The Economics of a Prompt: Google Reveals the True Cost of AI Queries

Google has published the first major disclosure of AI query energy consumption, revealing that a median AI-assisted search consumes 0.24 watt-hours—roughly equivalent to one second of microwave operation. The number is smaller than expected, but the methodology deserves scrutiny.

The Energy Architecture

Google’s breakdown shows AI chips consuming 58% of total query energy, with the remaining 42% supporting infrastructure — as explored in the economics of AI compute infrastructure — : processors, backup systems, and cooling. The company reports a remarkable 33x efficiency improvement from May 2024 to May 2025 through model optimization and hardware advances.

Variable costs matter significantly. Complex reasoning tasks and image generation consume multiples of the baseline, while routine cached queries cost substantially less. This variance creates interesting business model implications for AI pricing.

The Carbon Accounting Question

Google claims 0.03 grams of CO₂ per prompt—but this figure uses market-based accounting that incorporates renewable energy purchases. Actual grid emissions, without offsets, would be approximately three times higher.

This isn’t deception; it’s standard corporate carbon accounting. But it illustrates why second-order thinking matters when evaluating sustainability claims. The renewable energy credits Google purchases represent real investments, but they don’t change the electrons flowing through data center circuits.

The Missing Variable

Google hasn’t disclosed total query volume. Without this number, calculating cumulative environmental impact is impossible. A 0.24 Wh query multiplied by billions of daily queries produces very different conclusions than the per-query figure suggests.

This selective disclosure follows a pattern common in corporate communication: share impressive unit economics while obscuring aggregate scale. The efficiency gains are real; the total footprint remains opaque.

For investors and regulators, the question isn’t whether AI is becoming more efficient—it clearly is. The question is whether efficiency gains outpace volume growth. Google’s disclosure, while unprecedented in detail, carefully avoids answering that.