Reserve inference capacity, rather than whole GPUs. Match your reservation to your workload and the hours you need—with guaranteed latency, an SLA, and a price agreed upfront.
The problem
Shared inference
Response times on shared capacity may vary with demand.
Illustrative response timesFixed GPU reservations
A fixed reservation can leave capacity unused during quiet hours and fall short when traffic peaks.
Illustrative traffic and fixed capacityReserve the concurrent calls or requests per second you need, for the hours you need them. Keep your latency guarantee without paying for a peak-sized GPU reservation all day.
Dedicated Lanes reserves less capacity during quiet hours, reducing the cost of your reservation.
Deployment regions
Deploy Dedicated Lanes close to your users to reduce network latency.
Let’s find your lane
Bring your model, traffic pattern, latency target and user regions.
Serverless models / Powered by WPK
Connect to our hosted models for live speech and conversation. Our WPK inference engine powers the stream. No GPU reservation required.
Turn live speech into text, or build a voice conversation that can listen while speaking.
Introductory price. Billed while the session runs.
56×
Read the case study98%
Read the case study0 missed frames
Read the case studyBenchmarks measure GPU serving efficiency. The cost comparison assumes the same GPU-hour price and filled capacity. ASR: 0 missed frames out of 7.8M per run at 4,800 concurrent streams, across three runs.
WPK keeps the model’s ongoing work running on the GPU, reducing repeated scheduling overhead as new audio arrives.
| 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 |
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
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 articleStart transcribing today, or request access to VoiceChat.