workloads · voice

a token waits on the same physics as a frame

Bring your recognition, language and synthesis models. We run them in live sessions near the speaker and hand the heavy turns to Gemini, OpenAI or a model of your own.

talk to us about a pilot pilot places open by conversation

one conversation, drawn

Two hops at most. The light work stays on the machine near the speaker; the heavy turn goes out to the frontier model you chose and comes back into the same session.

  1. the person speaks

    the browser or the app

    Audio goes up the same live connection a game uses for input. No download, no plugin.

  2. the node hears

    a GPU near the speaker

    Your recognition model runs on a machine near the speaker, in a placed session that stays open for the whole conversation.

  3. the node thinks

    the same GPU, or one hop out

    Your language model answers the light turns on the node. A heavy turn goes to Gemini, OpenAI or a frontier model of your own, and comes back into the same session.

  4. the node speaks

    back down the same connection

    Your synthesis model renders the reply and it streams back while it still feels like one conversation.

what it is

  • your models, packed into one session and placed near the speaker
  • a stateful connection that stays open for the whole conversation
  • heavy turns routed to any frontier model you choose, and back
  • metered per session, like a game, and settled when it ends

built for

  • companions and characters that must stay in persona, on models you own
  • tutors with their own speech and pronunciation models
  • products that need sovereignty: models and audio inside the jurisdiction you name
  • games and worlds with on-brand voices for their characters

Consumer nodes run best-effort, without a service-level agreement. Regulated and safety-critical calls belong on operator-grade capacity, which we scope with you.

per session, like a game

A voice session is metered like a game session: up to an hour, settled when it ends. Tell us which models you run and we quote it.

from $0.04

a voice session, per session, up to one hour, to $0.24 in the top compute class

how every session here is priced

bring the conversation

Tell us which models you run, where your speakers are and how many conversations you carry. We come back with where the network can place them today and what a pilot looks like.

talk to us about a pilot

who is already asking

  • voice teams that own their models

    companions, tutors and character voices are the first through this door