About AI Cost Lab

Why this exists

AI Cost Lab is an independent project built to make the real cost of AI legible — both kinds of cost. Each tool here takes a question that's widely discussed but poorly understood, usually reduced to scary headlines or dismissive hand-waving, and replaces it with a transparent estimate you can inspect and adjust.

Two tools live here today:

• The environmental footprint calculator estimates the energy, water, and carbon of using AI — and it depends on what you actually do: which model, how long the answers are, whether you generate images, and which electricity grid powers the datacenter.
• The cost-to-serve calculator estimates the financial side: the marginal compute cost to run an AI query, shown next to the published API price, so you can see the gap between what a query costs to serve and what providers charge.

Rather than hand you a single dramatic number, each tool shows a modeled estimate with an honest uncertainty range, built from public research.

Our sources policy

Every figure is cited and adjustable. For the environmental tool, the defaults come from peer-reviewed and widely cited public work — Epoch AI's inference estimates, Luccioni et al. (Power Hungry Processing, arXiv:2311.16863), Li et al. (Making AI Less Thirsty, arXiv:2304.03271), and EPA data for grid carbon intensity; the full reasoning is on the energy methodology page. For the cost-to-serve tool, the defaults come from public GPU rental rates, throughput benchmarks (NVIDIA TensorRT-LLM / MLPerf), and the published API prices of frontier, mid-tier, small, and open-weight models; that reasoning is on the cost methodology page.

We deliberately avoid relying on unverifiable vendor claims. Where a number is uncertain, we say so and present a range, and advanced mode lets you override the load-bearing assumptions to match your own situation or a more recent source.

Limitations

These are estimates, not measurements — providers rarely publish per-query figures, so they're inferred and carry wide (roughly 10×) ranges. A few specifics worth stating plainly:

• Environmental: the carbon default uses the global-average grid (~440 g CO₂e/kWh), with a US preset (~348 g) selectable; your real grid may be much cleaner or dirtier. The popular "Google search" comparison rests on a dated 2009 figure, included with that caveat because it is the most-requested comparison.
• Cost-to-serve: it models the marginal compute cost only — it deliberately excludes training, R&D, and free tiers, and it does not compute a profit margin. Throughput is a per-GPU figure (the easiest thing to get wrong by an order of magnitude), and the overhead and prefill assumptions are labeled as assumptions, not measurements.
• Both tools cover inference — running your queries — not the one-time cost of training a model or manufacturing the hardware.

The goal is honest order-of-magnitude clarity, not false precision.

Suggest a correction

This methodology is meant to improve as better public data emerges. If you spot an error, have a newer source, or think an assumption is off, we genuinely want to hear it — corrections and pointers to published figures are welcome and help keep every number defensible. General suggestions and questions are welcome too.