ElevenLabs adds NVIDIA's Nemotron: voice agents 70% cheaper than other open models

A voice agent for a call center runs into three problems: it has to respond quickly, calculate without errors, and not send other people's data who knows where. ElevenLabs and NVIDIA have announced an expanded partnership that targets all three at once. This is a press release, and the figures in it are stated by the companies themselves. But for those building voice agents, there is plenty to unpack.

Nemotron inside ElevenAgents

The main news for developers: NVIDIA Nemotron 3.5 Super VL is now on the list of models available on the ElevenAgents platform. Nemotron is a family of open models designed for agentic scenarios: reasoning, calling tools, and carrying out multi-step tasks.

In ElevenAgents' internal benchmark on realistic customer-support tasks, the model showed the best end-to-end task completion accuracy. Financial calculations and invoice analysis were singled out separately. Those are exactly the areas where a voice bot usually confuses amounts and deadlines. According to ElevenLabs, it is also 70% cheaper than other leading open models.

There is a practical detail as well: Nemotron's reasoning mode can be turned off. For simple queries like "what time does the clinic open," a long chain of thought is unnecessary and only adds latency. Thanks to this, the model ranks among the three fastest on the platform. In voice, this is critical: a two-second pause feels like a frozen call.

Another plus: Nemotron is easy to fine-tune on corporate data. A company can build its own version of the model and run agents on it in ElevenAgents. This is a way to make sure a model trained on your terminology and scenarios remains your asset.

Hardware: another 15 MW

ElevenLabs is expanding its own on-premises infrastructure. On top of existing Blackwell Ultra systems, 15 MW of capacity on the Vera Rubin platform will be added. Part of it is needed for the largest customers with strict security requirements and in-country data storage, and part is allocated to research teams.

For a product manager, the takeaway is simple: "a cloud somewhere in the US" is no longer the only option. If you sell voice agents to banks, government bodies, or healthcare, the ability to deploy the solution locally becomes a selling point in a deal.

Poland: an agent for booking doctor appointments

The first sovereign case is Centralna e-Rejestracja, the central appointment booking system in Poland's national e-Health platform. The rollout is being carried out with partner EasyCall.

The scenario is straightforward. The agent calls patients who only have a landline, reminds them of an upcoming visit, and helps reschedule it. The scale: about 40 million appointments a year, with a no-show rate of 10–20%. By rough estimate, that is 4–8 million missed visits a year. Even a small reduction has a tangible effect.

The system runs on-premises on NVIDIA Confidential Computing. Citizens' data and audio stay within government infrastructure and are protected even during processing. Local deployment also reduces latency, which matters for a live conversation.

Robots remain at the planning stage

Voice for physical AI is a separate item. ElevenLabs and NVIDIA are exploring running voice models on NVIDIA edge devices: controlling industrial equipment, robots on the factory floor, talking to a car. The wording in the release is cautious: "potential applications," "exploring." There are no specific products or timelines, so it is too early to count this as an announcement.

What to do with this

• If your voice agent runs on an expensive closed model, compare cost and accuracy against Nemotron on your own scenarios. The ElevenLabs benchmark is internal and is no substitute for your own test.
• Split requests into simple and complex ones: enable reasoning only where it is needed, and save on latency.
• If you work with the public sector or healthcare, study the Polish case as an architecture template: local deployment, confidential computing, data never leaving the perimeter.

The limitation is obvious: all the figures are published by the partnership participants themselves, and the release contains no independent measurements. But the direction is clear. Voice agents are becoming infrastructure, which means questions of cost, latency, and control over data matter no less than voice quality.