At the end of the month, a support team's AI avatar has talked to customers for 500 hours. Most of those hours were ordinary: a delivery date, a password reset, a refund that hasn't arrived.
The invoice from its AI avatar provider has one line that matters, and it is counted in minutes. That line is not an accident of pricing. In the first piece in this series we showed that a cloud avatar keeps data-centre hardware busy for every second of the conversation. Every minute on the invoice is a minute their servers spend drawing your avatar.
This piece follows the same 500 hours through both ledgers: the data centre's and the customer's. What the face stops taking from both, once it moves to the device, is what we call headroom.
Where the face is drawn
A cloud-rendered avatar is a video stream. A GPU in the provider's data centre draws each frame and sends it to the customer's screen, for as long as the session is open. MeshLite draws the face on the customer's own device instead. A small audio model in the browser listens to the voice and picks the mouth shape for each moment, and the browser places the matching pre-made image on the face. MeshLite's servers run no inference for the avatar during the conversation, and no live video stream is sent. We explained why we built it this way in an earlier engineering note.
The look of the face is settled once, when the avatar is created in a one-time GPU job. During the conversation the browser doesn't generate the face; it places images that were already made. So this piece compares where the face is drawn and what that costs, not how it looks. For that, put the avatar on meshlite.io next to any cloud-rendered one you have seen.
Two things still use the cloud in both setups: that one-time creation job, and the speech recognition, language model and voice, which run on servers from providers the customer chooses. They have their own costs. The comparisons below cover ongoing avatar rendering only.
One support desk, one month
Even the most efficient model with a published benchmark holds about a fifth of an NVIDIA A10, an entry-level data-centre GPU, for each live session. At that rate, 500 conversation-hours hold 100 GPU-hours for the faces alone.
| 500 conversation-hours a month | Cloud-rendered avatar | MeshLite, during the conversation |
|---|---|---|
| Data-centre GPU-hours held | ~100 | 0 |
| Electricity, including cooling | ~23 kWh | 0 |
| The same, in phone charges | ~1,060 | 0 |
| Water, including power generation | ~111 L | 0 |
How we got these numbers: 500 hours × 20% of a GPU = 100 GPU-hours. The 20% comes from Hedra's published Livatar-1 benchmark (141 frames a second on one NVIDIA A10); it is a share of the GPU's frame output, not a measured utilisation. We assume the A10 draws its rated 150 W, and use the industry-average data-centre efficiency (PUE 1.56, Uptime Institute 2024), LBNL's water factors (0.36 L per kWh on site, 4.52 L per kWh through electricity) and 0.022 kWh per phone charge (Luccioni and colleagues). We count GPU power only; servers, memory and networking add more. Larger photoreal models on bigger GPUs hold several times as much.
On its own, that is small: about a thousand phone charges a month. Over a year it is 1,200 GPU-hours, roughly 280 kWh and about 1,300 litres of water. What matters is how it grows. It grows with every minute of conversation and with nothing else. A contact centre running 10,000 conversation-hours a month holds about 2,000 GPU-hours, close to three entry-level GPUs running day and night to draw lips and blinks.
The device that draws the face instead is already switched on, and with a cloud avatar it would be decoding a live video stream anyway. We have not yet measured the device's share, so we present the net saving as analysis, not as a measured figure.
The same month in dollars
The fairest comparison is a per-minute plan that covers the same ground as MeshLite: the avatar only, with the customer bringing speech recognition, the language model and the voice. One large AI avatar provider sells one for $475 a month, including 100 hours of session time, and each further minute costs $0.09: $5.40 for every extra hour.
MeshLite Business is a flat $499 a month, billed annually ($5,988 a year; $599 billed monthly), for one company's internal use, with unlimited avatar minutes and no per-minute rendering fee.
| Avatar hours a month | Per-minute plan (avatar only) | MeshLite Business (billed annually) |
|---|---|---|
| 100 | $475 | $499 |
| 200 | $1,015 | $499 |
| 500 | $2,635 | $499 |
| 2,000 | $10,735 | $499 |
List prices, avatar only. Per-minute plan: $475 + (hours − 100) × $5.40. The same provider's plans that include its own voice and language model charge more per minute and are not the same scope.
For our support desk, 500 hours a month comes to $31,620 a year on the per-minute plan and $5,988 on the flat one.
The comparison runs the other way at low volume, and buyers should know where the line is. The two plans cost the same at about 105 hours a month, roughly three and a half hours of conversation a day. Below that, the per-minute plan's included hours are cheaper, and a team that uses its avatar for an hour a day should take them. Against a flat $0.09 a minute with no included hours, the crossover comes earlier, at about 92 hours. Paying MeshLite monthly moves it later, to about 123.
One product, a year of users
The arithmetic changes shape when the avatar sits inside a product a company sells. Usage is no longer one team's working day. It is every user's practice session, onboarding call and follow-up question.
Take a training app with 600 monthly users who each talk to its avatar for about 22 minutes a week, 96 minutes a month. That is 960 conversation-hours a month. Rendered in the cloud, the faces hold about 190 GPU-hours a month, around 2,300 a year, and use roughly 22–45 kWh of electricity a month.
On the same per-minute plan, those 960 hours cost $5,119 a month, or $61,428 a year. MeshLite Starter covers up to 1,000 monthly users in a single product for $11,988 a year ($999 a month, billed annually). The crossover is about 197 hours a month, roughly 125 users at the same 96 minutes. If the app grows from 600 users to 1,000, the per-minute bill grows by two-thirds. The flat fee doesn't move until the product outgrows its tier, and then it steps to the next one.
What a discount can and can't change
These are list prices. Large buyers negotiate, and some providers advertise enterprise rates as low as a cent a minute at volume. A discount moves the dollar line. It does not change the capacity each session holds: for a provider that renders on GPUs, every minute of every session still holds a share of one.
That is what we mean by headroom. Every conversation drawn on the device leaves GPU time, power, cooling and grid capacity in the data centre for work that only a data centre can do. The customer sees the same headroom as an invoice that doesn't follow the minutes.
The cheapest GPU-hour is the one a conversation never asks for.
Next
Every figure in this piece rests on one number we had to estimate ourselves: how much of a GPU a single conversation holds. It decides the energy, and none of the AI avatar providers we checked publishes it. Next, we turn that gap into questions a buyer can put to any provider before signing: which GPU draws the face, how many conversations share it, and what happens to the bill on the busiest day of the year. After that, we look at why a talking face weighs so much more than the chat reply behind it.
Sources: AI avatar providers' published pricing and billing documentation (read 2026-10-19) · MeshLite pricing · Uptime Institute, Global Data Center Survey 2024 · LBNL, 2024 United States Data Center Energy Usage Report · Luccioni, Jernite and Strubell, Power Hungry Processing, ACM FAccT 2024 · NVIDIA L4 and A10 specifications.