Energy Costs of Voice AI
Someone asked me yesterday:
Is Voice AI going to burn the planet?
Profligate use of AI certainly contributes to emissions, especially if we don't put data centres in the right places, but I've worked through the case for Voice AI quantitatively so lets look at the numbers...
We can get ~5-10 concurrent calls on Llama 70B per H100 GPU which has a power draw of 700W. So each call is taking, say, 100W an hour ballpark.
There is a bunch of compute, SIP gateways, tools calls, maybe some extra AI doing RAG or similar lookups. Lets bump that up to 120W per call hour just so I can say 2Wh per call minute!
Now lets look at a human agent working in a call centre...
Commute: average 5 mile commute at 250Wh/mile in an EV (or much worse in an ICE vehicle) = 2.5kWh. Even if they take an e-bike 20Wh/mile, or train 100Wh/mile transport energy is significant. If they take a bus it is actually worse than an EV! Lets say they do a mix of all three and transport is 1.5kWh per agent for an 8 hour shift.
Then they sit at a desk in front of a computer (100W) and Monitor (50W) in a heated and lit office (100W/person) - 250W
So human agent in a call centre "costs" about 3.5kWh for an 8-hour shift or 7Wh a minute.
The AI Agent is over 3x more energy efficient than this scenario by my rough calcs.
Even if I'm out by a factor of 2 of either of those (quite likely), I still think it is fair to say that human agents in call centres (in the west anyway) are likely use more energy than a well implemented AI agent doing the same job.
There are plenty of social impacts around having humans talk to AI rather than other humans, but I'm pretty sure accelerating climate change isn't one of them!
Thoughts?
Energy Costs of Voice AI www.pickering.org/energy-costs...
— Rob Pickering (@rob.pickering.org) 15 May 2025 at 10:26
[image or embed]
