ElevenLabs is broadening its collaboration with Nvidia around ElevenAgents, its suite of voice agents. The expansion pairs ElevenLabs’ audio models with Nvidia’s Nemotron 3.5 Super VL model and Nvidia’s Vera Rubin architecture, with the companies positioning the combined stack for enterprise, public-sector and potentially robotic deployments.

The announcement puts the emphasis on scale rather than a single consumer-facing voice feature. ElevenLabs says new ElevenAgents capabilities include memory, routines and tools intended to help users design new agents. Nvidia’s contribution spans the language-model layer, accelerated computing and a security stack, while ElevenLabs supplies the audio and agent-facing side of the system.

What the expanded partnership covers

At its core, the arrangement combines two distinct pieces of an AI-agent system. ElevenLabs is known here for audio models and voice-agent technology: the components that can process spoken interaction and produce a spoken response. Nvidia is providing Nemotron 3.5 Super VL, described as an open model intended to support an agent’s understanding of workflows and human reasoning, alongside the computing capacity needed to run agents at larger volume.

An AI agent, in this context, is more than a system that answers a one-off question. It is software designed to carry out a sequence of steps toward a task. A voice agent may need to understand a spoken request, identify the relevant workflow, collect necessary details, retain appropriate context, take or recommend an action, and communicate the result aloud. The quality of speech is only one part of that chain.

For more context, explore Anyma Will Return to the Pyramids of Giza for Two ÆDEN Shows With Jean-Michel Jarre in 2027.

The newly outlined features reflect that broader role:

  • Memory refers to an agent retaining context that can make later interactions more coherent. The announcement does not specify the duration, scope or controls for that memory.
  • Routines suggest repeatable, structured sequences for recurring tasks, rather than requiring every interaction to be improvised from scratch.
  • Agent architecture tools are meant to help users construct new agents, presumably by defining how an agent handles a task and which steps it follows.

Those claims should not be read as a detailed feature specification. The companies have not set out every workflow, product interface or deployment configuration in the information available here. What is clear is that the partnership is moving beyond text-to-speech infrastructure toward systems meant to participate in business and government processes.

Why the computing layer matters

ElevenLabs says it will use Nvidia’s Vera Rubin architecture for the compute required to deploy its agents. Compute is the processing capacity needed to run AI models. For voice agents, demand can rise quickly when many people are speaking to a service at once, especially if each interaction uses multiple models or processing stages.

A typical voice interaction can involve turning speech into machine-readable text or representations, interpreting the request, generating a response, and producing natural-sounding audio. If the agent is connected to a workflow, it may also need to retrieve information or complete a task during the exchange. Each stage can add processing work and delay.

That helps explain why the announcement repeatedly frames the collaboration around large-scale deployments. A small demonstration can be useful with limited infrastructure. A service used by a major business or a public system has different pressures: high concurrent usage, reliability expectations and the need to keep interactions responsive enough that a conversation does not feel broken by long pauses.

Nvidia’s role is therefore not simply branding around a model name. The company is supplying a language-model component and the accelerated hardware infrastructure intended to make broad deployment possible. Accelerated computing generally means using specialized processing hardware, such as GPUs, to carry out the large volumes of parallel calculations AI workloads require more efficiently than conventional general-purpose processing alone.

A public-service example already in view

The companies pointed to Centralna e-Rejestracja, an appointment-making platform within Poland’s national online health system, as an example of voice-agent use. This is an important detail because scheduling is a bounded but meaningful workflow. It involves gathering information, navigating appointment availability and communicating a result clearly—precisely the kind of repetitive, multi-step interaction a voice agent is intended to handle.

It also illustrates the higher stakes of moving voice technology into public services. In an entertainment setting, an awkward response may simply be frustrating. In appointment-making for healthcare, reliability, clarity and appropriate handling of user information matter much more. The announcement says the companies want to make additional voice agents available to governments worldwide using Nvidia’s AI architecture, but it does not identify specific future government deployments.

The mention of a security stack is similarly notable. The statement presents security as part of the platform proposition, although it does not provide technical details about safeguards, data handling, access management or deployment controls. For organizations considering systems that may interact with customers, citizens or patients, those details would be central practical questions rather than optional extras.

What “critical infrastructure” means in this pitch

ElevenLabs CEO Mati Staniszewski characterized voice agents as becoming “critical infrastructure.” That is a strategic description, not a technical designation laid out in the announcement. The phrase signals that ElevenLabs sees voice interfaces shifting from a novel add-on to a foundational way organizations deliver services and complete routine work.

There is a practical distinction worth keeping in mind. A voice agent can be critical to an organization’s day-to-day operations without being suitable for every type of decision. The supplied information focuses on scale, workflows and infrastructure; it does not claim that ElevenAgents should independently make sensitive judgments. For users and institutions, the value of an agent will depend heavily on the specific task it is assigned, the design of its routine, and the oversight around its operation.

From a 2022 hardware relationship to an investor partnership

The companies’ relationship dates to 2022. ElevenLabs used Nvidia GPUs for its first text-to-speech model, while Nvidia used ElevenLabs voice technology for video voiceovers at company events. Nvidia later became an investor in ElevenLabs.

That history offers useful context for the current announcement. This is not presented as a newly formed vendor relationship. It extends an existing technical connection from the foundational text-to-speech stage into agent systems that join audio, language-model reasoning and deployment infrastructure.

ElevenLabs’ financial growth provides a second measure of the company’s ambitions. The company completed a $300 million employee tender offer in September at a $22 billion valuation. That was double the $11 billion valuation reported after a funding round in February. A tender offer involving employees is not the same thing as a new product capability, but it does indicate the scale of the company being positioned behind the expanded infrastructure effort.

Robots are part of the longer-term ambition

ElevenLabs and Nvidia also say they are seeking to bring voice agents to robots. The available information does not identify a specific robot, product timetable or capability set, so this should be treated as an area of intended development rather than a detailed launch plan.

Still, the direction fits the structure of the partnership. In a robot, voice can act as the conversational layer through which a person gives instructions or receives updates. The agent layer can interpret the request and connect it to a workflow, while underlying computing supports the models involved. A robot deployment would introduce additional complexity beyond a voice-only service, but the announcement does not provide enough detail to establish how that work will be implemented.

What to watch next

The most meaningful follow-up will be evidence of how these capabilities operate in real workflows. The headline promises of memory, routines and large-scale infrastructure describe the shape of the platform. The unanswered questions concern deployment: what kinds of organizations will build agents, which tasks will be prioritized, how administrators will govern them, and what specific protections accompany public-facing use.

For now, the announcement establishes a clear division of labor. ElevenLabs is advancing a voice-agent offering with features aimed at persistent and repeatable interactions. Nvidia is contributing language-model technology and the processing infrastructure intended to run those systems at substantial volume. Their shared goal is to make speaking to a service a faster route through a workflow—whether that workflow belongs to an enterprise, a government service or, eventually, a robot.