Gradium pulls in Nvidia as it extends its seed to over $100M for French voice AI

Gradium, a Paris-based voice AI company spun out of Kyutai in December 2025, has extended its seed round to over $100 million with a new tranche of approximately $30 million (€27.7 million) joined by Nvidia. The extension adds the chip giant to a cap table that already includes Firstmark (which led the original $70 million seed in December 2025), Eurazeo, and individual backers Xavier Niel, Rodolphe Saadé, and Eric Schmidt.

The total seed now exceeds $100 million — an unusually large number for a company less than eight months old, even by the standards of foundation-model startups.

From Kyutai to Gradium

Gradium emerged directly from Kyutai, the Paris-based AI research institute backed by Xavier Niel and Rodolphe Saadé. Its four co-founders — Neil Zeghidour (CEO), Laurent Mazaré (CCO), Olivier Teboul (CTO), and Alexandre Défossez (CSO) — led the development of Kyutai's voice AI research, including the Moshi model, before spinning Gradium out as an independent company.

The founding thesis is that voice AI is now replicating the dynamics of the large language model race from two years earlier: a small number of players capable of training foundational models at scale will determine the infrastructure the rest of the industry builds on, and everyone else will apply rather than train. Gradium is building for the training side.

The product suite spans text-to-speech, speech-to-text, translation, edge-optimised voice models, and a developer framework called GradBot. The target customers are developers building voice-enabled applications and enterprises deploying voice AI at scale — two markets that are moving quickly and signing contracts.

Why Nvidia is in the round

Nvidia's participation in a voice AI round is not purely financial. The company has been investing in AI model developers whose workloads — particularly at scale — run on its GPUs. A voice AI company training foundational models needs significant GPU compute; ensuring that Gradium remains on an Nvidia-first infrastructure path is part of the logic.

"The voice AI sector is strongly accelerating," Zeghidour said. "Thousands of startups and enterprises are using voice models to build new applications, but there are less than a dozen players capable of training these models at scale. There is therefore intense competition when it comes to the models themselves, but concentrated between a limited number of players."

The concentration argument is precisely why the capital is large. You cannot hold a position in foundational model training without sustained investment in compute and research talent. The $100M+ seed is, in part, an attempt to ensure Gradium stays inside that small group.

The San Francisco test

The new capital funds a San Francisco office — both a talent move and a market move. The US enterprise market for voice APIs is where the largest contracts are, and where ElevenLabs, OpenAI, and Google are fighting hardest.

The 18-month question is straightforward: can Gradium convert its research credibility into developer adoption and then into enterprise revenue, while larger competitors spend multiples more? Its founders built the models that proved French voice AI research is world-class. Whether that credibility compounds into a sustainable commercial position — or whether the capital advantage of US competitors overwhelms it — is what the San Francisco office is meant to find out.