What Comes Next: Efficiency, New Power and Better Data
AI's energy story is a race. On one side, each answer is getting more efficient. On the other, people are using AI far more, for heavier tasks. Where the race ends up depends on choices made by companies, governments and users.
This final lesson looks at the main forces that will shape AI's energy and water use, and how to keep judging new claims for yourself.
What You'll Learn
- How fast the energy per AI answer can fall
- Why efficiency does not always mean less total energy use
- Where the extra electricity for data centers is expected to come from
- What companies are doing about water
- Why better reporting matters, and what to look for
Efficiency is improving fast
Google reported that over one year, the energy used by its median Gemini text prompt fell by about 33 times, and the carbon footprint by about 44 times. It credited better software, more efficient chips and cleaner electricity.
You saw several of the tools behind gains like this earlier in the course and in AI Hardware Explained:
- More efficient chips that do more calculations per unit of power.
- Smaller or compressed models that give similar answers with less computing.
- Smarter serving, such as handling many requests together and sending simple questions to smaller models.
- Better cooling that wastes less energy and water.
The IEA's scenarios show how much this matters. In its high-efficiency case, data center electricity use in 2035 is about 20% lower than in its base case.
Why efficiency is not the whole answer
There is a well-known pattern in economics: when something becomes cheaper and easier to use, people often use much more of it. This is sometimes called the rebound effect or Jevons paradox.
AI shows the same pattern. As answers get cheaper:
- More people use AI, more often.
- Apps add AI to more features.
- Heavier tasks, like reasoning, agents, images and video, become normal.
So the energy per answer can fall while the total energy still rises. Both facts can be true at once. That is why the IEA still expects data center demand to grow strongly even with efficiency gains.
Where the new power is expected to come from
The IEA expects a mix of sources to meet growing data center demand:
- Renewables such as wind and solar are expected to meet about half of the growth.
- Natural gas is expected to provide a large share of the rest.
- Nuclear power is expected to add about as much new supply as gas.
AI companies are also signing long-term deals for new power. For example:
- Microsoft signed a 20-year agreement to buy power from a nuclear plant that had been shut down, so it could be restarted.
- Google signed an agreement with Kairos Power to buy electricity from several small modular reactors, with the first planned around 2030.
Small modular reactors are smaller nuclear plants designed to be built in factories and installed more quickly. The IEA expects the first ones to come online around 2030. They are promising, but new nuclear projects often take longer than planned, so it is wise to treat early dates as targets, not promises.
What is happening with water
Water is getting more attention as well. Common responses include:
- Low-water cooling, such as closed loops that reuse the same water, or using outside air when it is cool enough.
- Choosing sites where water is less scarce.
- Using recycled or non-drinking water instead of fresh water.
- Replenishment goals, where companies fund projects that return water to local rivers and aquifers. Google, for example, has a goal to replenish more water than its data centers use by 2030.
Replenishment helps, but it is not the same as using less water. A company can fund a water project in one place while a data center consumes water in another. When you read water claims, check both: how much is used, and where.
Better data is the next big step
One big lesson from this course is that numbers vary because reporting varies. Only a few companies publish per-prompt figures, and each counts different things.
Better reporting would include:
- Clear scope. Chips only, or the whole data center? Cooling water only, or power plant water too?
- Per-task figures for text, reasoning, images and video, not just one average.
- Training figures for new models.
- Location and time detail, since carbon and water depend on both.
Researchers are working on shared benchmarks. For example, Hugging Face and others run an AI Energy Score project that rates models on the energy they use for the same tasks. Mistral's 2025 life cycle study is an early example of a company measuring a model's full footprint, including hardware.
How to judge new claims
You will keep seeing new numbers about AI and the environment. Use this short checklist:
- What task? A short text reply, a reasoning answer, an image or a video?
- What scope? Chips only, the whole data center, or the full life cycle?
- Which model and when? Figures change quickly as models and hardware change.
- Measured or estimated? And by whom?
- Per answer or total? Small per-answer figures and large totals can both be true.
Where to go next
- For how the chips, memory and data centers actually work, take AI Hardware Explained.
- For the wider ethical questions around AI, including fairness, privacy and responsibility, take AI Ethics & Responsible AI.
Key Takeaways
- Efficiency is improving fast. Google reported a 33-times drop in energy per median text prompt over one year.
- Because of the rebound effect, total energy use can still rise as AI gets cheaper and more popular.
- The IEA expects renewables to meet about half of new data center demand, with gas and nuclear making up much of the rest.
- Water replenishment is not the same as using less water. Check both.
- Better, clearer reporting is key. Always ask about the task, scope, model, date and method.
Sources
- Google, Measuring the environmental impact of delivering AI at Google scale (2025)
- International Energy Agency, Energy and AI: Executive summary (2025)
- Google, New nuclear clean energy agreement with Kairos Power (2024)
- Constellation, Constellation to launch Crane Clean Energy Center, restoring jobs and carbon-free power to the grid (2024)
- Google, 2026 Environmental Report (2026)
- Hugging Face, AI Energy Score
- Mistral AI, Our contribution to a global environmental standard for AI (2025)