TL;DR: Gradium, a Paris-based voice AI startup spun out of Kyutai, extended its seed round to $100M with NVIDIA joining as a new investor. The company builds ultra-low-latency audio models for AI voice agents — the kind that respond instantly, without the awkward pause that plagues most AI conversations. This is one of Europe’s largest seed rounds and puts French AI voice research on the global map.
Introduction
On July 9, 2026, Gradium announced it had extended its seed round to $100 million, with NVIDIA joining FirstMark Capital, Eurazeo, DST Global Partners, Eric Schmidt, and Xavier Niel as investors. The round — one of the largest seed financings in European history — signals that the voice AI infrastructure layer is becoming strategic enough to attract hardware giants like NVIDIA.
Gradium originally launched out of stealth in December 2025 with $70 million. The company spun out of French AI research lab Kyutai, itself backed by Xavier Niel with a €300M budget and a commitment to open-source research. Both Kyutai and Gradium were co-founded by Neil Zeghidour, a researcher who previously worked at Google Brain, DeepMind, and Facebook.
The company is using the new capital to open a Bay Area office, competing directly for talent in the heart of the world’s AI ecosystem — a notable move for a Parisian startup that signals it intends to play at the global frontier.
(Source: TechCrunch — Paris-based AI voice startup Gradium raises $100M seed, backed by Nvidia)
What Gradium Builds
Gradium develops audio AI models purpose-built for voice at scale. The core differentiator is ultra-low latency — AI voices that respond almost instantly, eliminating the awkward half-second pause that makes current AI agent conversations feel robotic and unnatural.
The technology stack builds on Kyutai’s research foundation, which produced some of the most advanced open-source speech models in the field. Gradium commercializes this research for enterprise applications: customer service voice agents, real-time translation, voice-based AI assistants, and interactive voice response systems that need to sound human.
Unlike text-to-speech systems that bolt voice onto a text pipeline, Gradium’s models are natively audio-native — they process and generate speech directly, without the text intermediary that adds latency and strips prosody.
(Source: PYMNTS — Nvidia Joins $100 Million Bet on Gradium Voice AI)
Why NVIDIA Invested
NVIDIA’s participation in a seed round — even an extended one — is noteworthy. The company typically invests at later stages or through strategic partnerships, not seed extensions. Several factors explain the decision:
Voice AI Is an Inflection Point
2026 has been the year voice AI went from gimmick to infrastructure. AI agents that can hold natural phone conversations, customer support that sounds human, and voice interfaces that don’t frustrate users are all dependent on the same bottleneck: latency. The company that solves the latency problem owns the voice AI infrastructure layer.
Hardware-Software Symbiosis
Gradium’s ultra-low-latency audio models are likely optimized for NVIDIA’s hardware. Running inference on NVIDIA GPUs means Gradium’s success drives GPU demand — a virtuous cycle that benefits both companies. NVIDIA’s investment may also include technical collaboration to optimize Gradium’s models for their latest silicon.
European AI Talent Access
Paris has emerged as a genuine AI research hub, producing world-class talent from labs like Kyutai, Mistral, and H Company. By backing Gradium, NVIDIA gains a foothold in this ecosystem and early access to French AI research talent.
(Source: Tech Funding News — NVIDIA backs the French researchers who built voice AI’s next generation)
The Voice AI Landscape in 2026
Gradium enters a competitive but growing market:
| Player | Focus | Stage |
|---|---|---|
| Gradium | Ultra-low latency audio-native models | Seed ($100M) |
| ElevenLabs | Text-to-speech, voice cloning | Series C ($80M) |
| OpenAI (Advanced Voice) | ChatGPT voice mode | Integrated |
| Google (SoundStorm) | Audio generation models | Research |
| PlayAI | Voice agents for customer support | Series A |
The voice AI market is bifurcating into two approaches: text-pipeline systems (ElevenLabs, PlayAI) that convert text→speech, and audio-native systems (Gradium) that generate speech directly. The audio-native approach promises lower latency and more natural prosody but is harder to build and requires more specialized research talent — exactly the moat Gradium is constructing.
What $100M at Seed Means for European AI
Gradium’s round is part of a broader pattern in European AI funding:
- Mistral AI raised over $1B across multiple rounds
- H Company raised $220M seed (2024)
- Kyutai launched with €300M non-profit budget
- Gradium now at $100M seed
Europe is producing world-class AI research but has historically struggled to retain it. Gradium’s decision to open a Bay Area office while keeping its Paris HQ reflects a hybrid strategy: keep the research anchored in Europe where the talent is educated, but place commercial operations close to customers. This may become the dominant model for European AI startups.
(Source: TechCrunch)
FAQ
Q: How does Gradium’s tech differ from ElevenLabs? A: ElevenLabs uses a text-to-speech pipeline (text → speech). Gradium builds audio-native models that process and generate speech directly, cutting latency and preserving natural prosody.
Q: Why does NVIDIA care about voice AI? A: Voice AI inference runs on GPUs. If Gradium becomes the default voice layer for AI agents, it drives significant GPU demand. NVIDIA’s investment is both financial and strategic — optimizing Gradium’s models for NVIDIA hardware creates a moat against competitors running on AMD or custom silicon.
Q: Is this open source like Kyutai? A: Gradium is a commercial entity, not a research lab. While it builds on Kyutai’s open-source foundation, the production models are proprietary. The company has not announced any open-source plans.