ElevenLabs Says It’s at $600M ARR and May Be Worth $22B — But Its Real Bet Is Voice AI’s Next Phase

Updated September 24, 2026 8:24 pm

ElevenLabs, the voice AI startup best known for turning text into natural-sounding speech, is now reportedly valued at $22 billion, according to investor pricing, as the company says it is running at about $600 million in annual recurring revenue. The surge underscores how quickly enterprise demand for AI voice systems has accelerated, even as the company faces rising competition from customers that are building rival products on top of its technology.

That was the backdrop for a wide-ranging conversation with co-founder and CEO Mati Staniszewski in Toronto, where he described a market that is moving fast, becoming harder to define, and increasingly shaped by questions about disclosure, model choice, margins, and safety.

ElevenLabs has become one of the clearest examples of how generative AI is moving from chat interfaces into the less visible infrastructure behind customer service, government call centers, audiobook production, dubbing, and multilingual content tools. In practice, the company’s software often powers the voice a customer hears when they think they are speaking to a support agent — or a system that sounds close enough to human to blur the difference.

But the company’s rapid rise also highlights a new competitive reality in AI: platform providers are no longer just selling tools to customers, they are sometimes becoming suppliers to would-be rivals. That tension was central to the conversation, as Staniszewski argued that the industry’s boundaries are getting fuzzy and that model companies, platforms and applications are converging.

What exactly is ElevenLabs selling?

ElevenLabs sells the voice layer of artificial intelligence: models that convert written text into speech that sounds natural, expressive and increasingly difficult to distinguish from a human voice. The company’s customers range from multinational enterprises to governments and individual creators.

Its corporate users include firms such as Klarna, Deutsche Telekom, Cisco and Adobe, along with public-sector customers that use the technology for citizen services and internal operations. On the consumer and creator side, ElevenLabs also provides tools for audiobooks, dubbing, voice localization and music-related projects.

That mix matters because it helps explain why the company has been able to grow so quickly. The same core technology can be deployed in multiple markets: customer support, media production, accessibility services and government communications. In each case, the product is voice, but the business need is different.

How the product is used in the real world

One of the clearest examples of ElevenLabs’ scale is customer support. Klarna, the buy-now-pay-later company, has used the technology to handle first-line phone support for tens of millions of U.S. customers. That kind of deployment illustrates how voice AI is moving from novelty to infrastructure, especially in high-volume, repetitive interactions.

Government deployments show another side of the market. In some cases, the technology is used to reduce missed appointments, improve call-center efficiency or provide language-appropriate information across public services. Those use cases are often shaped by local data laws, residency requirements and authentication rules, making each deployment more customized than a consumer app.

Staniszewski said the company’s growth comes from proving value in real customer workflows, even if that means accepting thinner margins in the near term.

How big is ElevenLabs now?

The company says it is now pacing at roughly $600 million in annual recurring revenue, a figure that would place it among the fastest-scaling private AI startups in the voice category. Staniszewski said a majority of that business — more than 55% — comes from traditional enterprise customers, while the rest comes from small and medium-sized businesses, developers, builders and creators.

That revenue mix is notable because it suggests ElevenLabs is not reliant on a single buyer type. Enterprise sales bring longer contracts and larger deployments. Smaller customers and creators, meanwhile, can help spread the product quickly and expand usage across new creative and technical workflows.

The reported $22 billion valuation, if accurate, implies that investors believe the company’s addressable market is still widening rather than plateauing. It also reflects a broader trend in AI investing: the market continues to reward companies that can show both revenue and strategic importance, even when gross margins are under pressure from model inference costs.

Why does ElevenLabs see the market becoming blurrier?

The company’s CEO believes the old boundaries between model developers, platforms and applications are collapsing. That shift is partly because customers increasingly want integrated products, not isolated model access. It is also because businesses using AI often end up building their own layers on top of the tools they buy.

That creates a strange competitive loop: ElevenLabs can power a customer’s support stack today, while that same customer may later decide to build a voice product that competes with ElevenLabs’ own offerings. The company says it is comfortable operating in that environment because the market is expanding fast enough for multiple winners.

Staniszewski pointed to Anthropic as an example of this trend, describing how a company that may have begun as a model provider is increasingly also a platform and, in some cases, an application company. That evolution, he suggested, is not a bug but the direction the whole sector is heading.

What does competition from customers look like?

Competition from customers usually appears when a company first adopts a platform and then starts internalizing parts of the stack. A customer may begin by using ElevenLabs’ voice models, then fine-tune a voice product around its own data, workflows and user experience. Over time, that can reduce dependency on the original vendor.

That pattern is already familiar in AI infrastructure. Cloud platforms, model APIs and developer tools often become the base on which customers build differentiated products. Once those products gain traction, the supplier and the customer may end up competing in adjacent markets.

ElevenLabs appears to accept that risk as part of the broader expansion of AI demand. The company’s argument is that the total market is still growing quickly enough that losing some customers to in-house builds does not outweigh the gains from widespread adoption.

How does ElevenLabs choose models for customers?

ElevenLabs lets customers choose among different reasoning layers and model types, and the company says the best option depends on the task. For low-risk informational calls, open-weight models may be sufficient. For regulated, transactional or high-stakes use cases — such as financial services — frontier models still have the edge.

That distinction matters because voice AI is not just about sounding realistic. It also has to behave reliably when it is handling identity checks, transaction details, appointment scheduling or other tasks where errors can create legal, financial or reputational harm.

Open-weight or frontier models: which wins?

The answer is neither one is universally superior. According to Staniszewski, open-weight models can work well when the system is primarily retrieving and presenting information from a knowledge base. But where authentication, compliance and precision are essential, leading proprietary models remain more attractive.

In other words, model selection is becoming a product decision rather than a philosophical one. Enterprises are choosing based on task sensitivity, data restrictions, cost and reliability, not ideology. That could benefit companies like ElevenLabs that are willing to remain model-agnostic and adapt to customer needs.

Why governments are becoming important customers

Governments have become a meaningful market for AI voice services because they operate massive call volumes, face citizen-service bottlenecks and often need multilingual communication. ElevenLabs says public-sector use cases already include health systems, appointment reminders and government support lines.

Staniszewski said requirements can differ sharply by country. Some governments want open-weight models, some prefer closed-source systems, and some bring their own fine-tuned models. Others require strict data residency so information never leaves national or regional boundaries.

One example he cited involved Poland’s public healthcare system, where missed appointments are a persistent problem. In that setting, the deployment is aimed at automated reminder calls, with the goal of reducing the portion of patients who never show up.

He said public-sector implementations are shaped by local legal and operational constraints, including data residency and model selection.

Should AI callers disclose themselves?

Staniszewski believes businesses should tell people when they are speaking with an AI system rather than a human. In his view, today’s consumers are not yet accustomed to agent-to-agent interactions, and undisclosed automation can feel deceptive if the caller believes a person is on the line.

He also suggested that the social norm may change over time. As more people use personal AI assistants to interact on their behalf, calling a business could eventually mean expecting to reach another agent rather than a live employee. But for now, he argues, disclosure is the right default.

Why disclosure matters now

Disclosure is becoming a practical trust issue as much as an ethical one. Customers may be more willing to engage with AI if they know what they are dealing with, especially when the system is fast, helpful and transparent about limitations. In some cases, companies can even turn disclosure into a feature by offering a choice between a human queue and an automated path.

Staniszewski’s view is that businesses should clearly present that option rather than trying to hide the technology. If customers see a long wait for a human and a shorter wait for an AI, many will choose the machine — and then judge it on performance.

What did Staniszewski say about margins?

He did not provide detailed gross margin numbers, but he made clear that the company is willing to accept pressure on profitability if doing so helps it expand more quickly. That stance reflects a familiar logic in high-growth AI companies: capture market share first, optimize margins later.

The economics of voice AI can be challenging because inference, model calls and customization can be expensive, especially at scale. But Staniszewski suggested that ElevenLabs’ research and fine-tuning capabilities help the company constrain models efficiently enough to preserve value for customers, even if the company shares more of the savings.

In his framing, the central objective is not maximizing each transaction. It is proving usefulness, building trust and expanding deployment over the next several years.

How does ElevenLabs train its models?

ElevenLabs says the quality of its system depends heavily on annotation work, not just raw volume of audio data. The company reportedly uses thousands of contractors to label speech, timing, emotion, cadence and pronunciation details, and it also brings in voice coaches to improve accent recognition.

That approach reflects a broader truth in AI: raw data alone is not enough. For voice systems, the challenge is not just transcription or synthesis. It is understanding tone, pace, emotion and the subtle cues that make speech sound natural rather than robotic.

In some cases, the company has worked directly with customers to develop custom models for particular use cases. That can make the resulting product better aligned with a client’s needs, especially where language, accent or conversational style matters.

What is synthetic data doing here?

The company did not give a precise breakdown of how much training data is synthetic, but the emphasis on annotation suggests that human-labeled examples remain central. Synthetic generation may be useful for scaling and experimentation, but high-quality speech systems still rely on careful labeling and review to capture nuance.

That also explains why ElevenLabs describes its research element as a competitive advantage. The more the company can refine behavior, pronunciation and emotional expression, the more value it can add beyond commodity text-to-speech output.

Is an IPO in the works?

Staniszewski would not confirm any specific public-listing timeline, including the reported idea of an IPO in 2028. He said the company is building a foundation that could support a public offering in the coming years, but the exact timing would depend on market conditions.

That answer is typical of a fast-growing private company that wants optionality. Public markets may eventually reward revenue scale and strong enterprise traction, but the company does not appear eager to tie itself to a fixed date before it is ready.

For now, the emphasis seems to be on durability. The company says it wants to build something long-lived, not just something that benefits from the current AI funding cycle.

How exposed is ElevenLabs to AI safety and security risks?

Staniszewski argued that ElevenLabs is not designed in a way that would allow the kind of self-replication or agent proliferation that has raised alarm around some AI systems. He said the company does not deploy the intelligence core of agents in the same way as some broader AI platforms, and that every customer goes through know-your-customer checks.

That does not eliminate cybersecurity risk, but it suggests the company sees itself as operating one layer removed from the most alarming frontier-model concerns. The broader risk, he said, remains a world that adopts more capable systems without enough precautions.

He also indicated that major AI players are broadly aligned on the need to move carefully, even if they disagree on how public their caution should be or how much regulation is appropriate. The tension, as he described it, is between speed and responsibility.

Why this moment matters for voice AI

Voice is one of the most consequential interfaces in AI because it collapses distance between people and machines. A text chatbot can feel obviously artificial. A voice system, by contrast, can create the impression of a person who is present, patient and responsive, even when no human is involved.

That makes the category commercially powerful and socially sensitive. Companies want the efficiency and scale. Consumers want clarity, reliability and fairness. Regulators are likely to pay more attention as these systems become common in banking, healthcare, telecom and government services.

ElevenLabs sits at the center of that shift. Its growth suggests the market is far beyond experimentation. But its next challenge may be harder than turning text into speech: building a business that remains trusted as voice AI becomes more ubiquitous, more regulated and more competitive.

Timeline of ElevenLabs’ rise

What happens next?

ElevenLabs is likely to keep balancing three things at once: rapid enterprise growth, product expansion across creators and SMBs, and the need to stay trusted in sensitive deployments. The company’s willingness to let customers choose models, accept margin pressure and disclose AI interactions suggests a pragmatic strategy rather than a doctrinaire one.

That may be exactly what the market rewards next. In voice AI, the winners may not simply be the most realistic voices. They may be the companies that can package reliability, transparency and operational flexibility into systems that businesses feel safe deploying at scale.

For now, ElevenLabs appears to be betting that the world will keep asking for voices that sound human — and that the company can supply them across enterprise, government and consumer use cases faster than competitors can catch up.