Google AI Queries Now 33x More Energy Efficient

Google announced a significant breakthrough in reducing the effects of AI on the environment, reporting a 33-fold reduction in the energy consumption of AI's requests over the past year.

According to him Last research workOne text request now consumes only 0.24 watt hours of electricity. This is equivalent to watching nine seconds of TV.

The company also claims to reduce general emissions by 44 times related to the tips for the text of artificial intelligence in its Gemini Apps package, mainly because of the efficiency of both software and equipment.

Google shares a technical article on a comprehensive methodology for measuring energy, emissions and exposure to water tips on Gemini.
During the 12-month period, energy consumption on the median application of the twins of the text application decreased by 33 times, and a carbon trace of 44x. This… pic.twitter.com/g0yoyzywt

– Yossi Matias (@ymatias) August 21, 2025

Growing problems with energy consumption of AI

AI consumes a lot of energy. The launch of large language models needs a huge number of computing power and infrastructure that use electricity.

It is not surprising that forecasts indicate that Electricity consumption of AI can increase by 50 percent Every year from 2023 to 2030.

Bloomberg also emphasized the growing demand for power in data processing centers. The company calculated that the demand for electricity to the United States from data processing centers may increase 20-40% in 2025 And probably see two-digit growth to 2026-2030.

An increase in energy consumption leads to higher emissions of harmful gases.

And we have estimates to prove it. Demand for electricity controlled by AI may increase 1.7 gigatons of greenhouse gas emissions All over the world from 2025 to 2030. This is the same amount that Italy will produce from the use of energy in five years.

According to the Gapta et al., It is expected that technological companies will require cancellation of water 4.2 to 6.6 billion cubic meters by 2027 Potentially launch and cool their data processing centers. This will be from 1.7 to 2.6 million Olympic pools.

These figures emphasize the urgent need for effectiveness and transparency, since the rapid growth of AI becomes a significant load on global energy and water resources.

Ecological mathematics Google

Google reports that the average text request for Gemini applications now gives:

• 0.24 WH of the use of electricity, which is equivalent to watching the TV for 9 seconds.
• 0.26 ml of water consumption, which is approximately 5 drops.
• 0.03 grams of equivalent emissions CO₂.

Compared to 2024, Google claims to have Reducing energy consumption to an invitation by 33 times And a decrease in general emissions by 44 times.

The research article does not include data 2024 for power consumption, use of water or equivalent emissions.

To achieve their conclusions, the Google research group measured consumption Active Accelerator AI energy, active processor and Drama Energy, idle energy and overhead energy.

The energy on Idle Machine refers to the power consumed by computers of artificial intelligence, which are stored in standby mode.

They may not be active, but remain ready to process traffic spikes or refusal. The energy that these systems use in reserve art also increases the overall use of AI energy.

In the category of energy costs, Google considered energy consumption using infrastructure supporting data centers such as cooling systems or energy transformation.

How Google has reached this feat

Google explained this increase in the effectiveness of the next combination of strategies:

1. Smart architecture model

Models of twins are based on Transformer DesignWhich is already 10-100 times more effective than other models.

BesidesGoogle uses effective methods, including A mixture of experts (MO)reasonable Calculation of attentionand hybrid reasoning to increase efficiencyField

2. Effective algorithms and quantification

Google increases the effectiveness of its AI due to the constant improvement of algorithms that feed AI models. He uses methods such as Accurate quantum training To reduce energy consumption without prejudice to the quality of the answer.

3. Optimized conclusion and portion

Google is constantly working to make the delivery of artificial intelligence smarter, faster and more efficiently. He uses technologies as Specify decoding, Which offers a faster and more profitable conclusion from LLMS.

He also uses distillation Create more effective artificial art models optimized by the service.

4. User equipment

Google Ten. Known for its power efficiency.

The company has also developed TPU and AI models. This allows the software and hardware to work freely, increasing energy efficiency.

In fact, the latest TPU (Ironwood) is 30 times more energy efficient than the first public TPU model.

5. Optimized idle

Instead of constantly maintaining its stack for AI in standby mode, Google uses an optimized idle time to minimize the downtime of the accelerator. The models are shifted near in real time, depending on demand, and not without necessity.

6. Software stack ml

Google XLA compilerPallas nuclei, and Ways Systemic optimization of artificial intelligence models written in such frameworks as JaxAllows them to work effectively on Google user accelerators.

7. Ultra -efficient data processing centers

Google data processing centers are highly effective, working on the efficiency of electricity (Pue) 1.09.

This means that for every 1 watt used by servers, only 0.09 watts are used for overhead costs (cooling, fans, etc.).

Given the average value of the global data processing center Pue at 1.56Google data processing centers are highly effective.

8. Purchase of pure energy

Pursuing his 24/7 without carbon ambitions, Google emphasizes the purchases of pure energy.

As a result, Google data processing centers managed to reduce its emissions from 2023 to 2024, despite an increase in electricity consumption during this period.

Google vs. Openai

AI enthusiasts are often discussed, which is more energy -efficient: Google models or models of its main rival Openai.

In a personal message on the Openai blog Sam Altman It is discussed CHATGPT energy consumptionIN

“The average request uses about 0.34 watt hours, that the oven will use a little more than one second, or a highly effective light bulb will use in a couple of minutes. He also uses about 0.000085 gallons of water; About one fifteenth part of a teaspoon. ”

Since his message on the blog does not explain how the numbers were measured or what was calculated, we cannot reasonably compare it with other estimates, such as Google's disclosure explained in this post.

It is clear that without transparent, standardized measurements it will be difficult to compare the requirements of the effectiveness of the AI ​​models.

Thus, detailed research and measurements of Google can serve as a standard for assessing the energy use of AI.

Google stands for a single approach to measuring AI efficiency

The Google study emphasizes that the exposure of AI on the environment requires an integrated approach instead of a limited standard.

Although the impact on Per-Prompt is insignificant, the widespread use of AI around the world makes continuous effectiveness necessary.

Without standardized, transparent framework, the organization cannot guarantee that rapid growth of AI corresponds to the goals of stability and accountability.

Nevertheless, this encourages that Google publishes its methodology in a research article for everyone, and not use it exclusively as a PR.

Research work has completed;

“We advocate the widespread introduction of these or similar complex measurements to guarantee that, as AI progresses, their environmental efficiency has also coped.”