Energy & cost

What does your local AI actually cost?

Pick your GPU and region. We’ll show monthly electricity, CO₂, and how it compares to just calling a hosted API.

Setup

GPU / chipRegionUsage pattern

Duty cycle: 60% at load

Of the running hours, what fraction is actually generating tokens? Most chat use is <30 %.

Throughput: 70 tok/s while at load

Energy / day: 1.06 kWh
Energy / month: 31.9 kWh
Electricity cost / month: $5.27
CO₂ / month: 12.5 kg
Effective $/1 M output tokens: $0.290
Pure electricity cost per million tokens generated. Excludes hardware amortisation.

A typical chat exchange ≈ 700 tokens. At 70 tok/s with 60% duty, you generate one such call every 16.7 seconds. Compare to API call costs:

DeepSeek V3

$0.00036 / call

$9.20

if you sent the same volume to API

Local saves $3.93/mo

Llama 3.1 8B (Groq)

$0.00004 / call

$1.06

if you sent the same volume to API

Local saves $-4.21/mo (you'd pay more locally)

Claude Sonnet 4.6

$0.00450 / call

$116.64

if you sent the same volume to API

Local saves $111.37/mo

GPT-4o

$0.00325 / call

$84.24

if you sent the same volume to API

Local saves $78.97/mo

Math is for a single user. If you run a small server with many concurrent requests, local economics get much better.