Fish Audio Is Taking the DeepSeek Route

On July 28, Fish Audio, an AI voice company, announced a $52 million seed round co-led by Coreline Ventures and Today Capital (今日资本). A seed round of that size is bigger than many companies' Series A. The company is 12 months old. ARR went from zero to $21 million, with 8 million users and a team of 22 people. There is one reference point. ElevenLabs closed a $500 million round in February at an $11 billion valuation, with ARR above $330 million, more than ten times Fish Audio's size. But watching how Fish Audio fights recalls another company. Open source, low prices, and a different architectural route: the last time that combination appeared was DeepSeek. Open source does the job of distribution here Fish Audio began as an open-source project. Co-founder and chief scientist Shijia Liao, born after 2000 and formerly a researcher at NVIDIA working on video, was dissatisfied with the expressiveness of synthetic speech on the market. He trained a model on a gaming GPU in his bedroom and put it on GitHub, mostly on VTuber content. The project is called Fish Speech and now has over 31,000 stars. Here open source is distribution. Fish Audio uses its own "Fish Audio Research License": the models can be downloaded and tested freely, and commercial use is paid. A developer who has it running on their own machine and has a sense of its performance meets much less resistance at procurement. The platform now has 2 million community-uploaded voice models and more than a million monthly active users, and not a cent was spent buying those users. ElevenLabs is fully closed source and offers no self-hosting. A price gap of four to eight times Fish Audio's API costs about $0.015 per 1,000 characters, against $0.06 to $0.12 for ElevenLabs. The subscription gap is just as clear. Fish Audio Plus costs $11 a month for about 200 minutes, and ElevenLabs Creator costs $22 for 100 minutes, a fourfold difference in output for the same budget. The real separation comes from self-hosting. Fish Audio's open-source models are under Apache 2.0, which allows commercial use, and once an enterprise generates more than 3 to 4 million characters a month, running its own is cheaper than calling the API. In high-volume scenarios such as call centers, batch dubbing, and live customer service, the gap keeps widening. That road is closed at ElevenLabs. Once the price gap exceeds four times, procurement logic changes. Customers who followed "good enough, don't switch" will run the numbers again, which is what DeepSeek did back then.