Heatwaves, droughts, dying crops, wildfires. Big Tech: more AI data centres, please

Europe has spent much of this summer being absolutely fucking cooked. Western Europe recorded its hottest June on record, averaging 3.05°C above the 1991-2020 June average. Germany had 252 weather stations break their all-time temperature records during June alone, while widespread dryness and increasingly parched soils were already worsening drought and wildfire conditions across the continent. (World Meteorological Organization)

The UK hardly escaped it. Temperatures reached 38.1°C at Kew Gardens on 13 August, making it the fifth-hottest day ever recorded in the UK. (Met Office) By 10 August, 71.3% of England was officially in drought, covering around 45 million people, with 27 million already living under water-use restrictions after an exceptionally dry summer. (Environment Agency / GOV.UK)

The damage spreads far beyond everyone sweating their bollocks off for another week. Across Europe, drought and heat have hammered agriculture, river transport, power generation and ecosystems. The EU's Joint Research Centre cut its estimates for grain maize and sunflower yields by around 6-7% in July, while persistent heat and lack of rainfall continued to hammer crop development across France, central Europe, Hungary and western Romania. By 5 August, 505,683 hectares had burned across the EU, already well ahead of the 379,392 hectares burned by the same point in 2025. (EU Joint Research Centre)

So while reservoirs shrink, crops die, electricity systems strain under cooling demand and huge parts of Europe catch fire, some of the richest companies on Earth are simultaneously demanding vastly more electricity and infrastructure for AI data centres.

Just a few hundred billion more in AI infrastructure, please:

This isn't some modest expansion of ordinary cloud computing. Amazon, Microsoft, Alphabet, Meta and Oracle are projected to spend roughly $750 billion in capital expenditure during 2026, with AI infrastructure responsible for much of the enormous increase in spending. (S&P Global)

Electricity demand is following the money. The IEA expects global data-centre electricity consumption to rise from about 485 TWh in 2025 to roughly 950 TWh by 2030. (International Energy Agency) For some sense of scale, total UK electricity demand was 323 TWh in 2025. (UK government, DUKES 2026) Within four years, global data centres could therefore be consuming close to three times the electricity currently used by the whole fucking United Kingdom every year.

Even the figures attached to individual projects are easy to read without appreciating what they actually mean. The UK government's own AI Growth Zone criteria are designed around sites capable of supporting at least 500 MW of AI infrastructure, with applicants expected to demonstrate sufficient access to both electricity and water at that scale. (UK government)

A continuous 500 MW load consumes 4.38 TWh in a year. Using Ofgem's benchmark of around 2,500 kWh of annual electricity consumption for a typical household, that's roughly the yearly electricity use of 1.75 million British homes. (Ofgem) A sustained 1 GW campus would consume 8.76 TWh, around 3.5 million homes' worth of annual electricity.

These centres will not all sit permanently at their maximum rated load, but the comparison shows why hundreds of megawatts cannot be treated like another meaningless corporate specification. These are industrial loads on the scale of entire cities, arriving while countries are also trying to electrify transport and heating, replace fossil generation, strengthen ageing grids and cope with increasingly extreme summer cooling demand.

There apparently wasn't enough pressure on the system already.

And what exactly needs all of this?

There are useful applications for machine learning and GenAI. Scientific research, engineering, coding assistance, translation and genuinely useful automation can justify serious amounts of compute. That does not explain the extraordinary volume of infrastructure now being committed to every marginal use the industry can dream up.

Automated advertising generation, synthetic marketing material, summaries of documents and emails already sitting in front of you, generated search answers, customer-service chatbots, endless image and video generation, agent systems repeatedly burning inference on trivial office tasks, and generative features being shoved into products whether users demonstrated any meaningful need for them or not are all part of the same demand explosion.

The result is an industrial-scale commitment of electricity, cooling equipment, chips, water and infrastructure to making disposable digital content dramatically cheaper to manufacture, despite humanity already having effectively infinite supplies of advertising, marketing copy, generic images, corporate filler and searchable information.

Even cleaner electricity does not make the resource question disappear. The IEA expects renewables to provide around half of the additional electricity required by data centres through 2030, but natural gas and coal together are still expected to supply more than 40% of the additional demand. Electricity-related CO₂ emissions from data centres are projected to peak at around 320 million tonnes a year by 2030 in the IEA's base case. (International Energy Agency)

There is also the opportunity cost. Build another wind farm or power station because an AI campus has created a huge new load and that generation is now servicing demand that did not previously exist. It cannot simultaneously replace an existing fossil load, supply new housing, electrify transport or provide additional resilience during extreme weather.

Efficiency improvements help, but cheaper and more efficient compute has generally created room for companies to run more compute, not an excuse to stop building.

Sorry, we're using that water:

The UK provides an almost comically good example of what happens when the PowerPoint version of AI growth meets an actual physical resource.

The government is actively encouraging AI infrastructure through AI Growth Zones, with enhanced access to power and planning support intended to accelerate development. Yet its own April 2026 National Framework for Water Resources acknowledges that some water companies have already refused applications from data centres because supply cannot be guaranteed. Some catchments are already closed to further abstraction, while officials admit they are struggling to obtain enough information from operators to accurately model future data-centre water requirements. (UK government, National Framework for Water Resources)

The same framework expects a large and rapid increase in English data centres before 2030 and explicitly acknowledges that many could be built before the major new strategic water resources intended to increase supply have come online. That's quite a planning achievement when nearly three-quarters of England has just spent August officially in drought.

The sequence is almost absurdly neat: accelerate AI data centres, discover they need huge quantities of water, discover the water isn't there, get applications rejected, realise the major new supplies will not be ready in time, then start looking for recycled water, alternative abstraction and more efficient cooling so the original expansion plan can continue anyway.

Money stops behaving like a cheat code at this point. A trillion-dollar company can finance pipelines, cooling systems and reservoirs, but it cannot order rainfall. If a catchment is exhausted, the reservoir is depleted and households and agriculture already need the available supply, putting another zero on the investment proposal does not create more water.

The grid has discovered it can say no as well:

The electricity side has reached the same confrontation, with this summer providing a particularly nasty demonstration of why.

During the July heat dome in the eastern United States, PJM, the country's largest electricity-grid operator, had to activate emergency measures as air-conditioning demand surged. Its operating reserves collapsed from around 22 GW in the morning to about 5 GW by evening, approaching its 3.2 GW minimum requirement. (Reuters) Peak demand reached roughly 162.7 GW, close to PJM's all-time record of 165.6 GW, and the operator later said its emergency demand reductions probably prevented a new record from being set. PJM has been struggling with soaring demand driven primarily by the expansion of data centres. (Reuters)

Then came the capacity auction. Prices hit their cap and PJM still ended up roughly 6.8 GW short of the capacity needed to meet its reliability requirement. (Reuters) A 6.8 GW shortfall is not some rounding error: sustained for a year, that amount of power would equal the annual electricity use of roughly 24 million typical British homes using the same Ofgem benchmark.

PJM's latest proposal finally makes the priority question explicit. During a severe shortage, utilities should be able to reduce or transfer electricity demand from data centres and other enormous users before cutting power to ordinary households, potentially forcing the centres onto their own backup generation. (Reuters)

Once a heatwave drives demand high enough and there isn't enough supply for everyone, the argument stops being abstract. Somebody gets the electricity and somebody does not.

Increasingly, the answer is not automatically the data centre.

New York has imposed a one-year moratorium on new data centres using 50 MW or more while environmental standards are developed. (Reuters) Texas Governor Greg Abbott has ordered a pause on new data-centre grid approvals pending an audit after ERCOT accumulated around 474 GW of proposed new electricity demand, more than five times the state's record peak load, with roughly 90% of those requests coming from data centres. (Reuters)

Denmark has proposed putting new data centres behind households, healthcare, ordinary industry, transport and renewable projects in the grid-connection queue. Amsterdam has barred new centres and expansions until at least 2030, while Dublin spent years effectively blocking new connections and now requires new centres to bring their own power generation. (Reuters)

Pennsylvania has now gone further as well, removing data centres from its Fast Track permit programme while demanding stronger environmental safeguards, transparency and local community approval for future projects. (Reuters)

This is governments and grid operators being forced to establish priorities that should never have been mysterious in the first place. Homes, hospitals, transport and existing industry cannot simply disappear because another hyperscaler wants a giant new block of capacity immediately.

Money cannot hyperscale a river:

For decades, the largest technology companies operated in a world where most scaling problems really could be solved with capital. More users required more servers. More storage required more hardware. More traffic required more network capacity. If Amazon, Microsoft or Google underestimated future cloud demand, another campus and several billion dollars could solve a lot of the problem.

AI pushes that chain far beyond their own systems. More compute now means more data centres, which means more generation, substations, transmission, transformers, cooling, water, land, construction labour and planning permission. Some of those systems take years to expand. Others depend on resources that simply are not available in unlimited quantities.

A company can pre-order GPUs and sign twenty-year energy contracts. It cannot compress a multi-year grid upgrade into a few months because its next model launch is approaching. It cannot make a drought end because a training cluster has already been installed. Once enormous private demand collides with resources required by everybody else, the decision is no longer exclusively Big Tech's to make.

That's the wall the industry is starting to hit now. The limits are physical, increasingly political, and considerably harder to brute-force than another software bottleneck.

The timing could hardly be worse. Europe has just lived through record heat, worsening drought, damaged harvests and huge wildfires while governments are already spending more money adapting infrastructure to exactly these conditions. At the same time, society is being asked to reserve extraordinary amounts of new electricity, water, construction capacity and grid investment for an AI arms race whose participants are spending hundreds of billions partly because none of them wants to risk being the company that slows down first.

Amazon, Microsoft, Alphabet, Meta, Oracle and everyone chasing them can each decide that continuing to build is rational because their competitors are doing the same thing. The physical consequences do not remain neatly separated by corporate balance sheet.

Everyone shares the grid. Everyone shares the water system. Everyone lives with the emissions.