Google has announced a significant breakthrough in reducing the environmental impact of AI, reporting a 33-fold decrease in the energy consumption of AI queries over the past year. 

According to its latest research paper, a single text query now consumes just 0.24 watt-hours of electricity. That is equivalent to watching nine seconds of television.

The company also claims a 44x reduction in total emissions associated with AI text prompts in its Gemini Apps suite, mainly due to efficiency gains in both software and hardware.

Google is sharing a technical paper on a comprehensive methodology for measuring the energy, emissions, and water impact of Gemini prompts.
Over a 12-month period, the energy consumption per median Gemini Apps text prompt decreased by 33x and the carbon footprint by 44x. This is… pic.twitter.com/G0yoyzyEWT

— Yossi Matias (@ymatias) August 21, 2025

Rising AI Energy Consumption Concerns

AI consumes a lot of power. Running large language models needs a huge amount of computing power and infrastructure, all of which use electricity.  

Unsurprisingly, projections indicate that AI electricity consumption could increase by as much as 50 percent annually between 2023 and 2030. 

Bloomberg also highlighted the growing demand for power from data centers. The company estimated that US power demand from data centers could increase by 20-40% in 2025 and likely see double-digit growth through 2026-2030. 

Increased power consumption results in higher emissions of harmful gases. 

And we have estimates to prove that. AI-driven electricity demand could increase 1.7 gigatons of greenhouse gas emissions worldwide between 2025 and 2030. This is the same amount that Italy would produce from energy use in five years.

According to a report by Gupta et al., tech companies are expected to require water withdrawals of 4.2 to 6.6 billion cubic meters by 2027 to potentially run and cool their data centers. This would equal 1.7 to 2.6 million Olympic swimming pools.

These figures highlight the urgent need for efficiency and transparency, as AI’s rapid growth risks becoming a significant strain on global energy and water resources.

Google’s Environmental Math

Google reports that the median Gemini Apps text query now yields:

• 0.24 Wh of electricity use, which is equivalent to watching TV for 9 seconds.
• 0.26 mL of water consumption, which is roughly 5 drops.
• 0.03 grams of CO₂ equivalent emissions.

Compared to 2024, Google claims it has reduced power consumption per prompt by 33 times and decreased total emissions by 44 times. 

The research paper does not include 2024 data for per-prompt power consumption, water usage, or CO₂ equivalent emissions. 

To reach its findings, Google’s research team measured consumption of active AI accelerator energy, active CPU and DRAM energy, idle machine energy, and overhead energy. 

Idle machine energy refers to the power consumed by AI computers that are kept on standby. 

They may not be active, but stay ready to handle traffic spikes or failover. The energy these standby AI systems use also adds to AI’s overall energy use. 

In the overhead energy category, Google considered energy consumption by infrastructure supporting data centers, like cooling systems or power conversion. 

How Google Achieved This Feat

Google attributed these efficiency gains to the following combination of strategies:

1. Smarter Model Architectures

Gemini models are based on Transformer design, which is already 10-100 times more efficient than other models. 

Additionally, Google employs efficient techniques, including Mixture of Experts (MoE), intelligent attention computation, and hybrid reasoning to enhance efficiency.

2. Efficient Algorithms & Quantization

Google enhances the efficiency of its AI by continuously refining the algorithms that power AI models. It utilizes methods such as Accurate Quantized Training to reduce energy consumption without compromising response quality. 

3. Optimized Inference and Serving

Google continuously works to make AI model delivery smarter, quicker, and more efficient. It utilizes technology like Speculate Decoding, which offers faster and more cost-effective inference from LLMs. 

It also uses distilling to create more efficient serving-optimized AI models. 

4. Custom-built hardware

Google’s Tensor Processing Unit (TPU) is known for its power efficiency. 

The company has also co-designed TPU and AI models. This enables software and hardware to work together seamlessly, improving energy efficiency. 

In fact, the latest TPU (Ironwood) is 30x more energy efficient than the first publicly available model of TPU. 

5. Optimized Idling

Instead of constantly keeping its AI serving stack on standby, Google employs optimized idling to minimize accelerator downtime. Models are shifted in near real-time based on demand rather than remaining idle unnecessarily.

6. ML Software Stack

Google’s XLA compiler, Pallas kernels, and Pathways system optimize AI models written in frameworks like JAX, allowing them to run efficiently on Google’s custom accelerators.

7. Ultra-Efficient Data Centers

Google’s data centers are highly efficient, operating at a power usage effectiveness (PUE) of 1.09. 

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

Considering the global data center average PUE at 1.56, Google’s data centers are highly efficient. 

8. Clean Energy Procurement

In pursuing its 24/7 carbon-free ambition, Google emphasizes clean energy procurement. 

As a result, Google’s data centers managed to lower their emissions between 2023 and 2024, despite an increase in electricity consumption during that period.

Google vs. OpenAI 

AI enthusiasts often debate which is more energy-efficient: Google’s models or those of its main rival, OpenAI.

In a personal blog post, OpenAI’s Sam Altman discussed ChatGPT’s energy consumption,

“The average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one-fifteenth of a teaspoon.”

Since his blog post doesn’t explain how the numbers were measured or what was counted, we cannot reasonably compare it with other estimates, such as Google’s disclosure explained in this post.

What’s clear is that without transparent, standardized measurements, it will be hard to compare AI models’ efficiency claims. 

Therefore, Google’s detailed research and measurements can serve as a benchmark for assessing AI’s energy use.  

Google Advocates a Unified Approach to Measuring AI Efficiency

Google’s research emphasizes that assessing AI’s environmental impact demands a comprehensive approach instead of a limited benchmark. 

Although the per-prompt impact is minor, the widespread use of AI globally makes continuous efficiency essential. 

Without standardized, transparent measurement frameworks, organizations cannot ensure that AI’s rapid growth aligns with sustainability and accountability goals. 

That said, it’s encouraging to see Google publish its methodology in a research paper for everyone’s benefit, rather than using it solely as a PR stunt.

The research paper concluded;

“We advocate for the widespread adoption of this or similarly comprehensive measurement frameworks to ensure that as the capabilities of AI advance, their environmental efficiency does as well.”