Owning the Real-Time Voice Stack

Control of the whole voice stack is becoming the moat, because the hardest part of voice AI is not making a nice voice, it is making a live call feel natural under production constraints. In practice that means hearing noisy phone audio fast, deciding when the caller is done, generating a reply with almost no dead air, routing the call, logging it, and keeping quality stable in deployment. Cartesia is competing to own that entire loop with Sonic, Ink, and Line, rather than sell a single voice endpoint.

• Full stack rivals are moving the same direction. ElevenLabs now sells bundled agents and cut conversational pricing to 10 cents per minute in February 2026, while Deepgram has expanded from transcription into Aura-2 TTS, Flux turn taking, and a Voice Agent API. The market leader is no longer the best standalone TTS voice, it is the vendor that can remove the most handoffs.
  1 sacra 2 sacra 3 elevenlabs 6 deepgram 7 deepgram
• The reason handoffs matter is latency and turn management. Cartesia positions Sonic at under 90ms time to first audio, while Deepgram says Flux reduces response latency by 200 to 600ms versus pipeline approaches. Those gains come from coordinating transcription, end of turn detection, model reasoning, and speech generation as one runtime, not as separate vendors stitched together.
  1 sacra 3 elevenlabs 7 deepgram 8 deepgram
• This also changes switching costs and channel dynamics. A buyer using Cartesia only for TTS can swap providers on price. A buyer using Line for telephony, deployment, logging, and rollbacks is replacing an operating layer. At the same time, partners like Vapi and Retell can route traffic across vendors, so Cartesia has to climb the stack without losing the platforms that feed it demand.
  1 sacra 4 sacra

The next phase of voice AI infrastructure will look more like contact center software than model APIs. The winning platforms will bundle speech, turn taking, orchestration, observability, and enterprise deployment into one system, while native speech to speech models from OpenAI push the market toward even tighter integration and fewer separate components.