Why I Built an AI Carbon Calculator (And What I Learned)

I started thinking about AI's carbon footprint when I noticed my API bills kept climbing. Every GPT-4 call, every image generation—it all runs on GPUs that burn electricity. Most of us don't think about it because it's invisible.

I built AI Carbon to make that invisible cost visible.

Why Calculate at All?

If you're just using ChatGPT occasionally, the impact is negligible—maybe a few grams of CO2. But if you're running an AI-powered product serving thousands of users, those grams add up.

Some rough numbers I've seen:

• A single GPT-4 API call: ~0.01 kWh (very rough estimate)
• Training a mid-sized model: hundreds of megatons CO2e

For me, it became relevant when choosing between running a local model vs. calling an API. The local option felt "cleaner" but I wanted numbers, not vibes.

How the Calculator Works

I pulled together data from a few sources:

GPU power draw: NVIDIA publishes TDP ratings. H100s run around 700W under load, A100s around 400W. These aren't perfect—actual consumption varies with utilization—but they're a reasonable baseline.

Carbon intensity: This is where it gets messy. The same GPU in Iowa (lots of wind) vs. Texas (lots of gas) has wildly different carbon costs per kWh. I use approximate regional averages.

Workload estimation: Training vs. inference have very different energy profiles. Fine-tuning a model is nothing like running a few queries. I tried to factor in these differences based on papers from MLCommons and various厂商.

What the Numbers Mean

I don't think anyone has this exactly right yet. The field is too new, the hardware too varied, the methodologies too inconsistent. But here's what I'm confident about:

1. Running ML workloads on renewables is meaningfully different than on fossil fuels. The difference between a Texas data center and a Swedish one is real.

2. Training is where the big numbers live. If you're fine-tuning models, that's where your carbon impact concentrates. Inference at scale matters too, but training dominates for most organizations.

3. Hardware efficiency is improving fast. An H100 is way more efficient per FLOP than what existed three years ago. Progress is real.

Practical Takeaways

For most developers: the carbon impact of your AI usage is probably small compared to, say, flying somewhere or heating your home. Don't stress too much.

But if you're building AI products and want to make better choices:

• Consider running batch jobs at night when grids might be cleaner
• Think about whether you need the biggest model for every task
• If you're serving enterprise customers who care about this, look for API providers who publish their carbon data

I won't pretend this calculator gives you precise numbers. It doesn't. But I think it gives you a rough sense of scale, which is better than nothing.

Try the AI Carbon Calculator if you want to estimate your own usage.