Counting AI Efficiencies in Bananas

My kids asked me at dinner whether AI is more efficient than people. I build with AI every day, and I didn’t have a straight answer. So I built an energy calculator to find out, and measured everything in bananas.

A medium banana holds about 105 Calories, roughly 122 watt-hours. That’s about an hour of sitting and thinking at a desk, 8–9 minutes of running, or about 31 long AI reasoning answers.

Here’s what I told my kids: AI wins on short, well-defined tasks, long agent work is close enough that how you count picks the winner, and a person wins whenever judgment stops wasted repetition. The IEA expects data centre electricity use to more than double by 2030. Anyone defending AI’s energy use to their leadership or in a sustainability report needs a comparison that survives a skeptic.

Each ratio below is how much further one banana’s worth of energy gets the winner. The first number counts only the person’s extra effort above resting, which is the calculator’s default. The number in brackets counts the person’s whole metabolism.

Where AI wins

• Restaurant tip: AI ~41x (~75x).
• Plan 6 errands, a 10-minute puzzle: ~2.7x (~4.8x). Whole body, the person uses 18.9 Wh and the AI 3.91 Wh.
• 1st-grade math homework: ~2.3x (~4.1x).
• A page of advanced math: ~2.0x (~3.6x).

AI also wins on time almost everywhere.

Where it’s a tie

Here the winner flips with how you count:

• Fix a software bug: person ~1.22x counting only the extra; AI ~1.48x whole body.
• Research brief from 150 sources: person ~1.33x / AI ~1.36x.
• A full 8-hour workday: person ~1.22x / AI ~1.49x.

Long agent work keeps re-reading a big context. On the bug fix, the agent reads about 2 million tokens in an hour, mostly cached.

Where the person wins

Take an agent stuck in a loop on a flaky test that a person fixes in 20 minutes. The person goes ~14x further (~7.7x) and wins on time and cost too. Judgment beats brute force.

On price, the person wins every task except the tip: ~44x on the bug fix, ~67x on the research brief. A banana costs about 25 cents (BLS average, $0.65 a pound, roughly 0.4 lb each). An AI answer’s price covers hardware and margin, not just electricity. Price is not energy.

What tips the race to the person

You can try each of these in the calculator:

• Judgment: knowing when to stop avoids the AI’s retries.
• Counting only the extra effort: the calculator default.
• Retries: each one repeats the whole exchange, so energy per correct answer favors the person.
• Heavy thinking on easy work: a higher think level multiplies the tokens written.
• A bloated context: a lean one helps the AI.
• A skilled person: finishing in a third of the time triples their efficiency.

Ask what’s inside the number

Google reported 0.24 Wh for its median Gemini Apps text prompt, including idle machines, host CPU and memory, and data centre overhead. Active chips alone come to 0.10 Wh. Sam Altman wrote that the average ChatGPT query uses about 0.34 Wh without publishing a method. Find out what a per-query figure includes before you compare it.

For the Nerds

• Person: METs (multiples of resting energy use) come from the 2024 Adult Compendium: computer work 1.3, running at 6 mph 9.3. At 1 MET ≈ 1 kcal/kg/hour, 75 kg at 1.3 MET for 10 minutes ≈ 18.9 Wh. Neither view isolates the energy of thinking.
• AI: Oviedo et al. (Joule, 2026) model 0.31 Wh for a typical frontier query and 3.91 Wh for a long reasoning answer (~5,000 output tokens). The calculator uses 0.000782 Wh per output token. Modeled, not measured.
• Banana: 105 kcal (USDA) × 4,184 J/kcal (NIST) ÷ 3,600 J/Wh ≈ 122 Wh.

I’m still building a model for energy per correct answer at matched quality. If you’re measuring this too, let’s compare notes.

Try your own task: the calculator takes custom token counts, and you can change any number in its “For the nerds” section. Nothing you type is stored on a server.

Where to start

• Good (this week): Run one AI task through the calculator in both counting views.
• Better (this quarter): Judge agent work by energy per correct answer, tracking retries and context.
• Best (this half): Write down each number’s boundary before AI energy reaches a board report.

---

David O’Neil is a CISO who builds things, including this calculator. If your numbers say otherwise, tell him.