Capacity, for a live avatar, means three things: how many conversations can run at the same time, how long each one can last, and what each additional one costs. Every provider answers these three questions in its plans. Few explain the reason behind the answers.
The reason, in most cases, is a GPU. In the first piece in this series we measured a cloud-rendered avatar at about a fifth of an entry-level data-centre GPU for every second the conversation is open, including the seconds it spends listening. Minutes, session limits and concurrency caps are different ways of selling that same held hardware.
This year the question has a second edge. The largest clouds say customer demand for AI capacity is running ahead of what they can build, and the hourly rental price of NVIDIA's H100, a chip launched in 2022, is still rising. When GPUs are scarce, every conversation that holds one competes with everything else that needs one, and a provider's capacity is only as good as the GPUs behind it.
The questions below are the ones we would ask if we were buying. Each comes with the reason it matters and what a clear answer sounds like.
1. What do you hold while a conversation is open?
Ask what runs on the provider's side for every live session: a GPU, a share of one, or nothing. Then ask for one number: data-centre GPU-hours per 1,000 conversation-minutes.
It is the number behind every other line on the invoice, and in our search none of the large AI avatar providers published it. A provider that knows its cost of service can tell you. Our measurement for a cloud-rendered avatar is about 3.3; larger photoreal models hold more.
2. Where do your GPUs come from, and are they reserved for my traffic?
Some providers own their GPUs, some rent them on long contracts, and some buy them by the hour on the open market. Ask which, and whether the capacity your plan promises is reserved for you or shared with every other customer's busiest hour.
Then ask what your price rests on. A per-minute rate is built on the provider's GPU cost, and that cost is not falling this year. Ask whether your contract fixes the rate for its whole term, and what happens at renewal.
3. Does the price follow minutes, sessions or users?
These grow with different things. Minutes grow with every conversation and with how long it runs. Sessions grow with your busiest hour. Users grow with your product's success.
Ask which one your plan follows, and what happens when you pass the allowance: a charge for every extra minute, a step up to the next plan, or a conversation about the contract. The first grows with each conversation. The others grow in steps you can see coming.
4. Is listening time billed?
In a conversation, the avatar spends much of its time listening. Many plans bill the whole session, silence included, because the face is still being drawn while the customer talks.
Ask what counts as a billable minute: the time the avatar speaks, or the time the session is open. Then ask what happens to a session nobody closes.
5. How long can one session last?
Some plans end a session after 20 or 60 minutes. That rarely matters for a short support chat. It matters a great deal for a reception screen that stays open all day, a long training role-play, or a tutor a student keeps open for an evening. A session that ends has to reconnect, and the customer sees it.
6. What happens on your busiest day?
Ask for the number of simultaneous conversations your plan allows, and what happens at the limit: does the next customer wait, get turned away, or get billed at a higher rate? Ask whether that number is contracted capacity or an advertised one, and whether it has been load-tested.
Plan for your peak, not your average. A launch, a sale or the first Monday of a training programme is when an avatar is most valuable and when a limit is most likely to be reached.
7. If we negotiate a discount, what changes?
Large buyers negotiate, and some providers advertise rates as low as a cent a minute at volume. A discount moves the price per minute. It doesn't move the hardware: the provider still holds a share of a GPU for every minute of every session, and its price has to cover that at whatever rate.
So ask what the discount is tied to. A lower rate in exchange for a minimum commitment is still a price that follows your minutes. If you already have a quote, compare it with the volume you expect next year, not the volume you have today, and ask how long the discounted rate holds.
8. Where is cloud rendering the right choice?
A fair buyer's guide says when the other answer wins. Cloud rendering makes sense when usage is low and occasional, so a plan's included minutes cover it. It makes sense when the provider bundles a voice and language model you want to buy from one place. And it makes sense when the device on the customer's side is too old to draw the face itself.
Our answers
In a GPU shortage, every GPU-second spent drawing a face is a GPU-second not spent on intelligence. That is why we built MeshLite to draw the face on the user's device, and why we think a provider should be able to answer these questions in one table. Here are ours.
| Question | MeshLite |
|---|---|
| Held during a conversation | No data-centre GPU for the face. The avatar renders on the user's device; MeshLite's servers run no inference for it and send no live video stream. |
| Where the GPU comes from | There is none to source for the face during the conversation: it renders on your users' devices. Speech recognition, the language model and the voice depend on your providers' capacity. |
| Price follows | A flat plan with unlimited avatar minutes. Each plan sets a capacity allowance: how many people can talk to the avatar at the same moment, for a company's own use, or how many monthly avatar users, for a product. |
| Listening time | No minute is billed. An open session counts toward the allowance whether anyone is speaking or not. |
| Session length | Sessions last as long as you need. |
| Busiest day | The allowance can be expanded, or you move to the next plan. Growing past it is never a per-minute charge, and there is no automatic overage. |
| What still costs money elsewhere | Speech recognition, the language model and the voice, from providers you choose, billed by them. Each avatar is created once, in a one-time GPU job. |
| Devices | Current browsers and devices; very old hardware may not run it. |
Next
Every question here leads back to the same fact: a live avatar's face is the heaviest thing in the conversation. Next, we look at why, comparing what a chatbot, a pre-recorded video and a live avatar each ask of a data centre, and why, as compute gets tighter, each job is moving to the smallest hardware that can do it.
Sources: Off the Grid 1, measurement and method · Microsoft, FY2026 Q4 earnings call, 29 July 2026 · Amazon, Q2 2026 earnings call, 30 July 2026, as reported by Fortune · Silicon Data, H100 Rental Price Index · AI avatar providers’ published pricing and documentation, read 2026-10-01 (session limits, concurrency allowances, silence billing, volume rates).