ElevenLabs

IWhat it does

ElevenLabs makes software that turns text into natural-sounding speech, turns speech into text, clones or designs voices, and translates audio into other languages. It started in 2022 with a text-to-speech model that the company later described as almost indistinguishable from a real human voice⁹. Its first funding round, in 2022, valued it at about $9 million².

The company now describes itself through three product lines¹. ElevenCreative is for people who make audio and video: generating voice-overs, editing, and dubbing into more than 70 languages at the time of the February 2026 announcement. ElevenAPI is the developer product: low-latency voice that other companies build into their own apps. Customers named by the company include Meta, Epic Games, Salesforce, MasterClass and Harvey. ElevenAgents is the newest line: voice and chat agents, meaning software that holds a conversation with a customer, answers questions and takes actions, for firms such as Deutsche Telekom, Square, Revolut and the Ukrainian government.

In September 2026 the company said its technology is in daily use at 5 of the 10 largest tech companies, 5 of the 10 largest insurers and 4 of the 10 largest telecoms, and that it supports more than 90 languages². These are company statements. I could not check them independently.

Other products listed on its site include Eleven Music, an AI music model, ElevenReader, an app that reads books and articles aloud, Voice Isolator, a Voice Library where people can earn money when others use AI versions of their voices, and Scribe, a speech-to-text model^14161718. In September 2026 it also launched Scribe v2 Medical for healthcare transcription and the Eleven v4 and v4 Turbo speech models².

The strategic idea is simple. Voice may become a main way people deal with software and companies. ElevenLabs wants to own the voice layer: the model that sounds good, the platform that runs the conversation, and the connections to a customer’s systems. The CEO says the aim is to pass what he calls a Turing test for conversation, which needs emotional intelligence as well as intelligence⁵. That is a goal, not an achieved result.

IIHow it makes money

There are three revenue streams, and the company does not publish a split between them. The first is subscriptions from individuals and small teams. Plans run from a free tier to Starter at $6 a month, Creator at $22, Pro at $99, Scale at $299 and Business at $990, billed in credits, with custom Enterprise contracts¹⁴. The second is usage-based API revenue: text to speech costs $0.04 to $0.08 per 1,000 characters depending on the model, and speech-to-text costs $0.22 to $0.39 per hour of audio¹⁵. The third is voice agents, priced from $0.08 per minute plus the cost of the language model behind the agent¹⁵.

Reported growth is being driven by the third stream. Press coverage of the May 2026 update says ARR growth came from enterprises deploying voice agents for support, sales, hiring and marketing⁶. The September 2026 company post says ElevenAgents ARR grew more than three times since February, agents handle more than 15 million conversations a week, and enterprise is 55% of revenue². Because ElevenAgents is the fastest-growing line and enterprise is now the majority, the business is shifting from selling a tool to people toward selling a service to companies. That tends to mean larger, stickier contracts but also longer sales cycles and more competition from big vendors.

Margins are not disclosed. The CEO declined to discuss gross margins in detail and said he does not mind margins falling if that grows market share and passes savings to customers⁵. That matters. A voice agent minute includes the cost of computing, of the language model behind it, and often of human-labelled training data: the company uses thousands of contracted annotators⁵. For comparison, Twilio, a mature communications platform that resells network capacity, reported a 48% GAAP gross margin in the June 2026 quarter¹². I do not know ElevenLabs’ gross margin and make no claim about it.

The run-rate figure needs care. ARR here means the latest month’s revenue multiplied by twelve⁸. It is a company-reported run rate, not audited annual revenue, and it can include usage that is not contracted for a year. Sources also disagree on the past: the CEO said over $330 million at the end of 2025 in January, and the company said about $350 million in May⁴⁸.

Reported run-rate revenue (ARR) and valuation markers

IIIThe market and the competition

The market is anything where software speaks or listens: customer service calls, sales outreach, appointment reminders, audiobooks, video dubbing, games, accessibility tools. The best-known near-term prize is the contact centre, where companies pay people to answer calls and chats. Voice agents try to take part of that work. One example in the CEO interview is a Polish healthcare deployment that reminds patients of appointments, where 18% of patients otherwise miss them⁵. I have not seen a reliable independent estimate of the total market, so I do not give one.

Competition comes from four directions. First, the large AI labs. OpenAI sells a live voice model, GPT-Live 1, that listens and speaks at the same time and handles interruptions, priced at $0.05 per minute of session with the backend model billed separately¹⁰. ElevenLabs lists its agent platform from $0.08 per minute plus the language model¹⁵. The two prices are not like for like, but they show how close a lab’s list price sits to ElevenLabs’. Other labs, including Google, also sell speech models, but I did not open a primary source for them and do not rate them here.

Second, specialist voice companies. Names that come up repeatedly in comparisons are Cartesia, Deepgram, Hume AI and Resemble AI. I found only aggregator data on their funding, which disagrees across sites, so I do not quote it. Third, customers who become rivals. The CEO’s own interview notes that Decagon, a customer-service AI platform, trained its voice product on ElevenLabs and now competes with it⁵. Fourth, open-source speech models, which keep pushing the price of plain text to speech toward zero.

The public comparables are poor fits. SoundHound AI, a listed voice AI company, reported first-quarter 2026 revenue of $44.2 million, up 52%, a GAAP net loss of $25.0 million and $216 million of cash, with 2026 revenue guidance of $225 million to $260 million¹¹. It is a fraction of ElevenLabs’ size. Twilio, which sells communications infrastructure, reported $1.50 billion of revenue in the June 2026 quarter, up 22%, and raised its 2026 free cash flow guidance to $1.135 billion to $1.155 billion¹². Twilio is a sensible ceiling for how a profitable communications platform is valued today, and ElevenLabs is growing much faster.

IVThe moat and how it could be disrupted

My rating is contestable. ElevenLabs has real advantages, but I cannot show that they last.

What it has. A quality lead in expressive speech, which independent benchmarks reportedly show for the v4 models, according to the company itself². I could not check those benchmarks. A large body of voice data and human annotation. A voice library where creators license voices and are paid, which is a small two-sided market³. Enterprise integrations, security reviews, data-residency options for governments and the know-your-customer checks it applies to every customer⁵. And a head start in sales: 14 local go-to-market hubs funded by the Series D¹.

Why that is not enough yet. Voice quality is the kind of lead that shrinks. Each new model from a rival closes the gap, and buyers often cannot hear small differences. Switching costs for the API are low: a developer can swap one speech endpoint for another with modest effort. Switching costs for agents are higher, because a company builds workflows, knowledge bases and phone integrations around the platform, but that is the same fight that every contact-centre software vendor already runs. The CEO himself says the lines between model companies, platform companies and application companies are blurring⁵, which means the labs that own the language model can move into voice, and voice start-ups can be squeezed from both sides.

Price is the clearest pressure. The CEO says he accepts lower margins to grow share and pass savings on to customers⁵. That is rational for a company in a land grab, but it also tells us that the pricing power is not strong. Compare the agent price of $0.08 a minute plus model cost¹⁵ with a lab offering $0.05 a minute for the voice session¹⁰.

How it could break. A frontier lab bundles very good voice into its model at near-zero marginal price. A large customer builds in-house. Cheaper open models match v4 quality. Regulation of voice cloning raises cost or limits the product, after misuse cases such as the January 2024 fake Joe Biden robocalls that were reportedly made with its tools²⁰. Or the agent market commoditises into a thin layer on top of a language model.

VThe numbers we can see

All revenue figures are company-reported run rates. None are audited, and I have found no filed accounts. The company is incorporated in the US and, as a private company, has no public filing duty that I could find, so I could not check numbers against a regulator document. The best I can do is put the company’s own statements next to each other and check they are consistent.

They are consistent in direction. ARR was over $200 million in September 2025, about $350 million at the end of 2025, above $500 million by spring 2026 and about $600 million in September 2026⁸⁴⁶⁵. Going from $350 million to $600 million in about nine months is growth of roughly 70% in that period, or something above 100% annualised. The CEO said it took five months to go from $200 million to $300 million of ARR⁷. Growth is clearly slowing in percentage terms as the base grows, which is normal.

Two details deserve a flag. The first is dates: one source places the $600 million milestone as early as July 2026 while the CEO’s TechCrunch interview is dated 24 September⁵. I use September. The second is the quarter arithmetic: the CEO said ElevenLabs added $100 million of net new ARR in Q1 2026 and ended the quarter near $450 million⁴, while tech.eu says it passed $500 million in Q1⁶. These can both hold only if the $500 million was reached after the quarter ended, or if the two outlets used different definitions. I treat $500 million as reached by around April.

Funding is also reported in two ways. The February announcement said $500 million at $11 billion, with $781 million raised in total¹. By May the Series D had grown past $550 million after a third close⁶. A further $100 million employee tender was announced in May⁴, and the $300 million September tender is separate². One data site counts total funding as $881 million because it includes the first $100 million tender⁸. I use the company’s $781 million primary figure and keep tenders apart.

Headcount is unclear in public data: the company says over 800², while a data site shows an earlier estimate of about 400⁸. I use the company’s figure. Profit is not disclosed at all.

Valuation history of ElevenLabs

VIWho runs it and who owns it

Mati Staniszewski is cofounder and CEO. He was previously a deployment strategist at Palantir. Piotr Dabkowski is cofounder and CTO and was a machine-learning engineer at Google¹⁹. The company has a Warsaw office¹, and TNW reports that Poland’s state investment fund took an $11 million stake in mid-2026⁷. The Series D added Sequoia partner Andrew Reed to the board¹.

Owners. The company has not published a cap table, so I cannot say who owns what. Investors named by the company include Sequoia, Andreessen Horowitz, ICONIQ, Lightspeed, BOND, Evantic and others in the Series D¹. Later closes added BlackRock, Wellington, D.E. Shaw, Schroders, Nvidia’s venture arm, Santander, Salesforce Ventures, KPN and Deutsche Telekom⁶⁴. More than 30 entertainment figures also invested⁶. The September tender was led by Wellington and T. Rowe Price, with new holders including BDT and MSD, EQT, GIC, Goldman Sachs, OTPP and Sapphire².

What an outside owner can take out. Today, nothing is paid out as dividends that I can see, and a private holder can only sell through company-run tenders, which are at the company’s discretion. The company said retail investors would be able to invest through Robinhood Ventures, with no details⁴. Preferred shares from a venture round usually carry rights, such as a liquidation preference, that ordinary or employee shares may not. I could not see the terms, so I cannot say whether a price paid in the tender is comparable to a price paid in the Series D.

Governance: founder stakes are not disclosed. The late-stage shareholder base now includes large mutual and hedge funds such as Wellington, T. Rowe Price and D.E. Shaw²⁶. Funds like these often buy late in private rounds ahead of a possible listing, though that is a pattern, not a stated plan here.

VIIValuation: is the last price supported?

Equity value in US$ billions. Not a share price: a private company has none. Latest mark: employee tender (secondary), reported by the company; pre- or post-money not stated.

What the last valuation assumes

At $22 billion and $0.6 billion of run-rate revenue, the price is about 37 times today’s revenue. To justify that on a 12% required return after dilution, the company has to reach roughly $4 to $5 billion of revenue by 2030 and still be valued at a double-digit revenue multiple, with margins good enough to support it. That sits above my bull case on revenue and at its multiple. In plain terms, the price assumes ElevenLabs becomes the default voice layer for companies, keeps pricing power against the large labs, and keeps growing faster than 30% a year for four more years.

Bear, base and bull

• Bear. ARR of about $0.72 billion at end 2026, then growth of 30%, 20%, 15% and 12% to about $1.45 billion at end 2030 as labs and open models cut prices. Terminal value 4x revenue, in line with the median public software multiple of roughly 4x¹³. Discount rate 15%, dilution 3% a year. Equity value about $2.9 billion, rounded to 3.
• Base. ARR of about $0.78 billion at end 2026, then growth of 60%, 40%, 28% and 20% to about $2.7 billion at end 2030, with enterprise voice agents carrying most of it. Terminal value 7x revenue, above the 6.6x software average and the roughly 5x I estimate for Twilio, below the 12x of top-quartile names such as Datadog¹³¹². Discount rate 12%, dilution 2% a year. Equity about $10.9 billion, rounded to 11.
• Bull. ARR of about $0.85 billion at end 2026, then growth of 80%, 55%, 38% and 28% to about $4.2 billion at end 2030, as voice becomes a main software interface. Terminal value 10x revenue, near Snowflake and MongoDB at 12 to 13x less a discount for lower margins¹³. Discount rate 10%, dilution 2% a year. Equity about $26.1 billion, rounded to 26.

Comparables used

Where I come down

The $22 billion tender price is about double my base value of $11 billion and close to my bull value of $26 billion. My margin-of-safety line is $8 billion, so the price is far above it. What you would have to believe for $22 billion to be right: run-rate revenue of about $4 billion or more by 2030, a terminal multiple of at least 10x, and little price erosion. Why the market might be right even though it looks high: growth has been unusually fast, with ARR up roughly 70% in nine months⁴⁵; ElevenLabs sits at the point where voice interfaces spread; late-stage buyers such as T. Rowe Price, Wellington, GIC and Goldman Sachs bought at that price²; and in AI markets the winner often has a much larger business than any early model predicts. Three further cautions on the price itself: a tender is a smaller, informal deal where buyers get limited information, the buyers are not all paying for new equity in the company, and a tender price is not the same as a priced round with full preferred rights. My model also ignores interim cash flows and any net cash on the balance sheet, which would nudge each case up modestly.

What I cannot judge

I cannot judge gross margin, profit, customer concentration, churn, contract length or the real split between subscriptions, API and agents, because none is disclosed. I found no filed accounts. ARR is a company-reported run rate. I could not open primary sources on Google, Cartesia, Deepgram or Hume, so I do not rate them. I did not verify current public multiples for Twilio or SoundHound, and the comparable multiples I quote from SaaStr are one author’s February 2026 estimates, so the terminal multiple is the weakest input. Because public data is this thin, treat my range of $3 billion to $26 billion as wide, which it is, and not as a forecast.

VIIIThe IPO

Reported to be preparing an IPO. No filing, public or confidential, has been reported. The CEO told TechCrunch in September 2026 that the company is preparing the foundation to go public in the next years, depending on time and place, and did not confirm a 2028 target that has been reported⁵. In March 2026 he said it aims to be IPO-ready in two to three years⁷. A London listing has been mentioned as a possibility if the UK improves its listing conditions; I did not open a primary source for that. No exchange, ticker, banks or size have been reported. Retail access through Robinhood Ventures was mentioned in May without detail⁴.

What to watch

• A confidential or public filing, or the appointment of banks, reported by Reuters, Bloomberg or the FT
• Audited financials or gross margin disclosed in any form
• Another tender or primary round and whether it prices above or below $22 billion
• The choice of exchange (London, New York or a dual listing including Warsaw)
• Price moves by OpenAI, Google and others for voice agents compared with $0.08 a minute¹⁰¹⁵
• Whether ARR growth stays above 30% a year and enterprise share keeps rising above 55%²

IXWhat could go wrong

What most likely destroys the business is not a collapse in demand but price. If voice becomes a feature that the big labs give away inside their models, a stand-alone voice vendor has to earn its margin from workflow, integration and trust, which are harder to defend than sound quality. The business would still exist, but at a much lower value than $22 billion.

What has to go right at the same time, in my base case. Four things, and I multiply my rough odds. One, growth stays high: ARR reaching about $1.25 billion in 2027 needs about 60% growth. I put that near 55%. Two, gross margin holds up while agent usage grows, which I put near 60% given the CEO’s stated tolerance for lower margins. Three, quality or integration advantages keep the company from being priced like a commodity, near 60%. Four, no legal, regulatory or safety event damages enterprise trust, near 85%. These are loose estimates and they are not independent, but treating them as independent gives about 17% for all four together. That is why I read the base case as a central path, not a likely one, and why the range from bear to bull is so wide.

Other risks. Concentration: if enterprise agents are most of the growth, a few large contracts could matter a lot, and the company does not disclose customer concentration. Misuse: voice cloning has been used in fraud and disinformation, including the January 2023 abuse of its beta to imitate celebrity voices²¹ and the January 2024 fake Biden robocall noted above²⁰; the company says it applies KYC to every customer and bars self-replicating agents⁵. Intellectual property and voice-rights rules for AI audio are still being written in several countries, and I have not assessed them. Key people: the founders are central. Liquidity: tender offers can stop, and the secondary price is only as good as the next buyer.

The next best alternative for the same dollar. Public comparables such as Twilio are cash-generating, trade on a free cash flow yield that I can compute from its guidance, and are growing at about a fifth¹². They offer much less upside. SoundHound is a smaller, loss-making listed voice peer¹¹. A buyer of ElevenLabs at $22 billion is choosing a potentially much larger winner at a price that already assumes a lot. That is a choice of risk, not an obvious mistake.