Artificial intelligence can feel weightless, but every response rests on data, electricity, water, advanced chips and rapidly expanding physical infrastructure. The evidence shows those costs are growing quickly, even as efficiency improves.
THE UNIVERSAL RECORD
Sourced reporting. No opinions.
Brad Socha | August 28, 2026 | 9:41 AM EST
Featured image: Frontier supercomputer at Oak Ridge National Laboratory in Tennessee. Photo: Oak Ridge National Laboratory, U.S. Department of Energy.
Artificial intelligence begins with data and ends with electricity.
Between those points sits an increasingly enormous industrial system: semiconductor fabrication plants, specialized processors, servers, fibre networks, cooling equipment, power substations, transmission lines and data centres. Behind an AI response is also something less visible, the immense collection of human-created text, images, code, audio and other information from which modern models learn.
The scale is becoming difficult to ignore. The International Energy Agency reported in April 2026 that electricity consumption at AI-focused data centres rose about 50% in 2025, while electricity use across all data centres increased 17%. Its updated central projection puts total global data-centre electricity consumption at roughly 950 terawatt-hours by 2030, nearly twice the 485 TWh consumed in 2025.
But those numbers require an important qualification: AI is not responsible for every server or every kilowatt-hour consumed by a data centre. Cloud computing, streaming, storage, search, financial systems and countless other digital services share the same infrastructure.
The more revealing question is therefore not whether AI has a physical cost. It clearly does. The question is what that cost actually consists of, and whether growing efficiency can keep pace with extraordinary growth in demand.
AI Begins With Data, and Its Origins Remain Murky
Before a model consumes electricity answering a question, developers must teach it patterns from enormous quantities of information.
That information can include publicly accessible websites, books and articles, code, images, licensed archives, datasets obtained through partnerships, human-generated training material, user-provided information and increasingly synthetic data produced by other models.
Exactly what went into many leading systems remains difficult for outsiders to determine.
Stanford’s 2025 Foundation Model Transparency Index found continuing weakness in disclosure surrounding data acquisition, copyright, licensing and personally identifiable information. The average transparency score among the companies evaluated was 41 out of 100.
Developers have nevertheless disclosed broad categories.
OpenAI says its models are developed using publicly available internet information, information obtained through third-party partnerships, and material supplied or generated by users, trainers and researchers. A California disclosure updated in August 2026 goes further, stating that its development datasets contain trillions of tokens across text, images, audio and audiovisual material and may include copyrighted works and personal information. The company says it takes measures to reduce personal information and allows certain users to opt out of training.
Anthropic similarly says recent Claude models were trained using a proprietary mixture of publicly available internet information, non-public third-party data, contractor-generated material, opted-in user data and internally generated information.
Those descriptions reveal categories, not a complete inventory of individual works.
That gap sits at the centre of a continuing copyright dispute. Authors, publishers, artists and other rights holders have challenged whether copyrighted works may lawfully be copied for model development without permission. The U.S. Copyright Office concluded in its 2025 report on generative-AI training that copyright questions depend heavily on circumstances, including the works involved, the source of copies, the purpose of the use and what the resulting system produces.
Regulation is beginning to force greater disclosure. The European Union requires providers of general-purpose AI models to publish summaries of their training content and maintain policies for complying with EU copyright law. The European Commission’s template specifically covers scraped online sources, public and private datasets, user data and synthetic material.
It is an important shift, but a training-content summary is still not the same thing as a complete searchable catalogue of every work used.
The Physical AI Economy
Once data becomes computation, AI moves firmly into the physical world.
Training frontier models requires clusters of specialized accelerators operating in parallel. After training comes inference, the repeated computation required every time somebody asks a chatbot a question, generates an image, translates a document or sends an AI agent to perform a task.
Training can represent an enormous concentrated computational event. Inference may be individually smaller but can occur billions of times. As AI adoption expands, inference can therefore become a major continuing source of demand.
The IEA estimates that data centres consumed about 485 TWh globally in 2025. Its April 2026 analysis projects approximately 950 TWh in 2030, around 3% of global electricity demand. Electricity consumption specifically from AI-focused data centres is projected to triple over that period.
The pressure is particularly concentrated geographically. A hyperscale AI facility can demand hundreds of megawatts, while planned campuses can reach gigawatt scale. Unlike a widely distributed consumer load, that demand may arrive at a particular point on a regional electricity network faster than new generation, substations and transmission infrastructure can be constructed.
The IEA says grid-connection delays, transformer availability, chip manufacturing and other supply constraints are already slowing some development.
In the United States, Lawrence Berkeley National Laboratory’s June 2026 update estimates that data centres could consume 11.8% of national electricity in 2030 under its central estimate, with scenarios ranging from 9.5% to 15.3%.
Water adds another layer.
Some data centres consume water directly through cooling systems, while additional water can be associated with electricity generation. Berkeley Lab estimated U.S. data centres directly consumed about 66 billion litres in 2023. Future consumption depends heavily on cooling technology, geography and operating choices, making a universal “water per AI question” figure unreliable.
That problem is visible in widely circulated claims about individual prompts.
A 2023 academic analysis estimated substantially higher water requirements for AI under particular assumptions about models, locations and cooling conditions. In 2025, Google published operational measurements estimating that its median Gemini Apps text prompt consumed 0.24 watt-hours of electricity and 0.26 millilitres of water.
Those figures are not necessarily contradictory. They measure different systems, periods and assumptions.
Google’s number also cannot automatically be applied to ChatGPT, Claude or other models. Prompt length, model size, hardware utilization, reasoning effort, output length, location, electricity source and cooling method can all change the result dramatically.
The IEA reported in 2026 that electricity consumption per AI task has been falling by at least an order of magnitude annually in recent years. But it simultaneously warned that video generation, reasoning systems and autonomous agents can consume hundreds or thousands of times more energy per task than simple text generation.
There is consequently no scientifically defensible single answer to “How much electricity does an AI question use?”
Who Pays, and Who Benefits?
The financial trail demonstrates the physical scale more clearly than almost anything else.
The IEA reported that capital spending by five major technology companies exceeded $400 billion in 2025 and was expected to rise another 75% in 2026.
Individual corporate projections are extraordinary.
Alphabet told investors it expected 2026 capital expenditures of $175 billion to $185 billion, largely reflecting investment in AI computing capacity and related infrastructure.
Amazon said it expected approximately $200 billion in capital expenditures during 2026 across opportunities including AI, chips, robotics and its satellite network.
Meta increased its expected 2026 capital expenditures to $125 billion to $145 billion, citing higher component prices and additional data-centre costs.
Microsoft told investors in April that it expected roughly $190 billion in calendar-year 2026 capital expenditures. In one quarter alone, Microsoft reported $31.9 billion in capital spending, with roughly two-thirds directed to shorter-lived assets, primarily GPUs and CPUs.
That money flows through an extensive industrial chain: chip designers such as Nvidia and AMD; semiconductor foundries; memory manufacturers; networking companies; construction contractors; utilities; power-equipment manufacturers; cooling suppliers and landowners.
Governments can also become participants through tax incentives, infrastructure investment and economic-development programs. The precise allocation of grid-upgrade costs varies by jurisdiction and utility regulation, making broad claims that either technology companies or ordinary ratepayers universally “pay for” AI infrastructure misleading. The answer depends on the project and local regulatory structure.
AI chips carry their own upstream footprint. Semiconductor fabrication requires electricity, chemicals and exceptionally pure water. A U.S. National Institute of Standards and Technology environmental assessment reported median ultrapure-water use across 29 semiconductor facilities at approximately 1.16 million gallons per day in 2021.
The supply chain is geographically concentrated as well. Advanced chip fabrication depends heavily on East Asia, while specialized materials introduce additional vulnerabilities. The IEA estimates data-centre demand for gallium could equal as much as 10% of today’s global supply by 2030; China currently accounts for approximately 95% of gallium refining.
Then comes replacement.
GPUs, CPUs, memory and networking equipment do not last indefinitely, and the economics of AI can encourage replacement before hardware physically fails. Microsoft disclosed that a large share of current capital spending is directed toward relatively short-lived processors.
A 2024 peer-reviewed study in Nature Computational Science estimated that generative AI could cumulatively produce approximately 1.2 million to 5 million tonnes of electronic waste between 2020 and 2030, depending on how the industry develops. That is a scenario-based projection, not a measured forecast. The researchers also found that circular-economy measures could substantially reduce the amount.
Efficiency could change the trajectory further.
Google says its latest TPU generations deliver dramatically greater performance per unit of electricity than earlier hardware. Microsoft is deploying direct-to-chip liquid cooling, which it says can reduce water requirements compared with traditional evaporative approaches. Models themselves are becoming more efficient, while smaller specialized systems can sometimes perform tasks without invoking the largest available model.
Cleaner electricity matters too. Technology companies are signing large renewable-energy agreements and pursuing nuclear, geothermal and other generation. But contractual clean-energy purchases should not automatically be confused with the electricity physically flowing to a particular data centre at every hour. The local grid mix still matters.
There is another side of the ledger.
The same AI systems consuming energy could potentially reduce energy use elsewhere. The IEA identifies applications in grid management, renewable-energy forecasting, industrial optimization, transportation, building efficiency, methane detection and materials discovery. Its analysis concludes that widespread AI-led optimization in the energy sector could potentially reduce emissions by more than data-centre operations create, but only if those applications are actually deployed at scale. Rebound effects could erase part of those gains.
That leaves the central finding less dramatic, but more consequential.
AI is neither immaterial nor inherently environmentally destructive. Its cost cannot be reduced to the electricity consumed by a chatbot prompt.
The real system begins with human-created information and extends through semiconductor factories, mines, electrical grids, water systems, data centres and eventually discarded hardware. Its footprint is growing rapidly, while the efficiency of the technology powering it is improving just as rapidly.
What remains missing is comprehensive, standardized disclosure separating AI from other data-centre workloads and allowing independent comparison of models, training runs and inference systems.
Until that exists, the most precise answer to what artificial intelligence really costs is also the most important: far more than the price shown on a subscription, and not yet enough is publicly measured to calculate the complete bill.
Sources:
International Energy Agency — Key Questions on Energy and AI — April 16, 2026
https://www.iea.org/reports/key-questions-on-energy-and-ai
International Energy Agency — Energy and AI
https://www.iea.org/reports/energy-and-ai/
International Energy Agency — Energy Demand from AI
https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
International Energy Agency — AI and Energy Security
https://www.iea.org/reports/energy-and-ai/ai-and-energy-security
International Energy Agency — AI and Climate Change
https://www.iea.org/reports/energy-and-ai/ai-and-climate-change
Lawrence Berkeley National Laboratory — United States Data Center Energy Usage Report: 2025 Update
https://seta.lbl.gov/publications/united-states-data-center-energy-2025
Lawrence Berkeley National Laboratory — 2024 United States Data Center Energy Usage Report
https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf
National Institute of Standards and Technology — Final Programmatic Environmental Assessment for Modernization and Expansion of Existing Semiconductor Fabrication Facilities
https://www.nist.gov/system/files/documents/2024/06/28/Final%20PEA%20for%20Modernization%20and%20Expansion%20of%20Semiconductor%20Fabs%206-28-2024%20-%20OGC-508C.pdf
U.S. Copyright Office — Copyright and Artificial Intelligence, Part 3: Generative AI Training
https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
European Commission — Template for General-Purpose AI Model Providers to Summarise Their Training Content
https://digital-strategy.ec.europa.eu/en/faqs/template-general-purpose-ai-model-providers-summarise-their-training-content
Stanford Center for Research on Foundation Models — Foundation Model Transparency Index
https://crfm.stanford.edu/fmti/
OpenAI — Our Approach to Data and AI
https://openai.com/index/approach-to-data-and-ai/
OpenAI — Training Data Summary Pursuant to California Civil Code Section 3111
https://help.openai.com/en/articles/20001044-training-data-summary-pursuant-to-california-civil-code-section-3111
Anthropic — Transparency Hub
https://www.anthropic.com/transparency
Google — Measuring the Environmental Impact of AI Inference
https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference/
Google — 2025 Environmental Report
https://sustainability.google/google-2025-environmental-report/
Microsoft — 2025 Environmental Sustainability Report
https://www.microsoft.com/en-us/corporate-responsibility/sustainability/report
Microsoft — Fiscal Year 2026 Third Quarter Earnings Conference Call
https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q3
Alphabet — 2025 Fourth Quarter Earnings Call
https://abc.xyz/investor/events/event-details/2026/2025-Q4-Earnings-Call-2026-Dr_C033hS6/default.aspx
Amazon — Fourth Quarter 2025 Results
https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Fourth-Quarter-Results/default.aspx
Meta — First Quarter 2026 Results
https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-First-Quarter-2026-Results/
Wang et al., Nature Computational Science — E-waste Challenges of Generative Artificial Intelligence
https://www.nature.com/articles/s43588-024-00712-6
About the Author
Brad Socha is the founder of The Universal Record, focused on sourced, factual global reporting. Coverage includes international news, geopolitics, technology, and major developments.
