Paris is a great place to build AI—except when it isn’t.
Yes, the city buzzes with research talent, deep academic roots, and a surprisingly healthy startup scene. But in the real-time voice AI race, where every millisecond counts and ecosystems move at Lightning speed, being tethered to Europe while your closest rivals—Anthropic, Google, Meta, OpenAI—are humming inside the same ZIP codes as the silicon and capital that keep them fed? That’s not sustainable.
So Gradium did what smart startups do: it upgraded its address.
The Paris-based voice AI startup has extended its seed round to $100 million with Nvidia coming aboard as a new investor, and it’s now opening a San Francisco Bay Area office to compete for talent at close range. As Gradium puts it, this move is about “strengthening its position at the heart of the world’s leading AI ecosystem.” In plain English, that means showing up, showing off, and showing how fast your voice model can actually talk without tripping over its own latency.
And get this: this isn’t a cash grab for vanity. It’s a strategic pivot.
Gradium originally launched out of stealth in December 2025 with $70 million from a handpicked group that included FirstMark Capital, Eurazeo, DST Global Partners, Eric Schmidt, and French telecom billionaire Xavier Niel. The round was led by existing backers who doubled down because the product—ultra-low latency speech synthesis, transcription, live translation, and voice cloning—already had traction. Fast-forward to today, and Gradium’s customer list includes Renault, which quietly rolled out voice features across some of its EVs last year.
Why does that matter? Because Renault didn’t pick Gradium for flair. They picked it because the system answers in milliseconds, not seconds—exactly what matters when a driver asks a car’s assistant to change navigation and has to wait while the car stare-blanks like it’s buffering a YouTube video.
But let’s back up for a second. If you’re new to Gradium, here’s the real story—not just theheadline.
Gradium didn’t spring fully formed from a garage. It emerged from Kyutai, the French AI lab backed by Xavier Niel, and co-founded by Neil Zeghidour. Yes, that Neil Zeghidour—the researcher who cut his teeth at Google Brain, DeepMind, and Facebook AI Research before landing in Paris to co-create Mimicry (later Kyutai), the lab that ultimately birthed Gradium.
Zeghidour’s background is telling. You don’t land at DeepMind and Facebook AI Research without knowing exactly how broken most voice assistants sound in the wild. He also lived the startup-to-corp cycle, which means he saw how fast things get watered down once they enter the enterprise. Gradium was built as a counterpoint to that: small, lean, obsessed with latency—not marketing.
At the heart of Gradium’s stack is its belief that voice AI shouldn’t feel like listening to a robot who just woke up from a nap. The team engineered its models for real-time streaming, meaning the audio comes out in chunks small enough to feed back into an LLM without the AI agent running out of breath—or, worse, causing the conversation to stall. This is especially important when you consider most voice agents today still rely on “stop-and-respond” patterns, where the user speaks once, waits while the server talks to itself, then hears a canned reply. Gradium’s models aim for conversational rhythm: the kind you get when two humans talk, where overlapping speech and backchanneling feel natural.
That’s why the Nvidia check isn’t just money in the bank. It’s a signal that Gradium’s tech stack aligns with the infrastructure layer of next-gen AI. Nvidia already powers most of the world’s large language models. Now it sees Gradium not just as a customer, but as a complement to its larger AI narrative: if you want real-time agents, you need ultra-low latency. And if you want that, it helps to have your voice model team seated where the hardware and platform people already live.
Which brings us to the Bay Area office.
Opening an office in San Francisco isn’t just a rent bill—it’s talent triage. Google, Meta, and OpenAI poach researchers like sharks smelling blood in the water. Gradium can’t afford to lose its voice AI engineers to a four-hour commute and a 40% raise. So it’s setting up shop in the same ecosystem it needs to defend against, not outside of.
Here’s what that actually looks like in practice:
- Recruit engineers who’ve built speech models at Microsoft Research or Apple’s Natural Language group but still want to work on startups
- Close deals with enterprise buyers (like Renault) who want faster voice response times than existing vendors offer
- Learn from Anthropic’s safety practices while dodging their hiring tactics
- Build integrations with tools that already run on Nvidia GPUs
- Test real-world constraints: noise cancellation in cars, mobile network drops in dense cities, multi-speaker overlap detection
All of that requires physical proximity—not just to venture capital, but to engineers who already speak the language.
The competition? Oh, it’s fierce. ElevenLabs, valued at $11 billion in February 2026 after its own round, dominates the creator side of voice cloning and prosody control. Google’s Gemini platform bundles its TTS/STT stack into every device, making it hard for rivals to win on raw capability alone. Then there’s Amazon’s Lex and Microsoft’s Azure Speech Services, both locked into enterprise workflows.
But here’s the nuance Gradium is banking on: no one else has combined ultra-low latency and open architecture. Most commercial voice APIs are black boxes—great for simple interactions, terrible for building custom agents or chaining multistep voice flows.
That’s where Gradbot comes in.
Gradbot, Gradium’s open-source voice agent framework, lets developers build agents with natural turn-taking, silence detection, and live web retrieval. Earlier this month, Gradium announced an integration with Keenable search to bring real-time facts into voice conversations without breaking the flow. That’s huge: today, most agents either answer from memory (risking outdated facts) or cut off mid-sentence to call an API.
With Gradbot + Keenable, agents can ask for context mid-conversation (“Let me double-check that…”), retrieve live data from the web, and then resume—not unlike how a human pauses to Google something mid-discussion. It turns voice agents from punchcards into conversation partners.
And here’s the kicker: Gradbot is designed to run on-device where possible, meaning lower latency and better privacy—another nod to the Renault use case where data stays in the car unless the driver asks for cloud help.
So what’s next?
Gradium isn’t announcing a product launch this time—it’s raising funding because it already has one. Instead, the company is investing in distribution: customer success teams for enterprise deployments, a go-to-market team focused on automotive and contact center verticals, and partnerships with hardware vendors who want their silicon optimized for streaming audio inference.
Nvidia’s involvement isn’t just about writing a check. Expect joint demos at GTC 2027, co-marketing with Gradium’s open-source framework, and likely early access to new SoCs designed for generative audio.
Gradium’s $100 million seed isn’t a raise—it’s an on-ramp. A signal that voice AI is entering the high-stakes, Silicon Valley track where latency, personality, and proximity to talent determine who builds the next generation of assistants—and who gets left saying “Sorry, I didn’t catch that” into a silent void.
Let’s see who blinks first.