Picture a customer on a support page, checking how long a refund takes. She could read an AI-written explanation, watch a recorded presenter explain the process, or hear the explanation from a live AI avatar. All three can explain the same thing. But behind the screen, they ask the machines to do different work.
What is already done
The written explanation takes computing to produce. The language model reads the question and generates the words. Once they are on the page, the customer can pause, read them again or leave them open while she checks her order. The model does not have to generate those words again just to keep them visible.
A recorded explanation has done another kind of work in advance. The presenter’s voice, mouth movements and gestures are already in the video. The file still has to reach the customer, and her device still has to decode and display it. But watching the explanation does not require a server to make the presenter perform it again.
The text is ready to read. The performance is ready to play.
What has to happen now
Now the customer asks whether weekends count.
A text chatbot can answer that follow-up. A live AI avatar can deliver an answer too. The ability to respond comes from the conversation system behind them, not from having a face.
What the avatar adds is a moving picture. Its mouth has to follow the words of the new answer. In a cloud-rendered system, the provider processes the mouth movements and draws the face on its servers, then streams the resulting video to the customer’s screen.
The recorded presenter moves too. But those movements were made before the customer pressed play. The live face has to move with what is being said now.
A live face usually keeps moving while the customer talks, too: blinking, nodding, listening. In a cloud-rendered system, that drawing doesn’t stop between answers.
| On the screen | What is already made | What happens during use |
|---|---|---|
| An AI-written explanation | The words, once generated | Display the answer; generate another if the customer asks a new question |
| A recorded explanation | The presenter’s picture and sound | Deliver, decode and play the video |
| A cloud-rendered live avatar | No finished video of the new answer | Process the mouth movements, draw the face and stream the live picture |
Once the chatbot has written its answer, the words can stay on the page. While the live avatar delivers its answer, the picture has to keep changing.
Where the face is drawn
A live face has to be drawn somewhere. It doesn’t have to be in a data centre.
MeshLite draws it in the customer’s browser. The mouth is matched to the voice and the face is put together on her own phone or laptop, so no separate video of the avatar is streamed to her from MeshLite. For each conversation, MeshLite’s servers do no lip-sync and no drawing of the face.
That shows up on the invoice too. Within the capacity agreed in the licence, MeshLite charges no per-minute fee for drawing the face. The voice, the language model and the rest of the product still have their own costs and limits.
Back on the support page, the customer only wants to know when her money arrives. The written answer could sit on the page. The recorded video was made last month. The live face has to be made while she watches. So for any product that wants one, the question isn’t only how the face looks. It’s where that face is drawn for every second it speaks, and who pays for it.
Next
Knowing where the face is drawn is the start of an infrastructure claim, not the end. Next, we look at what it would take to check a data centre’s electricity and water figures, and why an average for a country cannot tell a buyer what happened at one facility.
Further reading: Off the Grid 1 — where live avatar rendering happens · Off the Grid 2 — the avatar-rendering bill · MeshLite public plans.