Does Your Brain Beat ai.go?
The real energy cost of asking an AI to do a job—versus doing it yourself
Every time you fire off a prompt, somewhere a data center sips a little electricity. It's become a familiar worry: “Isn't AI terrible for the planet?” But here's a question we rarely flip around — how much energy does you doing the same task cost? Because a human brain isn't free either. It runs on breakfast, lunch and dinner, and these have a surprisingly heavy carbon-and-energy backpack.
So we ran the numbers on a simple thought experiment: one workflow task, done two ways. A person knocks it out in 10 minutes. A GPT-5-class model does it in 60 seconds. Who wins on energy — and by how much? The answer is more interesting (and more nuanced) than either the AI-doomer or the AI-booster camp usually admits.
The setup: comparing apples to slightly different apples
To make an honest comparison, you can't just measure the AI's electricity bill. You have to count the whole chain on both sides. For the human, that means not just the calories the brain burns, but the fossil energy it took to grow, process, transport, and cook the food that fuels those calories. For a typical Western diet, that food-system multiplier is roughly 8× the food's actual calorie content.
Here are the assumptions we settled on:
| Parameter | Value | Note |
|---|---|---|
Human power draw | 100 W | Whole body, not just the ~20 W brain |
Human task time | 10 min | |
Food-system factor | 8 : 1 | Western, meat-inclusive diet |
LLM-modell | GPT-5-size | Frontier-scale |
LLM task time | 60 s | |
LLM inference | 6 Wh | Scales with runtime |
LLM training + hardware | ~1.5 Wh/task | Amortized over billions of queries |
Quick honesty check: we're totally stacking the deck in humans' favor here. For the AI, we're counting all the training costs (hardware, energy, the whole nine yards), but for humans we're only looking at the food supply chain. All the time (and all the meals) it takes a person to actually learn enough to do the task? Yeah, we're just conveniently ignoring that.
Round one: the per-task showdown
Add it all up and the single-task footprint looks like this: the human comes in around 150 watt-hours, the AI around 7.5 watt-hours. That's roughly a 20× difference in the machine's favor.
Energy per task—Human vs. LLM (watt-hours)
One task, done once. The human's bar is dominated not by biology but by the food supply chain behind those calories.
And that leads to the single most surprising finding of the whole exercise: on the human side, the brain is almost a rounding error. The energy isn't really going into thinking — it's going into feeding the thinker.
Where the energy actually goes (watt-hours per task)
The human's 150 Wh is ~89% food-supply-chain overhead. The LLM's 7.5 Wh is mostly live inference compute, with training and hardware nearly invisible once amortized.
Round two: the plot twist nobody mentions
Here's where the AI-is-cheap story gets complicated. That “training energy” line is tiny per task — ;but only because it's spread across billions of queries. Training a frontier model is a genuinely enormous, one-time expense: on the order of 50 gigawatt-hours. That's enough electricity to power thousands of homes for a year, spent before the model answers a single real question.
So imagine a model that only ever did one task. It would be catastrophically inefficient — 50 GWh to do a 10-minute job. The AI only “wins” once it has amortized that upfront cost over enough work. The question is: how much work?
Cumulative energy: when does the AI overtake the human?
The purple AI line starts sky-high (its training bill) but climbs almost flat. The gray human line starts at zero but rises steeply — 150 Wh every single task. They cross at roughly 350 million tasks.
The crossover sits at about 350 million tasks. Below that, a from-scratch model is actually the worse choice on pure energy. Above it, the AI pulls away and never looks back. It sounds like an impossibly high bar — until you remember that popular models field billions of queries. In practice, frontier models blow past break-even in a matter of days.
The takeaway from the curve: AI's energy story is all about scale. A niche model used a handful of times may never justify its training energy. A widely-used one amortizes that cost into near-irrelevance. Utilization is everything.
How solid are these numbers? (Spoiler: wobbly)
| LLM = 4 Wh | LLM = 7.5 Wh (base) | LLM = 15 Wh | |
|---|---|---|---|
| Lean diet (5:1) → human ≈ 100 Wh | ~25× | ~13× | ~7× |
| Western diet (8:1) → human ≈ 150 Wh | ~38× | ~20× | ~10× |
| Heavy diet (10:1) → human ≈ 183 Wh | ~46× | ~24× | ~12× |
Notice something reassuring: in every single cell, the AI still wins. The efficiency edge ranges from a modest ~7× to a dramatic ~46×, but the direction never flips. That's the robust conclusion here — not the exact multiplier, but the fact that the machine is consistently cheaper per task.
So… should we stop worrying about AI's energy use?
No — and this is the important nuance. This comparison measures a very specific thing: the marginal energy of one task, one way versus another. It does not mean AI is shrinking humanity's energy footprint. Here's the catch, and it's a big one:
When something gets cheaper and faster, we don't do the same amount of it — we do vastly more of it. This is the Jevons paradox, and it's the real energy story of AI.
A human might do 20 of these tasks a day. An AI enables millions. So even if each task is 20× cheaper, unleashing 1,000× more tasks means total energy demand goes up, not down. The data-center construction boom happening right now is the physical proof of that math. Per-task efficiency and total-system consumption are two very different conversations — and honest debate needs both.
The bottom line
- Per task, the AI wins big — roughly 20× less energy (≈7.5 Wh vs. ≈150 Wh) under our base assumptions.
- For humans, food is the story. ~89% of the human footprint is the supply chain behind the calories, not the brain itself.
- Training is a huge fixed cost that only makes sense at scale — break-even lands near 350 million tasks.
- The efficiency edge is robust — the AI stays cheaper across every sensitivity scenario we tested.
- But cheaper ≠ greener overall. The Jevons paradox means efficiency can drive total consumption up. The energy debate isn't about one prompt — it's about a billion of them.