Summary
The energy hunger of Artificial Intelligence is growing dramatically: The International Energy Agency predicts that electricity consumption at data centers could double by 2030 – to just under 3 percent of global electricity demand, as much as Japan consumes today. An average chatbot query actually consumes only about 0.2–0.3 watt-hours, not the oft-cited tenfold energy compared to Google searches. The real scandal is not the individual query, but the data center arms race on a gigawatt scale and the associated return to fossil energy sources and nuclear power. In Germany, megaprojects like the 200-megawatt facility in Lübbenau are currently emerging, while 12 percent of planned natural gas power plant capacity is allocated to AI supply – a climate policy disaster.
People
- Moritz Metz (AI Infrastructure Specialist, Deutschlandfunk)
- Friederike Hildebrandt (Bits and Trees Network)
- Sam Altman (OpenAI CEO, Project Stargate)
Topics
- Data centers and AI infrastructure
- Energy consumption and climate impact
- Natural gas power plants and fossil return
- Nvidia monopoly and chip dependency
- European sovereignty vs. US dominance
Clarus Lead
The debate over AI energy consumption is marked by contradictory figures: while an individual chatbot query at 0.2–0.3 watt-hours is significantly more efficient than often claimed, overall consumption explodes due to data centers on a gigawatt scale. The core problem lies not in the efficiency of individual queries, but in a global arms race for AI infrastructure that binds enormous amounts of capital, raw materials, and energy – and brings fossil energy sources back into play, even though the energy transition is more urgent than ever.
Clarus Original Work
Clarus Research: Mapping of German AI data centers reveals concentrated megaprojects (200 MW Lübbenau, 480 MW Nierstein, 1.2 GW Dummersdorf speculation) that could mean electricity consumption of 10 percent of the German power grid by 2037 – an increase from currently 4 percent to an estimated 9–10 percent.
Classification: The gap between building costs (3 billion euros 2025) and IT hardware (12 billion euros) reveals a critical dependency relationship: 70 percent of all investments flow back to US chips (primarily Nvidia), while European sovereignty is pursued only rhetorically. The "unholy alliance" between the dying gas industry and AI promises endangers climate goals for 25–40 years.
Consequence: Action-relevant for energy policy: data centers may in future only be built with additional renewable energy sources (not existing ones), waste heat utilization anchored regionally (not dissipated), Nvidia dependency reduced through European chip development.
Detailed Summary
The Data Center Arms Race in Global Context
The USA dominates with 70 percent of global AI computing power. Amazon is investing 118 billion dollars in 2025 for infrastructure alone, Google 85 billion, Microsoft 80 billion. Massive projects are emerging: Amazon's Colossus-2 data center in Memphis (Elon Musk's XII) draws as much energy as Berlin; Meta's Louisiana Hyperion with 2 gigawatts (Berlin + Munich) is supported by three new natural gas power plants. OpenAI's Project Stargate plans 250 gigawatts of computing power in seven years – as much as India's entire electricity consumption, with double CO2 emissions of ExxonMobil.
Germany: Between European Ambition and US Dependency
Germany is Europe's largest data center market, currently consuming 4 percent of electricity for data centers. The Federal Network Agency expects an increase to 10 percent by 2037. Planned megaprojects:
- Lübbenau (Spreewald): 200 MW, 11 billion euros, Schwarzgruppe/Lidl, 100,000 AI chips, groundbreaking late 2025, operational in phases from 2027. Waste heat for 75,000 households, but 90 percent locally unabsorbable (infrastructure mismatch).
- Nierstein (Rhineland-Palatinate): 480 MW, NTT Global, potentially Europe's third-largest data center.
- Wustermarkt (west of Berlin): 204 MW, Virtus, 4 billion euros.
- Brandenburg/Finsterwalde & Barutmark: Amazon double, 200 + 100 MW, 7.8 billion euros.
- Dummersdorf (Mecklenburg-Western Pomerania): 1.2 gigawatt speculation, potentially Europe's largest, could match Berlin's electricity consumption alone.
- Dietzenbach (Hesse): Google, 5.5 billion euros by 2029, AI-specialized.
- Bergheim & Bedburg (Rhineland Mining Region): Microsoft, 3.2 billion euros, BUND protests against 5,000-soccer-field land sealing.
Investments 2025: 15 billion euros (3 billion buildings, 12 billion hardware). But 70 percent of chip spending flows to Nvidia back to the USA – no European value creation.
The Core Problem: Fossil Counteroffensive in AI Guise
Friederike Hildebrandt and Moritz Leiner (Bits and Trees / Urgewald) document: 12 percent of currently planned natural gas power plant capacity in Germany is connected to data center supply. This means 25–40-year operating lifespans for gas infrastructure, even though decarbonization is urgent. Meta in Louisiana is building three natural gas power plants next to its Hyperion facility (3 billion euros for gas), even though natural gas is locally available in unlimited quantities. In the USA, Microsoft is reactivating the defunct Three-Mile Island reactor (America's worst nuclear accident 45 years ago), Google is investing in old nuclear power plants in Iowa and Pennsylvania. This development shows: AI companies are dodging their self-imposed climate goals and indirectly subsidizing fossil and nuclear industries.
Power Supply Bottleneck Rather Than Genuine Scarcity
Germany does not face fundamental energy scarcity, but rather an infrastructure supply bottleneck. Electricity is expensive and grid capacity is limiting. Political solution: the Federal Digital Ministry plans to relax permit regulations and expand the industrial electricity rate (subsidy) to data centers – a lobbying success for operators. New rules (from summer 2025): new buildings must achieve PUE efficiency of at least 1.2, 100 percent renewable electricity from 2027, waste heat utilization mandatory. But these standards lag behind the arms race.
The Myths About Chatbot Energy: Debunked
Viral Claim: "ChatGPT needs 10x as much as Google search" (3 watt-hours vs. 0.3).
Origin & Problems:
- John Hennessey (Alphabet's Chairman, Feb. 2023) spoke of costs, not electricity consumption.
- Costs ≠ electricity (personnel, infrastructure, profit included).
- Referred to Alphabet/Google, not ChatGPT (different models, infrastructures).
- The comparison study cited a blog post from 2009 for Google consumption value – 14 years old.
Current Reliable Numbers (Google, OpenAI, Summer 2024):
- One text query (ChatGPT/Gemini): 0.2–0.3 watt-hours
- Equivalent: 9 seconds of television OR 1 second of oven heating
- Or: 7 seconds of cycling at 150 watts (Fedi/Moritz benchmark)
Criticism of These Figures:
- Only median values (hide outliers)
- Video generation significantly higher (graphic shows 10–100x more)
- Reasoning models (+30x energy for self-reflection, web search)
- Companies report text generation, not video/image/reasoning
- Efficiency gains do not lead to savings measures, but to AI Slob (more AI features everywhere, no absolute reduction)
Energy Intensity by AI Type
| Application | Energy Intensity | Example |
|---|---|---|
| Text Chatbot | Low | 0.2–0.3 Wh |
| Image Generation | Medium-High | Multiple of that |
| Video Generation | Very High | 10–100x text query |
| Reasoning/Agents | Very High | Multiple, non-transparent |
The Decoupling: Growth Without Proportional Use
Classically: Data center infrastructure growth = user growth + feature expansion.
With generative AI: Massive energy infrastructure investments (gigawatt data centers, trillion-dollar spending), but user numbers do not increase proportionally. This is the symptom of speculative arms racing, not genuine need. Friederike Hildebrandt (Bits & Trees): "The decoupling is the central problem – we are building for future hype, not current demand."
Hardware Costs & Nvidia Monopoly
An AI server rack (8 graphics processors):
- Value: ~2 million euros
- Power consumption: up to 140,000 watts (900+ cycling equivalents continuously, 50 full ovens)
- Cooling costs: ~1/3 of total energy (PUE factor 1.5 standard, 1.2 minimum from 2025)
- Supplier: Nvidia (world's most valuable company, 4.5 trillion USD market cap)
The Core Problem: European data centers subsidize Nvidia's US profits. 70 percent of hardware investments leave Germany/Europe without creating local value. European chip development is unrealistic in the short term.
Water Consumption & Local Collateral Damage
Lübbenau waste heat: 75,000 households could be supplied, but 90 percent locally unabsorbable (infrastructure mismatch: data centers in rural areas, heat consumers in distant cities).
Louisiana Hyperion: increases state electricity consumption by 1/3. Local air and water contamination not addressed. Elon Musk's Colossus-2 (Memphis): Illegal air pollution from double generator count (vs. permit), locally 4x increased cancer risk. Affected communities: primarily poor, people of color.
Key Statements
The chatbot energy myths are debunked: A TextGPT query costs 0.2–0.3 Wh, not tenfold Google's amount. The origin of this claim is a distorted 2023 interview (about costs, not electricity) and a 2009 blog post.
The real problem is the gigawatt arms race: 250 gigawatts (Project Stargate), 1.2 gigawatts (Dummersdorf), 2 gigawatts (Meta Louisiana) are not rationally justified by user demand, but by market dominance speculation.
Fossil counteroffensive in climate guise: 12 percent of all planned natural gas power plants in Germany serve AI data centers. Long-term