Electricity matters enormously to an AI company. It determines how much capacity can be built, where data centres can operate and whether local grids can connect new projects. It does not follow that a cheaper kilowatt-hour would make a £450 professional subscription cost £35. The retail price pays for an entire intelligence service, not the electricity consumed by one median prompt.
Google’s study, published in August 2025, estimates that the median Gemini text prompt used 0.24 watt-hours, emitted 0.03 grams of carbon-dioxide equivalent and consumed 0.26 millilitres of water. The company says energy per median prompt fell 33-fold and carbon fell 44-fold over a recent twelve-month period while answer quality improved. These are provider measurements of a median text prompt, not a frontier agent run or an industry average.[1]
Why prompt electricity and subscription price are different quantities
| Measure | Value | Context |
|---|---|---|
| One median Gemini text prompt | 0.24 Wh | Google’s measured median across its own applications and methodology.Source 1 |
| 1,000 such prompts | 0.24 kWh | Our multiplication of the published median; actual prompts vary widely.Source 1 |
| Illustrative electricity at £0.25/kWh | £0.06 for 1,000 prompts | Illustrative arithmetic at an assumed electricity price. Google’s 0.24 Wh already includes host systems, idle capacity and data-centre energy overhead; non-electric service costs are separate. |
| 10,000 such prompts | 2.4 kWh / £0.60 at the illustration | Still not representative of maximum-reasoning, multimodal or agent workloads.Source 1 |
| Global data-centre electricity in 2024 | About 415 TWh | Around 1.5% of global electricity use, according to the IEA.Source 2 |
| IEA base-case data-centre demand in 2030 | About 945 TWh | More than double 2024, with AI the largest growth driver.Source 2Source 3 |
The cost stack begins before electricity reaches the rack
A frontier service must finance training and post-training, salaries, data and evaluation, specialised accelerators, servers, networking, buildings, grid connections, cooling, storage, model serving, safety systems, search and code tools, product engineering, customer support, fraud prevention and compliance. It also needs spare capacity for peaks and failures. Electricity is one recurring input inside that stack.
The IEA estimates that servers account for around 60% of modern data-centre electricity, while cooling ranges from roughly 7% in efficient hyperscale sites to more than 30% in less-efficient enterprise facilities. Even the energy bill is therefore broader than the accelerator drawing power during one answer.[3]
The main routes from expensive inference to cheaper service
| Layer | Efficiency mechanism | Why the retail saving can differ |
|---|---|---|
| Model | Fewer active parameters, shorter reasoning and better stopping | Providers may spend the saving on harder tasks or higher quality |
| Software | Batching, quantisation, speculative decoding and caching | Gains depend on workload, latency target and utilisation |
| Hardware | More tokens per accelerator and per watt | New systems have high capital and deployment costsSource 4Source 5 |
| Data centre | Better power, cooling, networking and utilisation | Grid scarcity and local construction can dominate timingSource 3 |
| Product | Route tasks to the least costly adequate model | The interface, tools, safety and support still need funding |
| Energy supply | Lower electricity price or cleaner generation | Energy is only one line in total cost; lower cost may attract more use |
| Market | Competition and lower margins | A provider can retain savings or use them to expand allowances |
Efficiency can lower the unit cost while total electricity still rises
The rebound is visible in the IEA scenarios. Its High Efficiency case meets the same level of digital and AI service demand with data-centre electricity about 20% below the 2035 base case. Yet demand can still grow because cheaper and better AI is used more often, on more difficult jobs and by more people. Efficiency reduces energy per unit; adoption determines the number of units.[2][3]
This is already happening inside hardware generations. NVIDIA says its GB300 system produces far more tokens per megawatt than Hopper on its cited low-latency agentic benchmark. Software updates also raised output on the same Blackwell platform within months. Better engineering can therefore cut cost without waiting for a cheaper power plant, although the provider’s published comparison needs to be tested on the actual model and latency target.[4]
Intelligence per pound is the better consumer metric
A consumer cares whether the service solves a problem accurately, quickly and with tolerable review—not how many watt-hours appeared on the supplier’s internal meter. The useful denominator is cost per accepted result. If a more efficient model needs fewer retries or less human correction, it can be cheaper even when its raw token rate is higher. If it produces ten times more reasoning tokens without improving the outcome, a lower energy cost per token can still leave the task expensive.
AI can absolutely become much more energy efficient. Google’s prompt measurement, Stanford’s hardware trends and NVIDIA’s serving claims all point in that direction. The realistic conclusion is not that energy is irrelevant. It is that the journey from £450 to £35 is an engineering and market problem across the whole stack, with electricity as one important input rather than the single switch.
Sources & further reading
- Our approach to energy innovation and AI’s environmental footprint Google · 21 Aug 2025 · Accessed 10 Oct 2026
- Energy and AI: executive summary International Energy Agency · Accessed 10 Oct 2026
- Energy demand from AI International Energy Agency · Accessed 10 Oct 2026
- New InferenceX data on Blackwell Ultra performance and cost NVIDIA · 5 Mar 2026 · Accessed 10 Oct 2026
- Research and Development: 2025 AI Index Report Stanford HAI · Accessed 10 Oct 2026
- Measuring the environmental impact of delivering AI at Google Scale Google researchers / arXiv · 21 Aug 2025 · Accessed 14 Oct 2026