Executive Summary
AI’s energy consumption has emerged as one of the most critical sustainability challenges in modern technology. Global AI data center power draw reached approximately 50 TWh in 2025, equivalent to the entire electricity consumption of Sweden, and the International Energy Agency projects this will nearly triple to 130-150 TWh by 2028. A single ChatGPT query consumes roughly 10 times the energy of a traditional Google search, and as AI becomes embedded in search results, email composition, code suggestions, and enterprise workflows serving hundreds of millions of daily users, the aggregate impact is compounding rapidly. The hardware trajectory is exacerbating the problem: NVIDIA’s B200 GPU draws 1,000 watts per chip (43% more than the H100), while next-generation training clusters will contain hundreds of thousands of these chips operating around the clock for months. Major technology companies including Google and Microsoft have acknowledged that their net-zero climate commitments are at risk specifically because of AI scaling. Over 5 GW of new data center capacity is under construction in the US alone, representing one of the largest power infrastructure buildouts since the electrification era.
Detailed Analysis of Key Data Points
50 TWh of global AI data center power draw in 2025 represents a figure that has roughly tripled since 2022, when AI workloads consumed approximately 15-18 TWh. To put 50 TWh in perspective: it exceeds the electricity consumption of countries like Portugal, New Zealand, or Ireland. It represents roughly 2% of total US electricity generation and about 0.2% of global electricity generation. However, this share is growing far faster than overall electricity demand, meaning AI is becoming an increasingly significant factor in energy planning and grid capacity allocation. The IEA estimates that data centers as a whole (including non-AI workloads) consumed roughly 460 TWh in 2025, meaning AI-specific workloads account for approximately 11% of all data center electricity consumption, up from an estimated 4-5% in 2022.
Projected 130-150 TWh by 2028 implies a roughly 3x increase in three years, a growth rate that creates genuine infrastructure challenges. Adding 80-100 TWh of power demand in three years is equivalent to adding the entire electricity consumption of a country like the Netherlands to the global grid. The projection assumes continued scaling of both training (larger models, longer runs) and inference (more users, more queries per user, more compute per query as models become multimodal and agentic). If AI agent workloads scale as aggressively as some forecasts suggest, the actual figure could exceed 150 TWh.
A single ChatGPT query using ~10x the energy of a Google search is a widely cited comparison that captures an important reality but oversimplifies the picture. A traditional Google search consumes approximately 0.3 watt-hours of energy. A ChatGPT query using a GPT-4-class model consumes an estimated 3-10 watt-hours depending on the length and complexity of the response. The 10x figure is a reasonable midpoint, but the range is wide: a simple factual query might consume only 3-4x what a Google search uses, while a complex multi-step reasoning chain or a coding task that generates hundreds of lines of output might consume 20-30x. The concern is not any individual query but the aggregate: if AI-enhanced search replaces traditional search at scale, the global energy cost of information retrieval could increase by an order of magnitude.
NVIDIA H100 TDP of 700W represents the thermal design power of each GPU at peak load. A rack of 8 H100 GPUs draws approximately 10 kW, and a full DGX H100 system (8 GPUs plus networking, CPU, memory, and cooling) draws roughly 12-13 kW. A training cluster of 10,000 H100 GPUs consumes approximately 8-9 MW of continuous power, and when cooling overhead is included (power usage effectiveness ratios of 1.2-1.4 in modern data centers), total facility power reaches 10-12 MW. This means a single large training cluster consumes as much electricity as approximately 8,000 US homes.
NVIDIA B200 TDP of 1,000W represents a 43% increase in per-GPU power draw compared to the H100. The B200 delivers roughly 2.5x the training performance and up to 5x the inference performance of the H100, so its performance per watt is significantly better. However, organizations are deploying B200s in larger clusters rather than using fewer chips at lower total power. The result is that next-generation data centers are being designed for power densities of 50-100+ kW per rack, compared to 10-20 kW per rack for traditional data centers. Many existing facilities cannot physically accommodate this power density, requiring either expensive retrofits or entirely new construction.
Google’s 48% year-over-year increase in AI-related carbon emissions is particularly striking because Google has been one of the most aggressive technology companies in pursuing carbon neutrality. The company achieved carbon neutrality for its operations in 2007 and has committed to running on 24/7 carbon-free energy by 2030. However, the pace of AI compute scaling has outrun the pace of renewable energy procurement. Google’s total emissions increased by roughly 13% in 2024, with AI workloads accounting for the majority of the increase. This highlights a fundamental tension: AI companies are simultaneously the largest corporate buyers of renewable energy and the fastest-growing sources of electricity demand.
Microsoft’s 6.4 billion liters of water usage for data center cooling in 2024 represents a 34% increase year-over-year. Water is used in evaporative cooling systems that are the most energy-efficient way to keep data centers at operating temperature, particularly in hot climates. The water-energy tradeoff is real: air-cooled systems use more electricity but less water, while liquid-cooled systems (increasingly adopted for high-density AI racks) use water more efficiently but require specialized infrastructure. Microsoft has committed to becoming water-positive by 2030, meaning it would replenish more water than it consumes, but AI scaling is making this target harder to achieve.
5+ GW of new US data center capacity under construction for the 2025-2027 period represents one of the largest power infrastructure buildouts in recent US history. For context, 5 GW is equivalent to roughly five large nuclear power plants or 10 large natural gas power plants. The construction pipeline is concentrated in Northern Virginia (the world’s largest data center market), Texas, Ohio, and the Pacific Northwest. Utilities in these regions are scrambling to meet demand, with some reporting connection wait times of 3-5 years for new data centers. This power demand is already affecting electricity prices for residential and commercial customers in data center-heavy regions and creating political pressure around energy allocation priorities.
Historical Context and Trajectory
The energy impact of computation has been a concern since the mainframe era, but AI has accelerated the issue by orders of magnitude. Global data center electricity consumption was roughly flat from 2010-2018 (at approximately 200-250 TWh annually) as efficiency improvements from the shift to cloud computing offset growth in demand. The AI era broke this equilibrium: since 2020, data center power consumption has grown at approximately 15-20% annually, with AI workloads driving the majority of the increase.
The historical precedent most relevant to AI energy consumption is the electrification of manufacturing in the early 20th century. Between 1900 and 1930, electricity consumption in the US grew roughly 10x as factories electrified. AI is driving a comparable (though smaller in absolute terms) step-change in electricity demand, concentrated in a smaller number of very large facilities. The parallel is imperfect — AI data centers are far more energy-efficient per unit of useful output than early factories — but the pattern of transformative technology driving rapid power demand growth is consistent.
What’s Driving This
Three factors are compounding AI’s energy footprint simultaneously. First, model size and training compute continue to scale exponentially. Each generation of frontier models uses roughly 5-10x more compute than the previous generation, and while hardware efficiency improves approximately 2-3x per generation, the net effect is increasing total energy consumption per training run. Second, inference demand is scaling with the number of AI users, the number of queries per user, and the compute required per query. As AI moves from simple chatbot interactions to multimodal understanding, agent-based task completion, and continuous background processing, the compute per interaction is increasing even as the cost per FLOP decreases. Third, the geographic expansion of AI infrastructure is placing new data centers in regions where the electricity grid may rely heavily on fossil fuels, particularly in the US Southeast and parts of Asia.
Comparison to Adjacent Markets
AI’s 50 TWh annual consumption is currently comparable to the energy footprint of cryptocurrency mining at its 2022 peak (estimated at 100-150 TWh globally, now reduced to roughly 50-60 TWh after the Ethereum proof-of-stake transition). However, AI’s trajectory is steeply upward while crypto’s has declined. By 2028, AI data center power consumption is projected to exceed crypto’s peak, making AI the dominant discretionary electricity consumer globally.
Compared to other large electricity consumers, AI data centers currently consume less than residential air conditioning in the US alone (approximately 200 TWh annually) but more than the entire electric vehicle fleet’s charging demand (approximately 20-30 TWh in 2025). The comparison to EVs is particularly relevant because both AI and EVs represent growing demands on the same grid infrastructure, creating potential competition for power generation and transmission capacity.
The carbon intensity of AI varies dramatically by location. AI workloads running in Iceland (nearly 100% renewable electricity), the Pacific Northwest (heavy hydroelectric), or France (75% nuclear) produce a fraction of the carbon emissions of identical workloads running in Texas (60% natural gas), India (75% coal), or Poland (70% coal). The choice of data center location can change the carbon footprint of a training run by 10-20x, making geographic decisions as important as hardware efficiency in determining AI’s climate impact.
What to Watch
The most important development to monitor is whether nuclear energy becomes a viable power source for AI data centers. Several major technology companies (Microsoft, Google, Amazon) have signed agreements or letters of intent to develop small modular nuclear reactors specifically to power AI infrastructure. Microsoft’s agreement with Constellation Energy to restart the Three Mile Island Unit 1 reactor signals how seriously the industry takes the need for carbon-free, reliable baseload power. If nuclear (including next-generation designs like small modular reactors or fusion) can be deployed at scale for AI, it would resolve the tension between AI scaling and climate commitments. If nuclear timelines slip (as they historically have), AI’s carbon footprint will continue to grow.
The efficiency of inference is a crucial variable. Techniques like speculative decoding, model distillation, and quantization have already reduced the energy cost per query by 50-70% for certain workloads. If inference efficiency continues to improve at this rate, total energy consumption could stabilize even as query volumes grow. However, the trend toward more complex agentic interactions (where a single user request may trigger hundreds of inference calls as the agent plans, acts, and iterates) could offset these efficiency gains by dramatically increasing compute per user interaction.
Regulatory intervention is increasingly likely. The EU is considering requirements for AI energy consumption disclosure, and several US states with large data center concentrations are debating limits on data center power usage or requiring renewable energy mandates. Any significant regulation that restricts power availability or increases costs for AI data centers could slow the pace of AI deployment in affected regions and shift infrastructure investment to more permissive jurisdictions.
Frequently Asked Questions
How much energy does AI use compared to other industries? Global AI data center power consumption reached approximately 50 TWh in 2025, which is comparable to the electricity consumption of Sweden or Portugal. This represents roughly 11% of total global data center electricity usage and about 0.2% of global electricity generation. While these numbers may seem small in percentage terms, the growth rate (roughly 40% annually) far exceeds overall electricity demand growth (roughly 2-3% annually), meaning AI is becoming an increasingly significant factor in energy planning.
Does AI use more energy than Bitcoin? As of 2025, global AI data center energy consumption (approximately 50 TWh) is roughly comparable to Bitcoin mining (approximately 50-60 TWh after Ethereum’s transition to proof-of-stake reduced total crypto energy consumption). However, the trajectories are divergent: AI energy consumption is projected to nearly triple to 130-150 TWh by 2028, while Bitcoin mining energy consumption is expected to remain relatively stable. AI will likely surpass cryptocurrency as the largest discretionary electricity consumer within 1-2 years.
What are companies doing to reduce AI’s energy footprint? Technology companies are pursuing multiple strategies: procuring renewable energy (Google, Microsoft, and Meta are among the world’s largest corporate buyers of wind and solar), improving hardware efficiency (each new GPU generation delivers more computation per watt), optimizing software and algorithms (model distillation, speculative decoding, and quantization reduce inference energy), and exploring nuclear power (Microsoft has signed a deal to restart a nuclear reactor, and Google has invested in geothermal energy). Despite these efforts, total AI energy consumption continues to rise because demand growth outpaces efficiency improvements.
Will AI energy consumption keep growing? Most credible projections expect AI energy consumption to continue growing rapidly through at least 2030, reaching 200-300 TWh annually. The growth rate will depend on several factors: how quickly model efficiency improves, whether AI agent workloads scale as aggressively as projected, whether new energy-efficient chip architectures achieve commercial adoption, and whether regulations limit data center power consumption. The consensus view is that AI energy consumption will roughly triple from 2025 levels by 2028 before efficiency improvements begin to moderate the growth rate in the 2029-2030 timeframe.