AWS and Cartesia: running a 90 ms voice model inside your own VPC

Amazon Web Services (AWS) published a technical how-to on 3 November 2026, co-written with staff from voice-AI company Cartesia, on running Cartesia's voice models for real-time agents. It says Cartesia's Sonic text-to-speech model produces its first audio byte in about 90 milliseconds, and that teams can host it in their own AWS account. The post is vendor-published and partner-authored.

What Cartesia and AWS claim

Cartesia builds Sonic (text-to-speech) and Ink (speech-to-text) on State Space Models (SSMs), a sequence architecture that carries a fixed-size internal state instead of caching every previous token the way transformers do. The post says this suits live audio, and cites native streaming and constant memory. It says Sonic covers over 44 languages and can clone a voice from a 10-second sample. It says Ink detects the end of a speaker's turn itself, without a separate voice-activity model, and emits partial transcripts so a language model can start answering before the caller finishes.

Two ways to run it

The managed route: subscribe to the Sonic 3.6 listing in AWS Marketplace, create a SageMaker AI model from the package, and start a real-time endpoint on a supported GPU instance. AWS says the endpoint runs inside your virtual private cloud (VPC) and inference data stays in your account. The model-package identifier is specific to each AWS Region.

The self-managed route: get Cartesia's container image and licence from Cartesia, copy it into your own Amazon ECR registry, and run it on Amazon ECS or EKS with GPUs. AWS recommends G6 (L4) or G6e (L40S) instances, one GPU per replica, a cluster placement group and a single Availability Zone for the shortest network path, and at least two replicas. It advises scaling on active stream count or GPU use rather than CPU, and warns that a larger batch size raises GPU use but adds first-byte latency.

• SageMaker route: AWS says deployment takes minutes after subscribing; cost is a software fee plus instance hours plus infrastructure.
• EC2 route: AWS says hours to days; you own provisioning, scaling, patching and monitoring, and can use Reserved or Spot capacity.

What the post leaves out

It gives no price per character or per minute, no concurrency per GPU, and no description of how the 90 ms was measured: which region, instance, text length or network path. A first-byte figure also leaves out the speech-to-text, language-model and telephony hops that make up the delay a caller hears. The post says Amazon Connect uses Cartesia's models in production, but it shares no volumes or customer outcomes.

For a team piloting a phone agent, the practical step is to run your own end-to-end timing from caller speech to first audio, on your instance type, before choosing between the two routes.