Real-time voice AI — the Sonic model family built for the speed of conversation
HQ: San Francisco, CA, United States | Founded: 2023 | Employees: ~100 | Stage: Reported $500M round at ~$11B valuation (late 2025); ~$590M+ total raised | Website: https://cartesia.ai
Cartesia builds real-time voice AI. Its Sonic model family generates speech fast enough for natural conversation — streaming audio with sub-second latency — and its models are embedded across the voice-agent ecosystem: platforms like Vapi and Retell list Cartesia as a selectable voice component, and enterprises use its API for agents, dubbing, and interactive media. The company also ships Ink, an on-device voice model small enough to run offline, targeting privacy-sensitive and latency-critical uses.
The company was founded in 2023 by researchers from Berkeley and Stanford, including Albert Gu — co-author of the state-space-model research line (Mamba) that rethought sequence modeling for efficiency. That architecture DNA is Cartesia's technical thesis: transformers struggle with the latency budgets of real-time speech, while SSM-style streaming generation is built for it. Kernel Folding, its inference technique, made low-latency speech generation practical at scale.
Funding tracked the voice boom: a $22M seed in early 2024, a $64M Series A in December 2024 led by Index Ventures (with Lightspeed, Felicis, Conviction and others), and — per category trackers — a reported $500M round at approximately $11B valuation in late 2025, placing Cartesia among the most valuable private AI infrastructure companies in the world. The team sits around 100 people, research-heavy by design.
Cartesia is the research-lab-turned-infrastructure play of the voice boom. Its founders — Berkeley/Stanford researchers who helped pioneer state-space models (the Mamba lineage) — built a speech model with a structural edge: it generates speech in a streaming fashion that's fast enough for real conversation. The result, Sonic, became one of the default voice components inside the agent platforms powering the 2024-2026 voice-AI wave, alongside rivals ElevenLabs and OpenAI's native audio.
The funding curve is the purest expression of the category's heat: $22M seed, then $64M Series A within months (Dec 2024), then — per category trackers — a $500M round at roughly $11B valuation in late 2025, among the steepest valuation climbs of the AI cycle. The bet: as every product gains a voice interface, the voice model itself becomes a commodity input, and the winner is whoever is fastest, most natural, and cheapest at scale — plus the only player shipping on-device voice for offline and privacy-sensitive use cases.
The risks are textbook model-layer: OpenAI, Google and ElevenLabs fight for the same workloads, open-weight models close the quality gap, and an $11B paper valuation demands flawless growth. But Cartesia's architecture-first approach (SSMs over transformers for streaming latency), its on-device differentiation, and its position as the speed benchmark give it the strongest technical identity in a noisy category.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki