In short: A cloud avatar keeps data-centre hardware busy for every minute of the conversation, including while it listens. Even the most efficient model with a published benchmark holds about a fifth of a data-centre GPU per conversation: about 3.3 GPU-hours per 1,000 conversation-minutes, at a time when data centres are short of power. MeshLite uses none: the face is drawn on the visitor’s device. The voice and the language model still run in the cloud.
Picture a customer on a support page, talking to an AI avatar for ten minutes about a refund that hasn’t arrived.
The language model behind the conversation works in bursts. It reads a question, writes an answer and goes quiet. The face never gets to rest. With a cloud-rendered avatar, a server in a data centre keeps drawing it for all 600 seconds: every blink, every nod, every frame of the minutes when the customer is talking and the avatar is only listening.
A chatbot uses a GPU in bursts, for a second or so per answer. A live avatar drawn on cloud GPUs needs them continuously, for every frame of the conversation. Most AI avatar providers bill by the minute.
The constraint has moved
For most of the AI boom, the scarce things were ideas and chips. Increasingly, the scarce thing is the socket in the wall.
- Gartner's June 2026 forecast: ~565 TWh in 2026 (all data centres including AI). IEA's central case: ~945 TWh by 2030 — more than double 2024's 415 TWh.
- Keeping them cool takes water. US data centres consumed 66 billion litres directly in 2023, and nearly 800 billion litres more through the electricity they bought, according to Lawrence Berkeley National Laboratory. That is consumption — water not returned to the local system. A Ceres analysis from August 2026 counts it a second way: data centres in seven US states depend on about 3.4 trillion gallons of freshwater withdrawals a year for their electricity — water drawn through power plants, most of it returned. Two-thirds of the water-dependent electricity generated in those states comes from regions under medium-high to extreme water stress. Withdrawals are not consumption, but they measure exposure, and Ceres projects 4.1 to 7.6 trillion gallons by 2030.
- New power is slow to arrive. In the US, the median project now waits about five years between asking to connect to the grid and switching on, up from under two in the early 2000s.
Every hour a GPU spends on one job is an hour of electricity, cooling and grid capacity that nothing else can use.
Every frame, for the whole conversation
Making images costs far more energy than making text. In the most widely cited measurement, generating images used on average more than 60 times the energy of generating text for the same number of requests (Luccioni and colleagues, ACM FAccT 2024).
A live avatar’s model is far lighter per frame than an image generator: about a joule per frame for an efficient model on an NVIDIA A10, against about 10,000 joules for one generated image. But it runs for every frame of the conversation, 25 times a second. The bill reflects it. AI avatar providers charge by session time, and some count the silences too.
Counting the faces
How much data-centre capacity does a talking face actually hold? Of the AI avatar providers we checked (LiveAvatar, Tavus, Anam, bitHuman and Spatius; pages read 1 October to 3 November 2026), none publishes how many conversations share one GPU, or the energy per minute. Anam names its GPU (an NVIDIA H200); LiveAvatar says its renderer runs on CPUs. So we worked it out from the most efficient model with a published benchmark.
Hedra’s Livatar-1 renders 141 frames a second on one NVIDIA A10, an entry-level data-centre GPU. A live face needs 25 to 30 frames a second, so at that rate one A10 draws about five conversations at once, and each holds about a fifth of the GPU. A thousand conversation-minutes is 16.7 hours of sessions, which works out at about 3.3 GPU-hours spent drawing faces. Common open models, and larger photoreal models on bigger GPUs, hold more.
| Efficient model: a fifth of an NVIDIA A10 (150 W) | Upper bound: one session per H100 SXM (700 W), not a measured provider | |
|---|---|---|
| GPU-hours held | ~3.3 | ~17 |
| Electricity, including cooling | ~0.8 kWh | ~18 kWh |
| The same, in phone charges | ~35 | ~830 |
| Water, including power generation | ~4 L | ~86 L |
How we got these numbers: the left column assumes one session uses a fifth of an NVIDIA A10 (150 W), from Hedra’s published Livatar-1 benchmark (141 frames a second on one A10). It is a share of the GPU’s frame output, not a measured utilisation. The right column is an upper bound: one session per GPU at H100 power (700 W). Energy and water are calculated from rated GPU power. Both columns use the industry-average data-centre efficiency (PUE 1.56, Uptime Institute 2024); large GPU clouds report 1.1–1.2, which would cut the electricity figures by up to 30%. Water uses LBNL’s US-average factors (0.36 L per kWh on site, 4.52 L per kWh through electricity); about 95% of the water figure is water used to generate the electricity. A phone charge is 0.022 kWh, the unit Luccioni and colleagues use. These are average-load figures: a provider also keeps spare capacity for peaks, which they do not count. We count GPU power only; host CPU, memory and networking add more.
One face is a rounding error. Faces at scale are not. Run 10,000 conversation-hours a month and the faces alone hold about 2,000 GPU-hours, close to three entry-level GPUs running day and night for nothing but lips and blinks. That comes to about 470 kWh a month with the efficient model, and up to roughly 10,900 kWh with larger models. For one company that is modest: about half of an average US household’s monthly electricity. The point is the shape of the load: it runs for every second of every conversation, and every deployment adds its own.
Taking the face off the grid
MeshLite takes the face out of that load.
The avatar comes alive on the device the customer is already holding. A small audio model listens to the voice and picks the mouth shape for each moment, and the browser places the matching pre-made image on the face. It is light work for the phone or laptop’s processor. MeshLite’s servers run no inference for the avatar, and no live video stream is sent. During the conversation, the figure in the table above is zero.
Some things still need a data centre, and we would rather say so than have a reader find out. Each avatar is created once, in a one-time GPU job. Speech recognition, the voice and the language model run on servers, from providers the customer chooses, as they do in any conversational AI product. What MeshLite removes is the face’s share of the load.
The device does its part on hardware that is already switched on, while a cloud avatar holds a data-centre GPU and still makes that same device decode a live video stream. We expect moving the face to the device to save energy overall. Until we publish a measured device figure, this is analysis.
The result is headroom. Every conversation rendered on the device frees GPU time, power, cooling and grid capacity for work that only a data centre can do. On the invoice, the same headroom shows up as no per-minute rendering fee.
Next
So far we have only counted faces. The harder questions concern the bill. What does a year of live avatars cost once the meter has run through every quiet minute? What should a buyer ask an AI avatar provider before signing, and why do so few publish the answers? And why did the most ambitious attempt yet to render every frame in the cloud, a Google product running in Google’s own data centres, close after little more than three years? We take them in turn over the coming weeks.
Sources: Gartner, Data Center Electricity Consumption to Grow 26% in 2026 (June 10, 2026) · IEA, Energy and AI (2025), executive summary · LBNL, 2024 United States Data Center Energy Usage Report · Ceres, Water Behind the Watts: The Hidden Risk of Powering Data Centers (August 2026) · LBNL, Queued Up: 2024 Edition · Uptime Institute, Global Data Center Survey 2024 · Luccioni, Jernite and Strubell, Power Hungry Processing, ACM FAccT 2024 · Hedra, Livatar-1: Real-Time Talking Heads Generation with Tailored Flow Matching (2025) · NVIDIA A10 and H100 specifications.