ElevenLabs hits reported $22 billion valuation four years after founding with $600 million ARR

ElevenLabs has reached a reported valuation of $22 billion (roughly 81 billion shekels) just four years after its founding, backed by an annual recurring revenue run rate of $600 million (about 2.2 billion shekels). The company, which builds models that turn text into human-sounding speech, has become the voice layer of the AI industry even though most end users don’t realize they’re interacting with it. Klarna runs first-line phone support for 35 million U.S. customers on the platform, alongside Deutsche Telekom, Cisco, Adobe and a growing list of governments.

The split between enterprises and creators

Chief executive and co-founder Mati Staniszewski says 55% or more of revenue comes from traditional enterprise customers, with the remaining 45% divided among small and mid-sized businesses, developers, builders and content creators. The platform powers audiobooks, dubbing and music, but the big money comes from contact centers and large-scale automated conversation systems.

When customers become competitors

The boundaries between model providers, platforms and applications are blurring. Decagon, a conversational AI platform, trained its voice product on ElevenLabs and now runs queries through its own models — making it a direct competitor. Staniszewski acknowledges the lines have blurred: “In the past there were clear splits where one begins and where another ends. Today the line is much more blurred.” He points to Anthropic as an example of a model company that became a platform and then a broad suite of applications.

Open models versus the frontier

At ElevenLabs the customer chooses the “reasoning layer” from a menu of options. Staniszewski explains this isn’t a binary choice: for purely informational conversations that don’t execute actions, open models can suffice because the knowledge base defines the experience. But in financial services, where authentication, transaction details or refunds are required, there is no room for error — and frontier models still lead.

Governments, different requirements for each deployment

Some open models originate in China, yet the U.S. government and European governments are customers. Staniszewski says the conversations differ with every deployment: the models and voices deployed depend on the specific use case. The Polish government or the Brazilian government each has its own set of requirements, and the company tailors its stack accordingly.

A Turing test for voice, margins under pressure

The stated goal: pass the Turing test for conversational AI. That demands not just intelligence but emotional intelligence — understanding the emotion on the other end, slowing down or speeding up. The prediction from a year ago, that audio models would become commoditized within two years, has not materialized; the quality gap at the model level remains significant, and Staniszewski estimates it will take another three to five years before the differences narrow. On gross margins he declines to elaborate, but makes clear he is willing to see them erode further if that is what it takes to grow market share. On transparency he is unequivocal: businesses should tell customers when they are speaking with AI.