.Wave

.wave builds continuous inference infrastructure for real time AI models, meaning models that keep state and process live input throughout a session rather than answering one bounded request at a time. Its engine schedules recurring steps from many live sessions together on the GPU so that each session meets its own timing contract, and it offers this through the hosted .wave Live API or as a managed deployment of a customer's own model. It is aimed at developers and companies building full duplex voice agents, streaming speech recognition and other models that must keep pace with live audio or data. The first hosted model on the Live API, in public beta, is NVIDIA's Nemotron VoiceChat 11B for full duplex voice conversation, and a streaming speech recognition model is listed as coming soon. Teams with their own continuous model can bring it for a managed deployment after talking to an engineer. The company publishes technical articles on duplex inference and on WPK, its persistent GPU kernel, and case studies reporting 56 times more voice chat sessions per H100 GPU and zero missed frames for streaming speech recognition, measured on its own benchmarks. The Live API is priced per session minute. Nemotron VoiceChat is listed at an introductory $0.005 per minute, billed in 80 millisecond increments while a session runs, and new accounts receive $10 of credit at sign up with no card required. Managed deployments are priced through sales. .wave is backed by Y Combinator as part of its Winter 2025 batch.