Nandini Roy Choudhury, writer
By TechSun News Desk | techsunnews.com | July 8, 2026 | Tech / AI / Explainers | 6 min read ⚡
Every time you type a question into ChatGPT, something happens that you can’t see. Somewhere in a data center — probably in Virginia, Texas, or Iowa — a rack of Nvidia GPUs lights up, draws power, generates heat, and consumes water to cool down. A response arrives on your screen in seconds. The electricity bill arrives at the end of the month.
A single ChatGPT query uses approximately 2.9 watt-hours of electricity — nearly ten times the 0.3 watt-hours a standard Google search requires, according to the International Energy Agency. That gap seems small in isolation. Multiply it by hundreds of millions of queries every day and it becomes one of the fastest-growing sources of electricity demand on the planet. (Source: IEA / Brookings Institution)
So where does it all go? And should you actually be worried about it?
Why AI Needs So Much Electricity
Standard software — the kind running your email or a spreadsheet — performs relatively simple, sequential operations. AI does something fundamentally different. Every response a language model generates involves billions of mathematical calculations, running simultaneously across thousands of processor cores.
Those processors are GPUs and specialised AI chips — and they are power-hungry by design. A single Nvidia H100 GPU, the chip that powers most of today’s frontier AI models, draws up to 700 watts of power. A standard training cluster for a frontier model uses tens of thousands of them running continuously, sometimes for weeks.
The electricity demand doesn’t stop when training ends. Every query you send to ChatGPT, every image you generate, every AI-powered search — all of it runs on inference hardware drawing power around the clock, 365 days a year.
What Happens When You Ask ChatGPT a Question
The journey from your keyboard to the response takes under three seconds. The energy consumed along the way is spread across several layers:
Your device: Minimal — your phone or laptop uses a fraction of a watt to send and receive the text.
The network: Routers, cables, and switches carry your query to the data center. Small but non-zero energy cost.
The data center: This is where most of the energy goes. Your query hits a load balancer, gets routed to a GPU server, and the model runs inference — generating your response token by token. The GPU cluster handling your request may draw several kilowatts during those seconds.
Cooling systems: GPUs generate enormous heat. Data centers use cooling infrastructure — air conditioning, liquid cooling, or outdoor cooling towers — that can consume as much energy as the servers themselves. The industry standard metric, called PUE (Power Usage Effectiveness), means a data center using 1 watt of IT load typically uses another 0.3–0.5 watts just on cooling.
According to Brookings Institution research, newer measurements suggest the median energy per text query has fallen to 0.24–0.3 watt-hours for standard prompts as efficiency improves — but longer reasoning tasks and multimodal requests (images, audio) can still push closer to the original 2.9 Wh figure.
AI Training vs AI Inference — Two Very Different Energy Problems
When people talk about AI’s energy cost, they’re usually conflating two separate things that work very differently.
Training is the one-time process of building an AI model. It requires thousands of GPUs running continuously for weeks or months, consuming gigawatt-hours of electricity. Training GPT-3 was estimated to have used 1,287 MWh — roughly equivalent to the annual electricity use of 120 US homes. Frontier models trained today are significantly larger and more expensive. (Source: AI Multiple)
Inference is what happens every time anyone uses an AI model. Each individual query is cheap compared to training — but inference runs at enormous scale, continuously. Some estimates suggest 80–90% of AI computing power is now consumed by inference rather than training, simply because of the volume of daily queries. For a model used by hundreds of millions of people, inference emissions can exceed training emissions within weeks of launch. (Source: Brookings Institution)
This distinction matters because solving AI’s energy problem requires different approaches for each: more efficient training methods for the one-time cost, and more efficient chips and model architectures for the ongoing inference load.
The Numbers Behind the Scale
| ⚡ KEY ENERGY FIGURES — 2026
2.9 Wh — energy per ChatGPT query (IEA estimate, peak figure) 0.3 Wh — energy per Google search (IEA) 460–490 TWh — global data center electricity consumption in 2025 (IEA) ~1,000 TWh — projected data center consumption by 2026 (IEA high-growth scenario) 1,287 MWh — estimated energy to train GPT-3 $355 billion — combined 2025 AI infrastructure capex from Microsoft, Google, Meta, Amazon, Apple 160% — Goldman Sachs projected increase in data center power demand by 2030 Sources: IEA, Brookings Institution, Goldman Sachs, WorldMetrics |
Water Usage — The Part Nobody Talks About
Electricity gets most of the attention. Water barely gets mentioned — which is why the figures tend to surprise people.
Cooling a data center full of GPUs requires enormous amounts of water, either directly (evaporative cooling towers that use water to dissipate heat) or indirectly (power plants that generate the electricity also use water). According to research cited by Presenc.ai, a 100-query ChatGPT session translates to roughly 0.5 litres of water in worst-case high-evaporation regions. (Source: Presenc.ai Research)
Microsoft reported consuming 6.4 million cubic metres of water in fiscal year 2022 — and that figure has risen since as AI workloads have scaled. Hyperscaler water consumption rose 25–40% year-over-year in 2024–2025 disclosures across the major cloud providers.
Liquid-cooled data centers — where coolant runs directly through server racks rather than cooling the air — reduce direct water use by 70–90%. Most new AI data centers being built in 2025–2026 use some form of liquid cooling for exactly this reason. But liquid cooling raises capital costs significantly, and the indirect water footprint from electricity generation remains.
In data-scarce regions — parts of the US Southwest, Northern Africa, the Middle East — this water demand is already creating tension between AI expansion and local water availability. Several proposed data center projects in Arizona and Nevada have faced local opposition specifically over water use.
How Companies Are Responding
The energy problem is large enough that the biggest AI companies are now treating it as a core business challenge, not just a PR issue.
Microsoft has committed to being carbon negative by 2030 and water positive by 2030, and has signed several large renewable energy deals to power its AI data centers. The company is also investing in nuclear power — signing an agreement to restart Three Mile Island as a dedicated power source for AI data center demand.
Google recently invested in a nuclear fusion startup as part of its effort to secure carbon-free power for AI workloads. The company already claims to match 100% of its electricity with renewable energy purchases on an annual basis — though critics note that matching on an annual basis is different from running on renewable power hour by hour.
Chip efficiency is improving significantly. New inference chips — including FuriosaAI’s RNGD chip that began shipping in early 2026 at 180W versus 600W+ for typical high-end GPUs — are delivering far more AI performance per watt. This is the same chip arms race we covered in our GPU, NPU, and TPU explainer — efficiency is now a competitive differentiator, not just an environmental consideration.
Model efficiency is also improving. Newer models are producing better outputs with less compute than their predecessors — the kind of efficiency gain that, if it continues, could partially offset the growth in raw AI usage.
Should You Be Concerned as a User?
Your individual AI usage — even heavy daily use — has a relatively small direct footprint. A hundred ChatGPT queries a day works out to less electricity than leaving a single LED bulb on for an hour.
The concern is structural, not individual. The issue isn’t that you’re using ChatGPT. It’s that hundreds of millions of people are, simultaneously, plus AI is being embedded into search engines, productivity software, customer service systems, and enterprise tools at a pace that means the aggregate energy demand is growing faster than renewable energy supply can replace fossil fuels in the grid.
The analogy that keeps coming up in policy discussions: the internet also seemed energy-intensive in 2000, and nobody stopped using it. Instead, data center efficiency improved, renewable energy expanded, and the per-query energy cost fell dramatically over two decades. AI researchers and energy analysts broadly expect a similar trajectory — but the speed of AI adoption is faster than anything the internet era produced, which is why the short-term pressure on grids, water systems, and energy infrastructure is more acute.
This energy demand is also directly connected to the global memory chip shortage affecting Apple and Microsoft — the same AI infrastructure buildout that’s consuming electricity is also consuming the world’s supply of high-bandwidth memory chips, driving up prices on consumer devices.
| 🟡 A NOTE FROM THE EDITOR
What stands out isn’t today’s electricity use — it’s how quickly demand is increasing. Data centers went from consuming 415 TWh in 2024 to a projected 1,000 TWh by 2026. That’s a doubling in roughly two years. If AI adoption continues at its current pace, the energy and water infrastructure the world currently has will not be sufficient. The companies that solve the efficiency problem — at the chip level, the model level, and the cooling level — won’t just be doing the right thing. They’ll have a structural cost advantage over everyone who doesn’t. |
💬 WE WANT TO HEAR FROM YOU
| Did you know AI uses this much electricity before reading this article?
A) No — this was genuinely surprising to me B) Yes — I’d read about it but didn’t know the scale C) Yes — and it makes me think twice about how often I use AI tools Tell us in the comments — and share this with someone who uses AI every day without knowing what powers it. |
❓ FREQUENTLY ASKED QUESTIONS
| Q: How does ChatGPT’s energy use compare to Google Search?
According to the International Energy Agency, a single ChatGPT query uses approximately 2.9 watt-hours of electricity at peak — roughly ten times the 0.3 watt-hours a standard Google search requires. More recent measurements suggest the median has dropped to 0.24–0.3 Wh for standard text queries as AI inference hardware has become more efficient. However, longer reasoning prompts, image generation, and multimodal requests (combining text and images) can still hit the higher figures. The key difference is that AI models generate responses token by token using GPU clusters, while Google’s search index lookup is a much simpler, faster operation. |
| Q: Is AI bad for the environment?
The answer depends on several factors, starting with where the energy comes from. AI data centers in regions powered by renewable energy have a much lower carbon footprint than those running on coal or gas. The major cloud providers — Microsoft, Google, Amazon — have made commitments to match their consumption with renewable energy, though the timelines and accounting methods vary. The broader concern is speed: AI adoption is growing faster than renewable energy can be built, meaning the short-term marginal power demand is often met with fossil fuels. Whether AI ends up being net positive or negative for the environment over the long run depends largely on how much efficiency improves and how quickly clean energy scales. |
| Q: Which AI companies use the most energy?
The five largest consumers of AI-related energy are Microsoft (Azure/OpenAI), Google (Cloud/Gemini), Amazon (AWS), Meta (AI infrastructure), and Apple (on-device AI plus cloud). Together, these five companies disclosed over $355 billion in AI-relevant infrastructure capital expenditure in 2025 alone — the largest single-cycle infrastructure investment outside of government in modern history, according to Presenc.ai. Microsoft and Google are the most exposed because they both run large public AI products (ChatGPT via Azure, Gemini) that generate enormous inference loads 24 hours a day. |
Sources: International Energy Agency (IEA) Electricity 2024 Report, Brookings Institution, Goldman Sachs AI infrastructure analysis, Presenc.ai 2026 Data Center Energy Research, WorldMetrics AI Energy Statistics, AI Multiple, Data Center Frontier, Global Electricity Review. All figures reflect information available as of July 8, 2026.
