Electricity use in the United States is rising at its fastest pace in decades. So far in 2025, power consumption is up nearly 4% compared to the same period last year. This shift follows years of flat demand and is largely tied to the surge in data centers, which now underpin the artificial intelligence (AI) boom.
But the growth has come with trade-offs. A significant portion of this added demand is being met with coal, whose share in U.S. power generation climbed 20% year-over-year as of May. This reliance raises concerns that the rapid adoption of AI technologies could be undermining global efforts to cut carbon emissions.
Measuring AI’s True Environmental Impact
Assessing AI’s environmental footprint is not straightforward. Researchers can measure the energy required by specific models, but extending those results to global-scale usage is more difficult. Real-world operations depend on variables like cooling systems, memory, storage, and how often AI models are called on in consumer applications.
That lack of clarity has left much of the conversation speculative. Unlike outside analysts, companies like Google have access to comprehensive data covering every layer of their operations—from server utilization to the energy intensity of electricity supplies. This puts Google in a rare position to provide more concrete numbers.
Google’s New Analysis
In a recently published report, Google shared detailed findings about AI’s environmental footprint, offering a rare look behind the scenes. The company said the energy required for a typical AI search prompt has dropped 33-fold over the past year.
The analysis factored in several elements:
- CPU, memory, and AI accelerator usage during both active and idle times.
- Cooling systems and water consumption at data centers.
- Proportion of data center resources allocated to AI.
- Carbon emissions tied to electricity supplies.
- Environmental costs linked to hardware manufacturing.
By including both Scope 2 emissions (direct emissions from energy use) and Scope 3 emissions (indirect emissions from supply chains), Google positioned the study as one of the most complete assessments to date.
Key Omissions
Despite its depth, the study excluded several areas. It did not evaluate the environmental toll of the network infrastructure that carries data between users and servers, which varies widely by connection type. Nor did it measure the energy consumed by end-user devices, such as smartphones or PCs, which also play a role in total usage.
The most significant omission was the energy-intensive process of model training. AI training can consume enormous amounts of power before deployment, yet those figures were left out of Google’s calculations, even though the company likely has access to the data.
Per-Request Numbers
To calculate typical values, Google monitored its data centers for 24 hours, recording activity and idle periods. For a median Gemini Apps text request, the company found:
- 0.24 watt-hours of electricity consumed
- 0.03 grams of carbon dioxide emissions
- 0.26 milliliters of water used (about five drops)
For perspective, the energy consumption of a single AI prompt is similar to nine seconds of television viewing.
While negligible at the individual level, the sheer volume of AI requests—since Google now processes them with every search—creates a significant cumulative footprint.
Efficiency Improvements Drive Progress
The dramatic drop in energy use per request stems from both cleaner energy sources and better technology. Expanded solar power has reduced the carbon intensity of Google’s electricity by 1.4x in the past year.
More impactful, however, are the software and hardware improvements:
- Mixture-of-Experts (MoE): Activates only necessary portions of a model for each query, cutting computation by 10 to 100 times.
- Compact models: Smaller versions of core AI systems reduce overall load.
- Custom AI accelerators: Hardware built in-house and paired with optimized software improves efficiency.
- Data center management: Ensures high utilization of active systems, while allowing idle hardware to run in low-power states.
Since AI accelerators account for more than half of per-query energy use, these optimizations have produced outsized results.
Google argues that these efficiency gains mean the energy costs of building and running its infrastructure are spread across many more processed queries, lowering the per-request footprint.
Unlike a simple marketing statement, the report was presented in an academic-style format, with the company emphasizing methodology. Google suggested that similar measurement frameworks should be adopted industry-wide to ensure AI’s environmental efficiency keeps pace with its rapid technological growth.
