Big Tech’s shadow power grid
AI guzzles vast amounts of energy. That’s why hyperscalers like Amazon, Google and Microsoft are building their own power plants – relying mainly on fossil fuels and nuclear power.
Amazon is planning a gigantic gas power plant in Texas to supply AI data centres with electricity. It’s set to have a capacity of 7.65 gigawatts, making it the largest natural gas power plant ever built in the US. For comparison: Switzerland’s four nuclear power plants together generate just under three gigawatts.
The authorities in Pecos County have approved emissions of up to 33 million tonnes of CO₂ per year. If this ceiling were to be fully used, the plant would be the largest single source of pollution in the US, exceeding even the largest coal-fired power plant. This is according to research by Cleanview, which established the link between the project – called GW Ranch – and Amazon.
Source: Render GW Ranch / Pacifico Energy
The case is a prime example of AI’s hunger for energy. It shows how tech companies are starting to build their own energy infrastructure – with far-reaching consequences for power systems, the climate and society.
How much electricity does AI really use?
At first glance, figures on the electricity consumption of individual AI queries look reassuring. A year ago, Google put the energy cost of a text query to Gemini at 0.24 watt-hours (Wh) (linked article in German). That’s the equivalent of nine seconds of watching TV. OpenAI puts the figure at 0.34 Wh. Other calculations go as high as 1 Wh.
Complex or agentic queries, however, quickly ramp up several times that amount. Images and videos are true power hogs too. MIT Technology Review did the math on AI, coming to the conclusion that generating a five-second video with an open-weight model takes almost one kilowatt-hour (1,000 Wh) – as much as a microwave running continuously for an hour. On top of that comes the energy needed to train new models.
All these figures should be taken with a pinch of salt. Although they may hold true for specific AIs and specific infrastructure, the efficiency of models and chips varies enormously. Depending on the number of parameters and servers used, energy needs can be considerably lower or higher. To assess it accurately, Google, OpenAI and Anthropic would need to disclose their true total consumption. Which they don’t do. So all that’s left are back-of-the-envelope calculations and approximations.
By 2030, data centres are set to consume an additional 515 TWh of electricity. That’s as much as the whole of Germany.
The International Energy Agency (IEA) estimates that data centres of all kinds consumed around 415 terawatt-hours (TWh) of electricity worldwide in 2024. By 2030, this demand could more than double to 945 TWh. AI is a key driver. For comparison: Switzerland currently consumes just under 60 TWh of electricity a year. The US’s annual consumption stands at around 4,200 TWh.
Source: International Energy Agency
A Greenpeace study (in German) also expects the electricity demand of AI data centres to be around eleven times higher by 2030 than in 2023. That would be roughly equivalent to today’s electricity consumption of all conventional server farms. In Switzerland, data centres are estimated to account for 6 to 8 per cent of electricity consumption today (link in German). By 2030, this share could rise to 15 per cent.
Fossil fuels and nuclear power
Amazon’s Texas project isn’t an isolated case. Tech companies are increasingly securing their own, large-scale industrial power sources with the aim of decoupling themselves from the public grid. Renewable energy is only of limited use, because server farms need electricity around the clock, which would require enormous storage capacity. The simpler, faster route relies on fossil fuels. Nuclear power is in high demand too.
Microsoft has signed a 20-year supply agreement with US utility Constellation Energy. In return, the Three Mile Island Unit 1 reactor in Pennsylvania is set to restart, supplying Microsoft’s data centres with nuclear power. The Trump administration has already approved a USD 1 billion loan for this and granted the relevant approval.
Source: Wikimedia Commons / formulaone
Google is banking on a new generation of small modular reactors (SMRs) (link in German). The Alphabet subsidiary has signed a contract with US start-up Kairos Power to bring a first SMR online by 2030, with more set to follow by 2035. Amazon is also involved in several SMR projects and has secured power purchase rights at a Talen Energy nuclear plant in Pennsylvania. Oracle is planning a data centre that’s set to be powered entirely by three SMR modules.
Data centre operators have announced their own power plants with a combined capacity of 90 gigawatts.
Meanwhile, more large gas power plants are being built. Alongside Amazon’s GW Ranch project, Microsoft is building a 2-gigawatt gas power plant with Chevron in the same Texas county. Amazon is also negotiating a further 4.5-gigawatt gas power plant in Pennsylvania (link in German).
Far-reaching ecological consequences
AI’s hunger for energy has direct consequences for the climate and the environment. In its study (in German), Greenpeace calculates that emissions from data centres of all kinds could rise from the 29 million tonnes of CO₂-equivalent in 2023 to 166 million tonnes in 2030, even with a growing share of renewable energy.
Source: Shutterstock
A UN report also points out (in German) that data centres are becoming a significant factor in global resource consumption. They don’t just cause CO₂ emissions – they also draw on large volumes of water and generate significant e-waste through short-lived server and chip generations.
Particularly problematic: AI data centres use more water than conventional server farms and are often located in dry regions (link in German). The UN report puts 2025 demand at 4.5 trillion litres. By 2030, it could rise to 9.3 trillion litres. However, it’s also true that other industries need a lot of water too. US golf courses alone use 2.7 trillion litres a year.
Manufacturing the hardware causes emissions too. AI chips are produced using large amounts of energy – often from coal power. As such, a large AI model’s carbon footprint is already substantial before it even answers its first query.
At the same time, the industry is working on efficiency gains. Swiss start-up Corintis, for example, is testing liquid cooling directly on the chip to cut electricity and water use and make waste heat usable (link in German). With cooling channels that are tailored to the thermal structure of individual chips, cooling energy needs can reportedly be cut by around 20 per cent. But efficiency doesn’t necessarily solve the underlying problem. The more efficient and cheaper a technology gets, the faster its use grows – and with it, often, absolute consumption.
Growing resistance
The massive expansion of AI infrastructure is meeting an increasingly critical public. In Switzerland, a representative survey (in German) shows that a large majority of the population views data centres’ electricity and water consumption critically. 72 per cent say new data centres should only be built if they run on renewable energy. 79 to 80 per cent want mandatory transparency on energy consumption and environmental impact.
Other surveys show that in the US, AI data centres are extremely unpopular, with Republicans and Democrats united on the issue (link in German). Cross-party alliances are fighting facilities in residential areas. Local residents are protesting against data centres out of concern they may cause noise, use a lot of water and pollute the air with their power plants. The state of New York has already imposed a one-year moratorium on building large data centres.
Source: Shutterstock
The AI industry, on the other hand, warns against blocking data centre expansion, arguing that AI drives productivity gains, higher incomes and progress in important research. Excessive regulation or blanket building bans could jeopardise these opportunities – and, geopolitically, cause regions with strict regulation to fall behind technologically.
Only 802 of the 3,969 announced data centres are currently under construction.
Many of the new data centres and power plants are, so far, just hot air. Which of them will actually get built is anyone’s guess. Construction delays aren’t unusual for such projects. According to Goldman Sachs, though, 72 per cent of data centres have historically gone online on time. For the next two years, the US bank forecasts that at most half of the promised servers will become reality. As of August 2026, construction has only actually started on just over 800 of the 4,000 planned data centres – and in many cases, that means nothing more than a few diggers and scaffolding so far.
Demand isn’t set in stone
The big energy projects from Amazon and the like are based on the scenario of explosive AI demand. But these forecasts are disputed. Investors are heavily subsidising growth so far. They’re burning through vast amounts of venture capital in OpenAI’s and Anthropic’s server furnaces. It’s by no means certain that demand for chatbots will stay this high once customers are eventually made to pay the real costs – for both development and operation.
This risk seems particularly high for the US hyperscalers. They’re building their servers largely for Anthropic and OpenAI. But Chinese open-weight models are close on the heels of the two market leaders and becoming ever more popular. Even in the US, nearly half of corporate customers’ tokens now go to DeepSeek, Qwen and other AIs from China. In Africa, they overtook Claude and ChatGPT long ago.
State regulation or public opposition could also throw a spanner in Big Tech’s works. What happens to the planned gas power plants, SMRs and reactivated nuclear reactors if demand turns out lower than expected remains to be seen. Such facilities are designed to run for decades. They could end up as fossil and nuclear «lock-ins» that make climate targets harder to reach in the long run, AI boom or not.
My fingerprint often changes so drastically that my MacBook doesn't recognise it anymore. The reason? If I'm not clinging to a monitor or camera, I'm probably clinging to a rockface by the tips of my fingers.
Interesting facts about products, behind-the-scenes looks at manufacturers and deep-dives on interesting people.
Show all
