Some AI companies publish carbon-neutral badges based on offsets we don't trust and metrics that don't survive scrutiny. This page is what we can honestly say instead.
What we don't do
- We do not claim carbon neutrality. Purchased offsets from most providers are of contested effectiveness; we don't want to make a claim we can't defend.
- We do not have our own datacenters. LADLE runs on Vercel (which runs on AWS) for edge + on Anthropic's infrastructure for inference. We don't control the underlying power sources.
- We do not have a published per-message CO2 figure. Anthropic hasn't published inference-level energy data; we'd be estimating from academic papers with 10x error bars.
What we do know
- Anthropic's inference runs primarily on AWS regions that are increasingly matched with renewable energy (AWS committed to 100% renewable by 2025; the current match rate is public but varies by region).
- Vercel's edge network runs primarily on renewables via their infrastructure providers.
- Per-message energy cost is estimated at ~0.3 to 3 Wh (wide range because it depends on message length and model activity). For comparison, a Google search is ~0.3 Wh and streaming a minute of HD video is ~50 Wh.
- LADLE's total datacenter power draw scaled to subscriber count is meaningfully smaller than the same subscribers' streaming or video-conferencing usage over the same period.
Rough magnitudes
Where the true cost is higher
Model training is where most of the industry's headline energy numbers come from — training a frontier model consumes megawatt-hours. But LADLE doesn't train models; we run inference on models Anthropic already trained. The training cost is amortized across all Anthropic API customers globally.
The one thing we do
Subscribers who ask us to reduce our infrastructure footprint have a clear path: use LADLE less. Fewer messages = less inference = less energy. We don't gamify usage or engagement-optimize; if you use LADLE only when you need it, that's the greenest usage pattern. We won't nudge you to send more messages.