In short: Data centers worldwide are heading for about 945 TWh of electricity a year by 2030, new US power projects wait about five years to connect, communities are pushing back against data centers over water, and GPUs are short. A chatbot uses a GPU in bursts. A live avatar drawn on cloud GPUs keeps one running for every frame of the conversation, even while it only listens. On a public-benchmark estimate, the most efficient published model still holds about a fifth of a GPU per conversation. As more conversations get a face, that waste grows with every minute. MeshLite removes that waste at the source: it draws the face on the user’s own device, so the conversation holds no data-center GPU for it.
Picture a customer on a support page, complaining to an AI customer service representative 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 center 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.
Everything AI needs is running short
For most of the AI boom, the scarce things were ideas and chips. Chips are still scarce. Now the socket in the wall is scarce too.
- Gartner’s June 2026 forecast: ~565 TWh in 2026 (all data centers including AI). IEA’s central case: ~945 TWh by 2030 — more than double 2024’s 415 TWh.
- New power is slow to arrive. In the US, the median generation or storage project built in 2025 took 61 months, about five years, from asking to connect to the grid to switching on, up from 22 months in 2008.
- Water is the next fight. UN University researchers estimate that the electricity used by the world’s data centers carried a water footprint of 4.5 trillion litres in 2025, rising to 9.3 trillion by 2030, before counting on-site cooling. In August 2025 Tucson’s council unanimously refused to bring Project Blue into the city, amid concerns including water and power; the developer moved ahead outside city limits with air cooling. In 2026, supervisors in Pinal County, Arizona, voted 4–1 against a data center, with two questioning its water use.
- GPUs are short as well. 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.
A GPU with no power connection does nothing. So every hour a GPU spends on one job is an hour of electricity, cooling and grid capacity that nothing else can use. A live face drawn in the cloud is a new demand on all of these, for every minute of every conversation.
What deserves a data center
Some work truly needs more data-center capacity. Finding the protein behind a disease no one can treat. Reading a child’s genome to name a rare condition. Breeding crops that can survive a drought before the next famine. These are compelling reasons to invest scarce electricity, water and GPU time.
But not every task needs a data center. Go back to the customer with the support ticket. Over ten minutes, the AI that understands them and writes the answers works in short bursts, a second or so per answer. The face is different. It needs a GPU for the whole conversation, including the minutes when the customer is talking and the avatar is listening.
Drawing the face in the cloud made sense when a lifelike face had to be generated frame by frame, work that needed a large GPU to keep up. That is no longer the only way. The mouths and expressions can now be prepared once, ahead of time. The customer’s phone then only has to choose the right image for each moment and place it on the screen.
At scale, it adds up. For a support team with 10,000 conversation-hours a month, the faces would hold about 2,000 GPU-hours, even with the most efficient model that has a published benchmark. That is close to three GPUs running all month, on a public-benchmark estimate.
The resources behind AI are precious. We would like to keep them for the work only they can do, and let a face be drawn where it is seen.
Taking the face off the grid
MeshLite takes the face out of that load. The avatar is drawn on the device the customer is already holding, with the same quality as a server-rendered avatar, on current browsers and devices except very old hardware. No data-center GPU draws the face during the conversation, and no live video is sent.
A GPU is still used once during the creation of each avatar. However, compared with drawing faces in the cloud, this is a one-time cost, not repeated for each conversation, and it saves a significant amount of resources in production.
This also turns a variable resource demand into a fixed one. As usage grows, the upfront investment stays flat, and the data-center capacity that would have been consumed remains available for other higher-value work.
We expect this to reduce total energy use as well. Until we publish a measured device figure, that remains our analysis.
Next
So far, we have only looked at where the face is drawn. 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? We take them in turn over the coming weeks.
How we got these numbers: Hedra’s published Livatar-1 model renders 141 frames a second on one NVIDIA A10. A live face needs 25 to 30, so one A10 draws about five conversations at once, and each holds about a fifth of the GPU. 10,000 conversation-hours × a fifth is about 2,000 GPU-hours, or close to three GPUs over a 730-hour month. This is a public-benchmark estimate, not a measurement.
Sources: Gartner, Data Center Electricity Consumption to Grow 26% in 2026 (June 10, 2026) · IEA, Energy and AI (2025), executive summary · UN University (UNU-INWEH), The Environmental Cost of AI (2026) · Arizona Luminaria on Project Blue (2025) · Arizona Capitol Times on Project Blue (2025) · AZFamily on Pinal County (2026) · LBNL, Queued Up: 2026 Edition · Hedra, Livatar-1: Real-Time Talking Heads Generation with Tailored Flow Matching (2025) · NVIDIA A10 specifications · 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.