The Energy Behind Aventide’s AI

Why we’re publishing this

Aventide runs on leading AI models, Claude today, to think through your business questions and generate your documents. Every AI interaction uses electricity and a little water for cooling, somewhere in a data center. We think small business owners deserve to know roughly how much, in plain terms, not buried in a footnote and not oversold either way.

The honest catch: the AI providers we use don’t all publish exact energy or water figures. Google has, for Gemini, and OpenAI has given an average for ChatGPT. Anthropic, whose Claude models power Aventide today, hasn’t published its own numbers (as of October 20, 2026). So the figures below are our best estimate, built from public research and what our own AI work actually costs us to run, clearly labeled as an estimate, not a measurement. We’ll update this page as our model mix changes and as providers publish harder data.

What this looks like at Aventide’s usage levels

Every action in Aventide, a question you ask or a document you generate, uses credits, and credits follow what that action actually cost us in AI. The two plan cards at the top of this page show what a whole month’s credits add up to if you spent them all on messages to Aventide, which is the most energy-hungry way to use them. Most people use far less.

For context: a typical U.S. home used about 865 kWh of electricity a month in 2024, or about 28 kWh a day (U.S. Energy Information Administration). Even Pro, used to the last credit, is under 1% of that month, and Starter is at most about a fifth of 1%. Most Aventide customers use well under their full monthly allowance, so real use is usually a fraction of the cards above.

Per-interaction, roughly

• A plain question to a chatbot uses about 0.24 to 0.34 Wh, going by the figures Google, OpenAI and Microsoft have published. An Aventide message uses more, because each one carries your business’s facts, the conversation so far and the tools Aventide can use, and a bigger request often takes several steps.

• A quick message (a lookup, logging an expense): roughly 0.4 to 1 Wh.

• A typical message: roughly 2 to 5 Wh.

• An in-depth message (several steps, or web research): roughly 5 to 13 Wh.

• A typical generated document: roughly 0.4 to 3 Wh. That’s often less than a message, because a document is one AI call without the chat’s tools. Market Research, which reads and writes a lot more, is roughly 4 to 16 Wh.

• Water: from about 10 drops for a quick message to about a tablespoon for a typical one, and up to a few tablespoons for in-depth work. Where in that range depends on how the data center is cooled, and on whether you count the water power plants use to make the electricity.

In everyday terms

If numbers in watt-hours don’t mean much to you, here’s the same thing in things you already use.

• Driving an electric car: about 2 to 5 miles for Starter, about 6 to 14 miles for Pro, every credit used.

• Running the shower: under a minute for Starter, under 3 minutes for Pro, every credit used.

The car is the middle of the EPA’s ratings for 2026 electric cars, about 37 kWh per 100 miles, charging included. The shower is the U.S. average, 2.1 gallons a minute for 7.8 minutes (Water Research Foundation), so even Pro used to the last credit is at most about a third of one average shower.

How we calculated this

We started from what our own AI messages cost us to run, measured on 198 real messages between September 3 and October 3, 2026. To turn cost into energy, we compared a plain chatbot question’s published energy (0.24 to 0.34 Wh) with what that same question would cost on the Claude model Aventide uses. That works out to about 25 to 65 Wh for every dollar of AI work. It assumes energy rises with the price of the work, which is a reasonable guess, not a measurement, so treat every figure here as a range that could be off by a factor of two either way.

For water, the low end is the cooling water used at the data center itself, about 1.1 liters per kWh in Google’s figures. The high end adds the water used to generate the electricity, about 4 liters per kWh at the U.S. average (UC Riverside). We don’t count training the models, making the hardware, or your own device.

What we do to keep the footprint small

• We choose AI models sized to the task rather than defaulting to the biggest, most compute-heavy model available.

• Videos stay in your Library until you delete them, and uploads that never finish are cleared within 24 hours.

• Our infrastructure scales down to near-zero when you’re not using it. No AI model sits idle waiting for you.

Sources

• Google, “Measuring the environmental impact of delivering AI at Google Scale” (August 21, 2025)

• MIT Technology Review, “In a first, Google has released data on how much energy an AI prompt uses” (August 21, 2025)

• Sam Altman, OpenAI, “The Gentle Singularity” (June 10, 2025)

• Oviedo et al., Microsoft Research, “Energy use of AI inference, efficiency pathways, and test-time scaling”, Joule (April 2026)

• Epoch AI, “How much energy does ChatGPT use?” (February 7, 2025)

• Li, Yang, Islam and Ren, UC Riverside, “Making AI Less Thirsty” (2023, revised March 26, 2025)

• Mistral AI, “Our contribution to a global environmental standard for AI” (July 22, 2025)

• U.S. Energy Information Administration, residential electricity use per customer in 2024 (May 12, 2025)

• U.S. EPA and Department of Energy, fuel economy ratings for 2026 electric cars (as of October 20, 2026)

• Water Research Foundation, “Residential End Uses of Water, Version 2” (April 2016), via the Alliance for Water Efficiency

As our AI provider mix expands, we’ll update these figures and link directly to any provider’s own disclosure the moment one exists.

Questions about energy

How much electricity does a month of Aventide use?

Used to the last credit, Starter comes to about 0.7 to 1.8 kWh a month and Pro to about 2.1 to 5.3 kWh. Even Pro is under 1% of the roughly 865 kWh a typical U.S. home used a month in 2024.

How much does one Aventide message use?

A quick message uses roughly 0.4 to 1 Wh, a typical message roughly 2 to 5 Wh and an in-depth message roughly 5 to 13 Wh. A typical generated document uses roughly 0.4 to 3 Wh.

Where do these figures come from?

They are our best estimate, built from public research and what our own AI work costs us to run. We update this page as our model mix changes and as providers publish harder data.