of a language model as an enormous collection of adjustable numbers, called parameters. During pretraining, it processes text divided into tokens—words or pieces of words—and predicts what comes next. Its prediction errors guide small adjustments to the parameters. This process repeats across the training data, teaching the model patterns that help it make better predictions.
Most of the arithmetic involved is matrix multiplication. For a conventional dense transformer, the training cost can be approximated as:
The factor of six accounts for both making predictions and calculating how the parameters should change. It is an approximation, not an exact count for every model architecture (Kaplan et al., 2020).
This gives two main ways to spend more compute: train a larger model, or train it on more text. More parameters make processing each token more expensive; more tokens mean more processing overall. The 2022 Chinchilla study found that, to get the best performance from a fixed training budget under its experimental conditions, model size and training data should grow in roughly equal proportions. Following that rule, ten times more compute would mean about 3.2 times as many parameters and 3.2 times as many tokens (Hoffmann et al., 2022).
The connection to electricity is physical. Arithmetic in a chip requires switching transistors, which charges and discharges tiny electrical capacitances and dissipates energy as heat.
For a rough hardware benchmark, an NVIDIA H100 SXM has a power rating of up to 700 W and a peak throughput of approximately dense, 16-bit floating-point operations per second. Dividing power by throughput gives an energy cost on the order of joules per FLOP. This is a peak-based GPU estimate, not a measurement of an entire training facility (NVIDIA H100 datasheet).
Using that illustrative efficiency, a training run requiring FLOP would consume:
Spread over three months, that corresponds to roughly 1.3 MW of average power under those assumptions. A facility actually drawing 130 MW for three months would instead consume about 280 GWh—a hundred times as much energy.
The hundredfold gap between the two examples is the difference between a 2023 frontier run and a 2025–26 one. Frontier training compute has grown 4 to 5 times a year, and after efficiency gains and longer runs are netted out, the power a frontier run draws has still grown 2.2 times a year (You and Owen, 2025). The signature scales with the run, and the run has been getting bigger faster than the chips have been getting cheaper to feed.
Heat
Almost all the electricity a data centre consumes ends up as heat. A large training run can keep chips under a fairly constant load for weeks, so the cooling system has to remove roughly the same amount of heat, day and night. Inference often varies with user demand. A sustained thermal signature can therefore be evidence of continuous, intensive use—not just that a site has the capacity for it. However it is important to note that a Prover may disguise energy consumption as other high-energy activities and obtaining detailed energy consumption data can be hard.
Read the release.
SatVu (2025) | 3 min
Commercial thermal satellites such as HotSat can resolve individual buildings. SatVu has published thermal imagery of an operating US data centre showing which cooling equipment, substations and electrical yards appear active. There are simpler indicators too: Baker et al. note that roofs where snow does not settle in winter can reveal waste heat (footnote 127). That may be visible in ordinary optical satellite imagery, without a thermal sensor.
But heat does not tell us what the chips are doing. It provides evidence of activity and a rough indication of its scale, not a reliable distinction between training and inference. A Prover could run inference between training runs to keep power consumption—and heat output—roughly constant. The workload would change; the thermal signature would not necessarily show it.
Verification Methods for International AI Agreements
Read the "Energy monitoring" entry under national technical means: the description, the historical precedents, and the evasion techniques. Note what they mean by "coarse": the method can catch a facility drawing far more than it declares, and cannot catch a run that fits inside an authorised facility's normal draw. The evasion list is the one to keep.
Wasil, Reed, Miller, and Barnett | arXiv (2024) | 5 min
Optional: Mechanisms to Verify International Agreements About AI Development
Read the national-intelligence building blocks and the locating-compute appendix (pp. 67–80), the source of the feasibility ratings quoted in the module and of the feasibility cards in 2.3.6.
Scher and Thiergart | MIRI (2024) | 30 min
Optional: xAI is facing a lawsuit for operating over 400 MW of gas turbines without permits
How the Memphis case was established: aerial photographs commissioned by the Southern Environmental Law Center in March 2025 showed 35 turbines around the Colossus site; thermal images about a month later showed at least 33 of them running; a June flight counted 26. The permit issued in July 2025 allows 15 generators, 247 MW.
Tim De Chant | TechCrunch (2025) | 5 min
Interactive: The Power Signature
The explainer below is built from Epoch's datacenter dataset. It places a real facility beside national grids for scale, shows which features of its load a watcher can and cannot read, and projects how fast the signature shrinks as performance per watt improves.