VoiceChat sessions per GPU
56×
Read the case study
Backed by YC W25 (Y Combinator, Winter 2025)Public beta
Continuous inference
Continuous inference
for real-time AI
Run stateful models that process live input throughout a session. Use the .wave Live API for hosted models, or bring your model to a managed deployment.
Capacity measured on the clock.
Lower GPU cost contribution
98%
Read the case studyASR streams kept on schedule
0 missed frames
Read the case studyMore live sessions.
Kept on time.
Compatible session work advances together, creating room for more live sessions within their timing contracts.
Swipe to compare
| What matters | Request serving | .wave engine | Benefit |
|---|---|---|---|
| Unit of work | Bounded requests | Recurring steps from live sessions | Share GPU capacity across live sessions |
| Timing | Response and token latency targets | A timing contract for each session | Meet recurring deadlines |
| Scaling | Requests or tokens per second at target latency | Live sessions sustained on schedule | Grow capacity without falling behind |
Models that must keep pace with the world.
Full-duplex voice and streaming ASR are different model classes, but both keep state and process live input throughout a session.
Articles
Continuous inference needs a different GPU engine
Engine
Why request-serving stacks fail continuous inference, and how WPK, our persistent GPU kernel, keeps recurring work on the GPU to serve 56× more real-time sessions on the same H100.
Read the article
Full-duplex inference runs on a clock
Systems
AI inference is moving beyond requests. New input arrives incrementally, model state persists, and each output is useful only while it is still current.
Read the articleReady to build
with .wave?
Start with Nemotron VoiceChat or streaming transcription on the .wave Live API, or bring your continuous model for a managed deployment.
