AI's Real Cost Isn't Your API Bill — It's Carbon
Google's and Amazon's emissions are climbing because of AI. Here's why the hidden cost matters to a Sri Lankan builder, and how to spend fewer tokens.
The real cost of AI showed up this week not on a billing dashboard but in two corporate climate reports. According to TechCrunch's "A warning sign about AI's real cost, courtesy of Google and Amazon", both companies' carbon emissions are rising fast, and AI is the reason. That matters to me as someone building on a small budget in Sri Lanka, because the same forces pushing their emissions up are the forces inflating my token bill.
Here is the part I keep coming back to: the price of a model call and the carbon behind it move together. Cut one and you usually cut the other.
📊 The numbers that should make you stop
Two of the largest cloud providers on the planet made net-zero promises, and AI has now put those promises out of reach. The figures from their own reporting:
| Company | Emissions change | The tell |
|---|---|---|
| Up 25% since last year | Scope 3 (indirect) emissions have doubled since its 2019 baseline | |
| Amazon | Up 16% | Added 1.2 GW of data center capacity in Q4 alone |
Google's indirect emissions alone rose by 2.1 million metric tons in a single year. Amazon's own words in the report were blunt:
"To meet strong customer demand, in 2025 we added more data center capacity globally than any other company, including more than 1.2 gigawatt (GW) in Q4 alone."
That 1.2 GW is roughly the output of a large power station, built into a single quarter, to serve AI demand.
🏗️ Why buying "green energy" doesn't fix it
The instinct is to assume a data center running on solar or wind is clean. The report's most useful point is that the emissions problem sits mostly in Scope 3 — the indirect stuff a company can't offset by buying renewable electricity. Where it comes from:
- Data center construction — steel and cement are among the dirtiest materials to produce.
- Chip manufacturing — GPUs and semiconductors are fabricated in Asia on grids that still lean heavily on fossil fuels.
- Chemicals in chip fabs — the process gases used to etch chips are potent greenhouse contributors.
- Natural gas for peak demand — when AI load spikes, gas plants get fired up to keep up.
Key takeaway: The carbon cost of AI is baked in before a single prompt runs — it's in the concrete, the silicon, and the grid. Efficient use of the model you already have is the lever an individual actually controls.
You and I can't change how a fab in Taiwan is powered. But every wasted token is demand we added to that pile for no benefit.
💰 The token-carbon link, from where I sit
I run small AI tools out of Sri Lanka on free tiers and thin margins. When I profile my own usage, the biggest waste is always the same: sending huge prompts, retrying blindly, and reaching for a frontier model when a small one would do. Each of those is both money and carbon.
A rough sense of the trade-offs I make in practice:
| Habit | Cost impact | Carbon impact |
|---|---|---|
| 8k-token prompt when 800 would do | ~10x the input cost | ~10x the compute |
| Frontier model for a classification task | High per call | High per call |
| No caching on repeated queries | Pay every time | Compute every time |
| Small/local model for simple jobs | Cents or free | Fraction of the energy |
None of this is sacrifice. A tighter prompt is usually a clearer prompt, and a smaller model that fits the task is often faster too. If you want to see how quickly per-call costs add up at scale before you commit to an approach, our AI cost calculator is a quick way to sanity-check the math.
🛠️ What a small builder can actually do
You don't need a sustainability team to spend less compute. The moves that cut my bill are the same ones that cut the footprint:
- Right-size the model. Reserve the biggest models for genuinely hard reasoning. Classification, extraction, and formatting rarely need them.
- Trim context ruthlessly. Send only what the model needs. Long, padded prompts cost the most and help the least.
- Cache and batch. Repeated or overnight work doesn't need to hit the API live every time.
- Prefer on-device where it fits. Small models running client-side, in WebAssembly or on a phone, move zero data to a data center.
- Measure before optimizing. Log your token usage for a week. You'll find the waste faster than you'd guess.
The cheapest, cleanest AI call is the one you didn't need to make. Solve it with a regex or a lookup table first, and reach for a model only when the problem genuinely calls for one.
🌐 What this means for you
If two of the richest companies in the world can't AI their way to net zero, the takeaway isn't despair — it's that efficiency is now the whole game. For a Sri Lankan student, freelancer, or small team, that's genuinely good news. The exact habits that keep AI affordable on a learning budget are the same habits that keep its footprint down.
Treat every token as something you're paying for twice: once in your currency, once in carbon. Build small, measure often, and reach for the frontier model only when the problem earns it. The companies in this story are being forced to reckon with the bill. Starting from a tight budget, you already have the right instincts — use them.