Nvidia just wrote a check to a seven-month-old Paris voice AI startup, and the size of it tells you the chipmaker isn't content to just sell the picks and shovels anymore.
Gradium, the Paris-based startup building AI models for real-time text-to-speech and speech-to-text, has added roughly $30 million to its seed round, according to a report from Sifted, pushing the total past $100 million. Nvidia led the new money. It's a striking vote of confidence for a company that only launched in September 2025.
The extension follows a $70 million first tranche that closed at the start of December 2025, led by FirstMark Capital and the French investor Eurazeo, as TechCrunch reported at the time. Xavier Niel, Rodolphe Saadé and former Google chief executive Eric Schmidt all put money in too. Gradium hasn't disclosed which other investors joined Nvidia in the latest tranche, and it isn't sharing a valuation.
You don't raise $100 million on a seed round for a company that's barely old enough to have a second birthday without a story behind the founder. Neil Zeghidour built Gradium after cutting his teeth as a voice researcher at Google DeepMind, and the company itself spun out of Kyutai, the nonprofit Paris AI lab that launched in 2023 with 300 million euros from the same trio of Niel, Saadé and Schmidt. That lineage is doing real work here. Investors are backing a team that already shipped open research inside Kyutai before going commercial.
What Gradium actually sells is speed. Its models handle real-time text-to-speech, speech-to-text and voice translation, tuned to run with ultra-low latency, including on edge devices like laptops and phones. At launch the models covered English, French, German, Spanish and Portuguese, with more languages promised. The company has also released an open-source framework for building voice agents, the kind of infrastructure move that makes it easier for outside developers to build on top of Gradium rather than a rival's stack.
Nvidia doesn't need to invest in voice AI startups to sell them GPUs. Every model Gradium trains and serves already runs on Nvidia silicon, seed round or not. So the check is about something else, extending Nvidia's reach into the application layer, where the company gets a front row view of how developers are actually building with voice, and a stake in the outcome if one of these startups becomes the default voice layer for AI agents. It's the same instinct that's driven Nvidia's venture arm into dozens of AI application companies over the past two years, and it fits a pattern where the compute seller increasingly wants equity in the layer sitting on top of its chips, not just a purchase order.
The competitive backdrop explains the urgency. ElevenLabs, the London-based voice AI company, is reportedly targeting a valuation near $22 billion. Cartesia is chasing the same real-time voice market from San Francisco. And frontier labs including OpenAI, Anthropic, Meta and Mistral are all shipping their own voice capabilities directly into their flagship models, which means Gradium is fighting more than other startups now: the labs that might make standalone voice models redundant. Zeghidour told Sifted the field is more concentrated than it looks.
Also read: Helsing's Equity Reshuffle Shows How Fast Europe's Defense Boom Is Moving • Mercor Buys Deeptune After Its Own Founder Quietly Backed the Startup • Meta Opens Muse Spark 1.1 to Developers While Locking Rivals Out at Home