Energy and environment
Phone agents designed to use less energy.
Most of every call runs on small AI models we train ourselves. A large model steps in only for the open-ended parts. Here is where the electricity goes, what the research says, and what we are measuring next.
Book a call
An engineering description, not a pledge.
Most of a phone call is handled by small models we train and host ourselves. The open-ended turns run on a large language model. Where a number on this page is someone else’s, we name them. Where it is ours, we will publish how we measured it.
Self-hosted speech recognition
Small models on our GPUs
Scripted turns skip the LLM
Cached speech
Reviewed September 2026
Where a call’s compute goes
Four jobs on every turn. Very different footprints.
A voice agent listens, decides, writes and speaks, dozens of times per call. Model size sets the energy of each job, and even the writing is done by fast, sparse models built for real-time streaming.
LISTEN
0.6–1.7B
parameters, speech recognition
Live transcription runs on speech recognition models we host ourselves, alongside a self-hosted commercial engine.
Runs on our GPUs
DECIDE
<1B
parameters, what happens next
The conversation flow and ultra-small, fast models we trained decide the next step: intent matching, yes-or-no and option detection. Scripted lines, including the greeting and the opening, are chosen by those models and spoken as written, with no large language model in the loop.
Runs on our GPUs
WRITE
~5B
parameters used per word, open-ended turns
Free-form turns run on fast models. The default one uses about 5 of its 120 billion parameters for each word it produces, and smaller models handle the side tasks. Repeated prompt text is saved at the provider rather than recalculated. Part of this runs at third-party providers.
Streaming language model
SPEAK
Cached
text-to-speech, third-party
Speech is synthesized by a text-to-speech provider. Scripted lines are pre-generated when a call starts, and identical utterances are served from a voice cache.
Third-party provider
Also on our GPUs
Entailment checks, embeddings, query rewriting, dialogue segmentation and voice activity detection. Small models we host ourselves, most under one billion parameters.
Calls are short
51% of cold calls end inside the first minute
In Klenty’s study of 25,000 cold calls, 12,845 ended within 60 seconds. Gong puts a successful cold call at about 5 minutes 50 seconds and an unsuccessful one at roughly half that. The start of a call, the greeting, the identification and the opening, is exactly the part that runs on scripted turns and small models. On a typical sales call, a large share of the turns never reaches the large model.
Klenty, How long should a cold call last, 25,000-call study. Gong, cold call statistics. Industry data, not ours. We will publish our own share of turns, see below.
Design choices
How we spend less of it.
01
Scripted turns run on ultra-small models
The greeting, the opening and every scripted line are chosen by ultra-small, fast models and spoken as written. The large language model is called only for open-ended turns.
02
Small models around the conversation
Intent, entailment, embeddings, segmentation and voice activity detection run on self-hosted models, most under one billion parameters, instead of a frontier model.
03
Speech recognition under two billion parameters
Live transcription runs on speech recognition models we host on our own GPUs, sized between 0.6 and 1.7 billion parameters.
04
Speech that is cached, not re-synthesized
Identical agent utterances are served from a voice cache instead of being synthesized again by the text-to-speech provider.
05
Optimized inference, not just smaller weights
Self-hosted models run in half precision with ONNX Runtime paths. Model instances are pooled across calls, not reloaded for each one.
06
Single-GPU instances that follow call volume
Inference runs on single-GPU cloud instances that scale with live calls. Dispatch is capacity-aware, so we never dial more calls than we can serve.
07
Bounded conversations
Every call has a hard maximum duration and a silence timeout. A stalled or runaway session cannot keep burning compute.
Hosting
On AWS, in regions AWS reports as renewable-matched.
Our regions, US East and Frankfurt, are on AWS’s 2025 list of regions whose electricity was matched 100 percent with renewable sources. AWS reports a global power usage effectiveness of 1.14.
AWS’s figures, on an annual-matching basis.
What the research says
Smaller models use far less energy. The evidence is public.
~100×
Size drives energy per token
Models in the 70-billion-parameter class consume on the order of 100 times more energy per token than small models.
The Price of Prompting, arXiv 2407.16893
70×
Spread on one short prompt
Across 30 commercial models, a short prompt ranged from about 0.1 Wh on a small model to about 7 Wh on a reasoning model.
How Hungry is AI?, arXiv 2505.09598. Third-party estimates.
90%
Task-specific beats general
Small models tailored to a task can cut energy use by up to 90 percent compared with one large general-purpose model.
UNESCO and UCL, July 2025
0.24 Wh
A median prompt, disclosed
Google reports a median Gemini text prompt at 0.24 Wh and 0.03 gCO2e. Mistral’s lifecycle assessment puts a 400-token response at 1.14 gCO2e.
Google, Aug 2025 (self-reported). Mistral with ADEME, July 2025.
None of these figures are ours. We cite them for direction, and to explain why model size is the lever we pull first.
What we are working toward
Numbers before adjectives.
Things we can measure with tools we already run, and one direction we have chosen. This page changes when each has a number behind it.
01
Energy per call minute
Measure GPU energy on our own inference per minute of conversation, and publish the method alongside the figure.
02
Hosting footprint by region
Report the emissions our cloud provider attributes to our workloads, per region, using its customer carbon reporting.
03
Share of turns without the LLM
Our analytics already tag every scripted line as spoken without a language model call. We will publish that share of turns, and how it moves.
04
More of the dialogue on models we run
Move dialogue inference toward fine-tuned open models served on our own GPUs. Chosen, not done. We will say when it is.
Ask us how a call is served.
We will walk your team through the models, the hosting, and what we measure.
Book a call