Nvidia’s Nemotron joins ElevenAgents, and Poland’s e-Health system is first

Nvidia is widening its work with ElevenLabs: Nvidia’s Nemotron 3.5 Super VL model is now selectable in ElevenAgents, the voice-agent platform. ElevenLabs also says it will add 15 MW of Nvidia Vera Rubin capacity for customers with strict security rules.

In a post on X, Nvidia wrote: “We’re expanding our collaboration with @elevenlabs to bring voice agents to enterprises and governments worldwide.” The post is cut off in the version available to us. The details below come from ElevenLabs’ own blog post on the partnership, published around ElevenLabs’ New York summit.

Nemotron 3.5 Super VL arrives as a model choice

According to ElevenLabs, Nemotron 3.5 Super VL is now available inside ElevenAgents. Nvidia designs its Nemotron open models for multi-step agent tasks that call tools, and they can be post-trained on a company’s own data. That gives organizations a choice of model beyond the ones already offered on the platform.

ElevenLabs reports that on its own customer-support benchmark, the model had the highest end-to-end task-completion accuracy, at about 70% lower cost than other leading open models. With reasoning switched off, ElevenLabs says it ranks among the three fastest models on the platform. Neither figure has been independently verified, and the blog post does not name the models it compared against or publish the underlying scores.

Pricing and regional availability are not given in either source. ElevenLabs does not say whether Nemotron costs extra or is included in existing ElevenAgents plans.

A national appointment system as the first showcase

The headline deployment is in Poland. ElevenLabs says models from the company, running on Nvidia infrastructure, power a voice agent for Centralna e-Rejestracja, the central appointment system in the country’s national e-Health platform, with EasyCall supporting delivery. The agent phones patients who only have landlines, sends appointment reminders and helps them reschedule.

The blog post puts the system at 40 million appointments a year and cites no-show rates of 10 to 20 percent. Those are the context for the deployment, not results: ElevenLabs does not report how many calls the agent has handled or how much it has reduced missed appointments.

The compute plan is the other concrete piece. The 15 MW of Vera Rubin capacity will sit on top of ElevenLabs’ existing Blackwell Ultra systems and is meant for large deployments with security or data-residency requirements. ElevenLabs and Nvidia also say they are exploring on-device voice AI for robots, industrial machinery and vehicles, but the post describes that as potential future work.

Confidential computing is the real selling point

The angle that matters most for this story is data protection. The Polish deployment runs on-premises using Nvidia Confidential Computing, which, according to ElevenLabs, keeps citizen data and audio inside government infrastructure while it is processed. Voice is among the more sensitive inputs a government can handle: recordings carry identity and, in a health setting, information about medical appointments.

ElevenLabs chief executive Mati Staniszewski said that as voice agents become critical infrastructure, businesses must be able to operate securely. Nvidia chief executive Jensen Huang, quoted in the same post, said “enterprises and governments can run voice agents securely at national scale.” Both statements are the companies’ own framing. The post does not mention an independent audit, certification or Polish regulator sign-off, and it does not describe how the system handles consent for recorded calls.

Hardware is already a thread in the relationship. In February, ElevenLabs announced a Google Cloud deal built around G4 virtual machines with Nvidia RTX PRO 6000 Blackwell GPUs, according to Google Cloud’s press release. The new Vera Rubin capacity points to a shift from rented cloud instances toward dedicated systems for customers that need data to stay in a given place.

Nvidia has been pushing the same message of controlled agents elsewhere. Last week it handed its agent sandbox to the CNCF, which we covered in Nvidia hands its AI agent sandbox OpenShell to the CNCF. Together the moves suggest Nvidia wants open models and agent tooling to be the default for buyers who cannot send data to a shared public service.

What remains open is how much of this a typical customer can use today. The model is available in ElevenAgents now, but the secure on-premises route appears aimed at large government and enterprise contracts, and the Vera Rubin capacity is still a plan rather than a live service.