Artificial Intelligence on Physical Ground: Electricity, Water and the Cloud

Error rates began to climb at 11:49 pm Pacific time on 19 October 2025, and Amazon Web Services did not say that everything had returned to normal until after 3 pm the following day, roughly fifteen hours later. According to the company, the cause was a problem with DNS resolution for the regional endpoints of its DynamoDB database service in the US-EAST-1 region. One fault in one region hit many services, a reminder that what we call "intelligence" stands on a heavy structure.

Three pillars of iron and electricity

AI needs data centres packed with servers, advanced chips, and power and cooling. The International Energy Agency estimates that data centres consumed about 415 terawatt-hours in 2024 (about 1.5% of the world's electricity) and, in its base case, expects this to roughly double to about 945 terawatt-hours by 2030. In a January 2024 report the Agency had estimated that consumption, after about 460 terawatt-hours in 2022, could exceed 1,000 terawatt-hours in 2026; estimates differ as assumptions differ, and projections are not promises.

Water and cooling

Cooling has a water cost. Researchers estimated that GPT-3 "drinks" a 500 ml bottle of water for roughly 10 to 50 medium-length responses, depending on the location of the data centre and conditions, and that training it in American data centres could evaporate about 700,000 litres; the authors describe these as conservative estimates, and they do not apply literally to every model or country.

When the foundation is concentrated

The cloud market itself is concentrated: in the second quarter of 2026, Amazon, Microsoft and Google held 63% of it, according to Synergy. In chips, Taiwan's TSMC makes more than 90% of the world's most advanced semiconductors, according to an analytical comment from King's College London. The fewer the points of dependence, the wider the effect of a fault at any one of them.

AI adds a technical twist of its own: losing a particular model is not like losing a kilowatt of electricity. An organisation may have built its operations around that model's behaviour, and no other model can stand in overnight. This is our own assessment, not a published figure.

The subtler danger: a failure no one sees

When electricity fails, everyone knows. A model may instead degrade and keep working at lower quality without warning. Translation or summarisation quality may slip after an update or under server load, and the user may not notice for some time. So the most useful thing to build is likely a design that reveals its incapacity instead of hiding it, with regular quality checks on samples that humans review.

What can be done?

• Diversify providers and models, and prepare the ability to switch between them when needed.
• A minimum of local capacity sufficient to keep critical services running if the connection abroad is cut.
• Safe degradation: when hit, the system falls back to a simpler, safer mode.
• Transparency: the system declares when it cannot cope.
• Human skill: the organisation keeps the ability to work manually if the tool is gone.

Trust in God does not replace taking precautions. Whoever builds a system that people depend on should ask, before feeling reassured: what happens when it fails?