CarbonWatch
A 500-token answer produced by Z.ai costs about 1 Wh, or 0.58 gCO₂e. Z.ai is a chat service whose inference runs in China, where power is worth 580 gCO₂e per kilowatt-hour.
What CarbonWatch records
| Kind of use | Chat |
|---|---|
| Inference region | China CN |
| Grid intensity | 580 gCO₂e/kWh |
| Energy rate | 2 Wh / 1 000 tokens |
| Bytes per token | 25 |
Three reference points
The estimate is linear: these three rows are the same calculation at three scales. They are orders of magnitude, not measurements.
| One 500-token answer | 1 Wh | 0.58 gCO₂e |
|---|---|---|
| A hundred answers — a working day with an assistant | 100 Wh | 58 gCO₂e |
| A thousand answers | 1 kWh | 580 gCO₂e |
The same usage elsewhere
The same exchange, the same model, billed at another region's grid. This is the point in the calculation where the choice of provider weighs heaviest — more than the length of the answers.
| Switzerland | 40 g/kWh | 0.04 gCO₂e |
|---|---|---|
| France | 50 g/kWh | 0.05 gCO₂e |
| European Union | 250 g/kWh | 0.25 gCO₂e |
| United States | 380 g/kWh | 0.38 gCO₂e |
| World | 480 g/kWh | 0.48 gCO₂e |
| China this service | 580 g/kWh | 0.58 gCO₂e |
How it is detected
Outgoing connections to these domains, resolved into addresses. The comparison is made on the address, never on the hostname: nothing is decrypted, nothing is read.
- chat.z.ai
What this figure does not say
The scope is operational: the electricity of inference, and nothing else. Not the manufacturing of the servers, not the training of the model amortised over requests, not the transport network.
The rate is the one for the service type, not for the exact model you called: CarbonWatch does not read content and cannot tell whether an answer came from a small model or a large one. A long chain of reasoning costs more than what is counted here.
The ratio between bytes received and tokens produced is the most uncertain parameter of the model. It is adjustable, service by service, in the registry.