With its valuation surging by 11 billion US dollars in half a year, why has the AI audio sector achieved profitability earlier than the AI video sector?

Some say AI programming is profitable, others claim AI Agents are lucrative, and many argue AI video generation brings in big money—but few realize that AI audio can also be a highly profitable business.

Foreign media reports indicate that ElevenLabs is in internal discussions for a secondary market share sale transaction, allowing employees to sell their held stocks. The deal is expected to close before September, valuing the company at approximately $22 billion.

Yet just five months prior, the company had completed a $500 million Series D funding round, at which point ElevenLabs was valued at $11 billion.

In other words, its valuation doubled in less than half a year.

The critical question is: How does an AI voice company justify a $10 billion valuation increase in just six months?

01

From AI Dubbing Tool to Voice Infrastructure

In 2022, ElevenLabs' two founders Mati Staniszewski and Piotr Dabkowski initially followed the TTS (Text to Speech) path.

Traditional TTS is essentially "concatenative synthesis": splitting real human recordings into small fragments and stitching them together according to text. While the words fit, tone, rhythm, and emotion are completely lost, resulting in a robotic reading of a script.

Through deep learning, Mati and Piotr enabled their model to understand the meaning of text, directly generating speech with emotion, rhythm, and natural pauses. The pair invested $100,000 of their own money for the first round of training, and the results were exceptional.

Although in product form it remained TTS, from that point onward, ElevenLabs already had a solid AI foundation.

Starting in 2024, ElevenLabs launched its Conversational AI platform (later renamed ElevenLabs Agents).

Its product logic works as follows: when a user speaks a sentence, the system first uses speech-to-text to extract the content, sends it to a large language model to generate a response, then converts the response back to speech, ultimately delivering the reply with rich expression.

The entire process takes roughly hundreds of milliseconds.

ElevenLabs' advantage lies in the fact that both the "listening" and "speaking" components in this process use its own proprietary models.

For "listening", it uses ElevenLabs' Speech to Text model, responsible for transcribing user speech into text. On the "speaking" side, it offers two distinct speech synthesis models with different specializations.

eleven_flash_v2_5 prioritizes speed, reducing latency to approximately 75 milliseconds, specifically designed for real-time conversation scenarios; eleven_v3 prioritizes quality, supporting over 70 languages with stronger expressive power, suitable for content creation scenarios that demand high quality and are less sensitive to latency.

After ElevenLabs Agents became its core engine, ElevenLabs expanded further downstream, launching Dubbing (multi-language dubbing) and Music (music generation) services.

That's not all. In July 2026, Netflix premiered a Wonka-themed reality show called *Wonka's The Golden Ticket*. The show's narration uses the voice of Gene Wilder, even though Wilder passed away back in 2016.

This voice was AI-reconstructed by ElevenLabs, with authorization from the Wilder Estate. Wilder's wife released a statement saying the family supports using this method to bring his voice to new generations of audiences.

Around the same time, ElevenLabs released an AI-narrated version of *The Odyssey*, with the narrator being a voice clone of Michael Caine, also fully authorized.

Celebrity voices are intellectual property, just like portrait rights—they can be licensed, monetized, and deployed at scale.

Subsequently, ElevenLabs launched the Iconic Voice Marketplace to turn celebrity voices into licensable IP, as well as Reception AI, an AI-powered virtual front desk employee.

By the end of 2025, ElevenLabs' ARR (Annual Recurring Revenue) neared $350 million. By April 2026, that figure had exceeded $500 million.

To date, over 2 million Agents have been created on the Eleven Agents platform. In the first half of 2026 alone, the platform processed more than 33 million conversations. Importantly, none of these conversations include demo data—all are real production-level calls.

Reading this, you might feel puzzled: Isn't ElevenLabs growing increasingly successful? Why would employees want to sell their stocks? Wouldn't it be better to hold onto them for further appreciation?

The situation is this: according to foreign media reports, ElevenLabs now competes with OpenAI and Anthropic for the same pool of engineers and researchers every day. These employees have worked at the company for two or three years, and the paper value of their stocks has multiplied several times, but they cannot realize this value before an IPO. Without the ability to cash out, the temptation to switch jobs grows stronger by the day.

So this transaction essentially represents the company unlocking liquidity for its core team. In fact, ElevenLabs already completed a $100 million tender offer concurrently with its Series D round, making this the second such move in less than a year.

Investors' willingness to repurchase employee stocks at a $22 billion valuation demonstrates their confidence that the company's value will eventually exceed $22 billion.

02

The AI Voice Market:

Why Voice is Monetizing Faster Than Video

Sora was shut down due to massive losses, and Tan Dai, President of Volcano Engine, once stated that all circulating revenue figures for Seedance are completely inaccurate and universally overstated.

In other words, whether domestically or internationally, AI video generation is simply not profitable right now.

So why is AI audio generation so much more lucrative?

First, it has a far leaner cost structure.

Speech generation processes a single time-series sequence, with the core task of converting text, semantics, and emotion into continuous audio; video generation processes continuous frames, requiring simultaneous consistency across characters, backgrounds, actions, camera angles, lighting, and consecutive frames.

The latter outputs far more information, with a much more complex inference pipeline.

This results in products like Sora having extremely high per-output costs. Moreover, the generated results are not guaranteed to be "usable". A single video might need repeated re-runs due to character distortion, unstable camera work, unnatural movements, or inconsistent styling.

Every retry consumes additional computing power.

Audio is different: its product form is mature, it does not require repeated generation attempts, and its per-unit cost is far lower than video.

More critically, usability standards for voice are easier to productize. Customer service voice can be integrated into enterprise workflows as long as it is sufficiently natural and low-latency. Video, by contrast, almost always requires post-editing before release.

So AI audio functions like infrastructure, while AI video remains more of an interesting "novelty".

This is the underlying logic behind ElevenLabs' rapid growth. It does not sell one-off "wow effects"—it sells persistent voice capabilities that enterprises can continuously integrate. Low cost, low latency, minimal retries, and easy integration are what allowed AI voice to become a profitable, sustainable business far earlier than AI video.

Second, its use cases are far more clearly defined.

Dubbing, audiobooks, short video narration, localized translation, customer service calls, outbound sales calls, employee training, online education, game NPCs... Off the top of my head, I can think of countless scenarios where AI audio can be applied.

AI audio does not create new demand—it essentially replaces existing dubbing methods and expands production capacity.

Video generation has no such clearly defined scenarios. What can you actually use it for? Advertising assets? Short films? Social media content?

These established scenarios have driven ElevenLabs' growth.

Third, the barrier to entry is far lower than video.

Video has countless problems to solve, including world models, objective reality constraints, portrait rights, and challenges across every technical layer.

But voice only needs to meet four key metrics: clear audio quality, natural emotion, low latency, and reliable stability.

Once these metrics are met, voice can directly enter production workflows without manual post-processing.

A lower barrier means a shorter path from "demonstratable" to "production-ready". A shorter path means faster monetization.

Fourth, and most critically, voice is the natural interface for Agents.

If AI Agents are to enter the real world, they cannot only type—they must be able to listen and speak.

Text interaction was the interface of the internet era; voice is the interface of the Agent era. Many real-world interactions never happened on a keyboard to begin with.

In daily life, voice is the default interaction method for countless scenarios.

ElevenLabs Agents do not just enable machines to participate in a complete conversation—they enable machines to fully understand user queries. This is why ElevenLabs emphasizes the "employee-like" nature of its ElevenLabs Agents.

Finally, the voice interface has another advantage: it can extend beyond screens.

Text Agents primarily rely on web pages, apps, and office software, but voice Agents can integrate with phone lines, headphones, car infotainment systems, and offline physical stores—places with no screens at all.

No matter how impressive your AI-generated videos are, they are completely useless once you step away from a screen.

03

Voice is the Core Interface of the AI Era

For us, the ElevenLabs story is far from mysterious.

Open any AI assistant app, start a voice chat, ask it to read a passage, switch to a different voice profile, and it can deliver the content with rich, natural expression.

The ability for "AI to speak" is not a rare feature in China. So why has no Chinese equivalent of ElevenLabs emerged?

Take Doubao as an example. Doubao's goal is to attract users into its ecosystem, keep them engaged, and drive usage across ByteDance's product portfolio. Its voice feature is merely an interaction method for a super app/assistant.

Doubao's voice capability is essentially part of ByteDance's broader AI interface strategy. The better it works, the more willing users are to stay within Doubao and rely on it as their primary assistant.

A major reason ElevenLabs succeeded overseas is that the international market is highly fragmented. Content companies, game studios, education platforms, customer service firms, creator tools, and developer platforms all operate on separate systems with distinct requirements.

ElevenLabs positioned itself perfectly at the center of these scattered demands, establishing itself as "voice infrastructure" and selling bundled services to all these different players.

But what about ByteDance? Short video dubbing is handled by Jianying; web novel audiobooks are served by Tomato Novels; AI assistant voice is powered by Doubao; enterprise voice calls are supported by Volcano Engine; live e-commerce narration is delivered through Douyin.

In other words, the use cases that ElevenLabs had to actively pursue one by one in overseas markets are already naturally embedded within ByteDance's own ecosystem.

It is highly unlikely that an independent Chinese company fully comparable to ElevenLabs will emerge. Because the most high-value voice scenarios have already been captured within closed domestic ecosystems.

At the same time, ElevenLabs faces another challenge: as model capabilities across the industry converge, what will its unique competitive advantage be?

Take Vapi, for example: it is a platform for developers to build AI voice Agents, allowing users to string together the entire pipeline—speech recognition, LLMs, speech synthesis, phone line integration, tool calling, interruption handling, latency control, testing, and deployment—all in one place.

Vapi's threat to ElevenLabs is that as voice model capabilities become increasingly similar, customers may no longer prioritize "which voice sounds most human"—they will prioritize which platform can get their voice Agent up and running the fastest.

ElevenLabs excels at voice quality and TTS models, but Vapi operates at the higher orchestration layer, directly serving developers and enterprises, where it can treat TTS as a swappable component.

Today it can integrate ElevenLabs, tomorrow it can switch to newer alternatives like OpenAI, Cartesia, or PlayAI. If better options appear, it can easily replace one provider with another.

But ElevenLabs? It can only use its own models. That means if future constraints like computing power shortages prevent ElevenLabs from regularly updating its models, it will inadvertently enable the success of aggregation platforms like Vapi.

This article originates from the WeChat Public Account "Face AI" (ID: faceaibang), author Miao Zheng, editor Wang Jing, republished with authorization from 36Kr.