Every time you ask ChatGPT a question, generate text with Gemini, or run a coding agent, servers somewhere draw electricity. The question many people now ask is simple: how much electricity does AI use for one prompt?

Early estimates varied wildly. Official disclosures and independent analyses from 2025-2026 now give a clearer picture. A typical short text prompt uses a fraction of a watt-hour — comparable to running a microwave for one second or watching television for a few seconds. More complex reasoning tasks and especially agentic workflows consume far more.

See our companion piece on the water side: How Much Water Does AI Use for One Prompt?.

Official and Recent Electricity Use Per AI Prompt

Here are the most reliable figures available in 2026:

Service / ModelReported Energy Per PromptSource / Notes
Google Gemini (median text)0.24 WhGoogle's August 2025 comprehensive measurement (May 2025 production data). Includes AI accelerators, host CPU/memory, idle backup capacity, data-center overhead. ≈ TV for <9 sec or microwave for ~1 sec. 33x reduction vs prior year.
ChatGPT / OpenAI average query~0.34 WhSam Altman's June 2025 disclosure — ≈ oven for >1 sec or LED bulb for a few minutes.
Independent benchmarks for typical chatbot queries~0.3 WhEpoch AI modeling + other benchmarks for GPT-4o-class models.
Longer / reasoning-heavy promptsSeveral to tens of Wh~1.8-5+ Wh for medium-to-long Claude / GPT-4o. ~18 Wh (up to 40) for medium GPT-5-class with extended thinking. Higher for DeepSeek-R1 on long prompts.
Agentic AI workflows (coding agents)~150 Wh per prompt avg (range ~60-290)Climate researcher Zeke Hausfather's tracking of Claude Code. ~600x a simple chat prompt because each user prompt triggers many internal model calls + heavy context re-processing.

A narrower “chip-only” view for Gemini drops to about 0.10 Wh, but Google emphasizes that full-stack accounting (0.24 Wh) better reflects real-world operation.

Advertisement

Why AI Prompts Use Electricity

Large language models run on specialized accelerators (GPUs, TPUs, or equivalent) inside data centers. Energy is consumed by:

  • The AI chips performing matrix multiplications and attention calculations
  • Supporting hardware (CPUs, memory, networking)
  • Idle capacity kept ready for reliability and load balancing
  • Data-center overhead (power conversion, cooling systems, lighting)

Google's breakdown for the median Gemini prompt shows:

  • AI accelerators ≈ 58% of the 0.24 Wh
  • Host CPU / DRAM ≈ 25%
  • Idle machines ≈ 10%
  • Data-center overhead ≈ 8%

Efficiency has improved rapidly through better model architectures, quantization, custom silicon and optimized serving software. Still, longer outputs, multi-step reasoning, image/video generation and especially autonomous agents multiply the energy cost because they generate far more tokens and trigger repeated model invocations.

Everyday Comparisons

  • 0.24-0.34 Wh ≈ 1 second of microwave use or ~9 seconds of television
  • Charging a smartphone typically uses ~10-20 Wh
  • A typical US household uses tens of thousands of Wh per day
  • One hour of laptop use is usually in the range of 20-50 Wh depending on the device

At the individual level, a few dozen ordinary text prompts barely register on a household electricity bill. At global scale the picture changes: billions of daily queries across major platforms add up to meaningful demand on power grids and drive rapid data-center expansion.

What Drives Higher Electricity Use

Energy per prompt is not fixed. It rises with:

  • Model size and capability (frontier reasoning models cost more than lightweight ones)
  • Prompt and response length (more tokens = more computation)
  • Reasoning or “thinking” modes that generate internal steps before the final answer
  • Multimodal tasks (images and especially video generation are significantly more intensive)
  • Agentic systems that loop, call tools and re-read large context windows many times
  • Hardware utilization, data-center PUE and geographic location

Training a new large model remains far more energy-intensive than any single inference, but training occurs relatively infrequently while inference happens continuously at massive scale.

Why Agentic AI Is the Big Story for 2026

The 600x delta between a simple chat prompt (~0.24 Wh) and a Claude Code agent invocation (~150 Wh) is the most important 2026 signal. As coding agents, research agents and multi-step workflows become the norm, aggregate AI electricity demand is scaling much faster than a naive “chatbot users x per-prompt cost” calculation would suggest.

For NRI tech professionals working on cloud infra + AI, this means:

  • Data-center power planning must account for agentic-workflow spikes
  • Enterprise cost models under-estimate if they only use average chat pricing
  • PUE and cooling efficiency become even more important on the margin

The Bigger Picture for 2026 and Beyond

Simple chat prompts are becoming remarkably efficient. The real growth in electricity demand is coming from heavier workloads — long-context reasoning, coding agents, research tools, and multimodal generation — as well as the sheer volume of global usage.

Hyperscalers continue investing in efficiency, custom chips, renewable energy matching and improved cooling. Transparency is improving (Google's detailed methodology is a notable step), though most providers still release limited or no per-prompt energy data. Independent researchers fill the gap with estimates based on public APIs, hardware specs and token volumes.

Advertisement

Why This Matters for the Indian Diaspora and Indian Tech

  • India's data-center boom. Mumbai, Chennai, Hyderabad, Pune, Bengaluru are among the fastest-growing global data-center hubs. Per-prompt energy translates directly into grid-planning pressure.
  • NRI energy-sector investors. Rising AI demand is one of the tailwinds for utility-scale renewables, nuclear SMRs, and grid modernization — sectors where several NRI-led funds are already active.
  • Career signal. Skills in AI efficiency (quantization, distillation, MoE serving), agentic-workflow cost engineering and data-center PUE are increasingly valuable for Indian tech professionals.

See our companion pieces: How Much Water AI Uses Per Prompt, AI Impact on IT Jobs 2026 and Best Tech Jobs for NRIs 2026.

Key Takeaways for Users

  • A typical short text prompt today uses roughly 0.24-0.34 Wh — a very small amount of electricity.
  • Complex reasoning, long outputs and especially agentic AI can use tens to hundreds of times more energy per interaction.
  • Individual impact is modest; collective impact is large and growing with AI adoption.
  • Efficiency gains on the provider side continue to lower the cost per unit of intelligence, even as overall demand rises.
  • If sustainability is a concern, prefer short, focused prompts and use agentic workflows judiciously.

Frequently Asked Questions

Does one ChatGPT prompt really use only ~0.34 Wh?

That is OpenAI CEO Sam Altman's stated figure for an average query as of June 2025. Independent estimates for typical text prompts land in a similar 0.3-0.5 Wh range. Longer or heavier prompts use more.

Which AI provider is most efficient per prompt?

Google Gemini has published the lowest verified figure (0.24 Wh median), based on Aug 2025 methodology. OpenAI's disclosed figure is ~0.34 Wh. Both are much lower than earlier viral estimates.

Why do agentic AI workflows use so much more energy?

Each user prompt triggers many internal model calls (planning, tool use, re-reading context, verification). Coding agents like Claude Code have been measured at ~150 Wh per user prompt on average — ~600x a simple chat.

How does AI energy compare to household use?

A US household uses tens of thousands of Wh per day. A hundred typical AI prompts use less than half a kilowatt-hour — roughly 1% of daily household usage. Aggregate demand at the platform level is a different story.

Is training more energy-intensive than inference?

Per event, yes. But training happens once (or periodically) while inference happens billions of times. Cumulative inference load can rival or exceed training over a model's lifetime.

How can I reduce my per-prompt energy use?

Prefer shorter, more focused prompts. Avoid heavy agentic workflows for tasks that a simple prompt can handle. Choose models sized to the task — frontier reasoning models are much more energy-intensive than lightweight chat models.

As AI tools become more powerful and more deeply integrated into work and daily life, accurate measurement of energy use will remain essential. The electricity required for one ordinary prompt is no longer a mystery — current data show it is measured in fractions of a watt-hour for simple queries, with clear upward scaling for more demanding tasks.

Primary sources for these figures include Google's 2025 Gemini environmental-impact methodology, Sam Altman's June 2025 disclosures, independent benchmarks from Epoch AI and academic teams, and Zeke Hausfather's tracking of agentic workloads. Numbers will continue to evolve as models and infrastructure improve. NRI Globe provides journalism and general information only.