The energy of an answer
by Alexander Moshchev
When you type a question into ChatGPT and receive a response, the exchange consumes approximately 0.3 to 0.4 watt-hours of electricity. According to OpenAI’s Sam Altman, an average query uses around 0.34 Wh[1], while independent measurements of GPT-4o recorded approximately 0.43 Wh for a typical prompt[2]. To put this in perspective, that’s comparable to powering a 10-watt LED bulb for two to three minutes – a modest amount for a single interaction.
The picture changes when these individual queries accumulate. A million AI responses consume between 300 and 400 kilowatt-hours. Based on ADAC testing data, this equals the energy a Tesla Model 3 would use to travel approximately 2,300 kilometres – roughly the distance from Berlin to Barcelona[3]. Scale up to ten million responses, and the consumption reaches 3 to 4 megawatt-hours, approaching the annual electricity usage of a typical three-person household in Germany (approximately 3,500 kWh)[4].
These figures represent the computational energy alone. Data centres require additional power for cooling systems, power distribution, and infrastructure operations. Leading facilities operate with a Power Usage Effectiveness (PUE) of around 1.1, meaning the total energy draw exceeds the computational requirement by approximately 10 percent[5]. As AI adoption accelerates across billions of daily interactions, these individually modest energy costs compound into a consideration for electrical grids worldwide.
Behind the computation: What drives energy use
The energy required for an AI response stems from intensive mathematical operations performed by specialised processors – GPUs and TPUs – that execute billions of calculations to generate each word. Three primary factors determine energy consumption:
Processing volume: Longer prompts and responses require more computation. Modern transformer architectures use an “attention” mechanism that compares each new word against all previous ones, causing energy requirements to grow with context length.
Model architecture: Dense models activate all parameters for every word, while Mixture-of-Experts (MoE) architectures activate only specialised subsets, reducing computation. For example, Qwen’s MoE design uses only 3 billion of its 80 billion parameters for each response.[6]
Hardware efficiency: Energy consumption varies based on batch size and optimisation strategies – counterintuitively, systems optimised for fastest responses often use more energy per query than those processing larger batches.
Beyond the processors: Data centre infrastructure
The computational energy represents only part of the total consumption. Data centres require substantial additional power for cooling systems, power distribution, lighting, and other infrastructure operations. The industry quantifies this overhead through Power Usage Effectiveness (PUE) – the ratio of total facility power to IT equipment power. A PUE of 1.10 means an additional 0.10 kWh powers infrastructure for every kilowatt-hour consumed by servers.
Leading hyperscale operators achieve fleet-average PUE values around 1.1, while the industry average remains at approximately 1.56 due to older, less efficient facilities[7]. This disparity means an AI query consuming 0.40 Wh at the processor level translates to 0.44 Wh in a state-of-the-art facility but 0.62 Wh in an average data centre.
Infrastructure energy consumption has declined from approximately 40% of total data centre electricity in 2014 to 30% in 2023, driven by enhanced PUE and migration to more efficient hyperscale centres. Despite these improvements, infrastructure overhead remains substantial.[8]
Environmental impact extends beyond energy efficiency. Carbon intensity depends on regional grid mix, while water consumption varies with cooling methods – evaporative systems use more water but less energy than dry cooling. These location-specific factors explain why credible scenarios incorporate regional variations in cooling approaches and grid composition.
The multiplication problem
Google processes ~14 billion searches daily (increasingly AI-enhanced)[9], GitHub Copilot writes ~46% of code for its 15+ million developers[10], and every smartphone assistant runs AI. Each interaction uses just 0.3-0.4 Wh, but billions of daily queries across thousands of services create power consumption that rivals small countries.
The IEA projects this could reach 945 TWh by 2030 – nearly 3% of global electricity.[11] The real number depends on three multipliers: energy per answer (falling as chips and models improve), usage volume (surging as features spread), and where and when the load lands on the grid (setting emissions and local strain).
Mapping the future: IEA’s energy scenarios
The International Energy Agency provides the most comprehensive modelling of data centre electricity demand through 2035. Rather than offering a single projection, the IEA presents multiple scenarios that adjust three primary variables – energy efficiency per query (Unit), adoption rates of AI services (Volume), and geographical distribution of data centres (Venue) – to establish a credible range of outcomes.[12]
Base case trajectory
Under the IEA’s central scenario, global data centre electricity consumption rises from approximately 415 TWh in 2024 (1.5% of global electricity) to 945 TWh by 2030, approaching 3% of worldwide electricity demand. This growth stems primarily from AI-accelerated servers, whose electricity consumption expands at roughly 30% annually, compared to 9% for conventional servers. While this represents less than 10% of global electricity demand growth, the concentration in advanced economies tells a different story – these regions see data centres accounting for over 20% of new electricity demand, explaining current grid tensions in technology hubs.[13]
The efficiency-adoption spectrum
Looking toward 2035, the scenarios diverge significantly based on technology and policy developments. A high-efficiency pathway, assuming aggressive improvements in PUE, hardware efficiency, and model optimisation, limits consumption to approximately 970 TWh (2.6% of projected global electricity). Conversely, a “lift-off” scenario with accelerated AI adoption and streamlined infrastructure deployment could see consumption exceed 1,700 TWh (4.4% of global electricity). This nearly twofold difference underscores how technological choices and deployment patterns will determine actual outcomes.[14]
The IEA estimates that approximately one-fifth of planned data centre capacity could face delays due to electrical grid interconnection queues and equipment lead times, suggesting that infrastructure limitations may govern near-term expansion in many regions. The span between 970 TWh and 1,700 TWh reflects different plausible trajectories for efficiency improvements, adoption rates, and geographic distribution – each determining not just total consumption but regional impacts and emissions intensity.[15]
Continue reading in Part 2!
Footnotes
[1] Altman, S. (2025). The Gentle Singularity. https://blog.samaltman.com/the-gentle-singularity
[2] Jejjam, Mohammad, et al. “How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference.” arXiv preprint arXiv:2505.09598 (2025). https://arxiv.org/abs/2505.09598.
[3] ADAC. “Tesla Model 3 – Verbrauch und Kosten.” ADAC Ecotest, 2024. Real-world energy consumption benchmarks: ~17–19 kWh/100 km. https://www.adac.de/rund-ums-fahrzeug/autokatalog/marken-modelle/tesla/model-3/1generation-facelift/329535/
[4] ADAC/Stromspiegel. “Stromverbrauch im Haushalt.” 2023–2024 https://www.adac.de/rund-ums-haus/energie/spartipps/stromverbrauch-im-haushalt
[5] Google. “2024 Environmental Report.” Google Sustainability, https://sustainability.google/reports/google-2024-environmental-report
[6] Qwen Team. “Qwen3-Next-80B-A3B.” Qwen Documentation, 2025 https://qwen.ai/blog?id=4074cca80393150c248e508aa62983f9cb7d27cd&from=research.latest-advancements-list
[7] Uptime Institute. “Global Data Center Survey Results 2024.” Uptime Institute, 2024. https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2024.GlobalDataCenterSurvey.Report.pdf?version=0
[8] U.S. Department of Energy. Referenced in Lawrence Berkeley National Laboratory, “2024 Report on U.S. Data Center Energy Use,” 2024. Infrastructure energy consumption declined from ~40% (2014) to 30% (2023) of total data centre electricity. https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report_1.pdf
[9] SparkToro. “New Research: Google Search Grew 20%+ in 2024.” SparkToro Blog, 2024. Google saw more than 5 trillion searches in 2024, approximately 14 billion per day. https://sparktoro.com/blog/new-research-google-search-grew-20-in-2024-receives-373x-more-searches-than-chatgpt/
[10] GitHub (2024). Measuring Impact of GitHub Copilot. Copilot contributes 46% of code on average across 15+ million developers, with 88% acceptance rate. https://github.blog/ai-and-ml/github-copilot/github-copilot-now-has-a-better-ai-model-and-new-capabilities/
[11] International Energy Agency. “Energy and AI.” IEA Special Report, 2025. Comprehensive modelling of data centre electricity demand through 2035; Unit × Volume × Venue framework. https://www.iea.org/reports/energy-and-ai.
[12] Same as 11
[13] Same as 11
[14] Same as 11
[15] Same as 11