A 40-person team at $35 million ARR: HeyGen made AI avatars a business, not a demo

Nathan Labenz’s Cognitive Revolution intro is read by his own HeyGen avatar. The rest of the setup is even stranger. During recording, Joshua Xu’s video track looked fine and then turned out to be corrupted. The team considered shipping audio only. Instead they regenerated Xu’s side of the conversation from the transcript with HeyGen’s own avatar product and cut it into the episode. After the intro, Labenz and Victor Lazarte are live; Xu is a digital twin. Labenz’s verdict: it came out amazingly well. A company that makes AI talking heads used the product to save its own podcast appearance. It is hard to invent a cleaner demo.

The business behind the stunt is already past the novelty stage. Labenz notes that while Sora, Google Veo, Kuaishou’s Kling, and Runway Gen-3 have been filling feeds with general video models, those tools remain unwieldy and inconsistent for day-to-day work. HeyGen skipped the foundation-model arms race and focused on practical business video: quality, consistency, and control, assembling its own avatar models with best-in-class script writing and text-to-speech. The company now serves more than 40,000 businesses and does $35 million in annual revenue, with about 40 to 50 employees. Benchmark recently invested $60 million; Lazarte, a general partner there and previously a co-founder of the Brazilian mobile-game company Wildlife Studios, led the round.

Labenz, who spent three years applying AI to video at Waymark, says the conversation is a reminder to think bigger than cheaper versions of today’s ads. Xu and Lazarte describe translation into many local languages, personalization down to a single viewer, and a coming shift from recorded clips to live, interactive avatars. Trust and safety is not an appendix. At a moment when many AI apps ignore misuse, Labenz ranks HeyGen’s consent and anti-impersonation stack as industry-leading — and invites listeners who like breaking guardrails to help him red-team more products.

Xu’s path runs through Snap. From 2014 to late 2020 he first built advertising systems that helped advertisers get return on Snapchat, then moved to the AI camera team. In 2018 there was not yet a consumer category called “AI,” but there were GANs: a computer could generate a picture of something that did not exist and still look realistic. Standing on that frontier, he had the feeling the technology would change how people create content.

The last 10 to 15 years of content platforms, he argues, rode the rise of the mobile camera. The iPhone 4 unlocked Instagram, Snap, TikTok, and a generation of creators. Working on camera software, he still saw the same bottleneck: many people cannot make good content with a phone camera. They are shy, they do not have time, or they are not good at performing. If you could replace that component — replace the camera — you could unlock visual storytelling for everybody. HeyGen was founded in December 2020 on that sentence: use AI to generate video instead of asking people to film it.

Labenz asks a skeptical question for a world already awash in content: where is more video actually needed? Xu’s consumer-side answer is appetite. People live in a video-first world; more than a billion hours of video are watched on YouTube every day, in his telling. When his Tesla had a problem he did not open the manual. He looked up a video, because video is easier to consume. Every business wants to make more of it. The constraint is that video is expensive. HeyGen’s job is to let companies produce at the pace their customers already watch.

He buckets usage into create, localize, and personalize. Create: pick a stock avatar or clone yourself, grab a template, type a script, generate a product explainer, demo, announcement, or training clip. Stock avatars are a user-journey device, not the destination — people try the tech and test whether the content works, then clone themselves and build a brand identity around the twin. Localize is the breakout. Upload an existing video and HeyGen can recast it in many languages while preserving the original speaker’s voice and expression, with lips that match the new audio. Traditional localization hires a voice actor at roughly $10 to $20 per minute, takes days or weeks to coordinate, and still looks dubbed. HeyGen’s version is, he says, more than 10x cheaper and about 100x faster, and it is the first translated cut that feels native. People used to prefer original-plus-subtitles; lip-sync changes the engagement math. Large companies with dedicated localization teams have used it to shrink workload, cost, and time by more than 10x. Labenz’s back-of-envelope: 100 product videos into 100 languages is 10,000 videos. Old-school, you simply would not do it. In HeyGen’s world the cost is GPU time.

Personalization is newer. Email already says “hey Nathan.” HeyGen is inventing the video version. Names are only the starting point; the content itself can change. One customer generated more than 100,000 personalized videos and sent them to employees as thank-yous. HeyGen is a horizontal platform — marketing, sales, learning and development, localization — from small businesses to Fortune 500 companies. About 1% of users are professional editors who take HeyGen footage into Adobe or Final Cut as a camera replacement. The other 99% never made video before and finish camera and edit inside HeyGen. Xu’s demand-expansion example is the one that sticks: a company that needs 100 videos a month might have used HeyGen for 10 spokesperson spots. Make those 10 cheap and fast, and they do not stop at 10. They make 100. Better tools replace old workload and enlarge the need.

Labenz puts Xu on the uncanny valley. Advertising viewers swipe in half a second; a mandatory corporate training video gets finished no matter how dull the presenter. Are we still in the valley, leaving it, or does it depend on the job? Xu splits the answer. External marketing and ads have a much higher quality bar because brands put real budgets behind them. On raw likeness, they have already cleared the valley in a lot of cases: look at some of his own avatars and you probably could not tell avatar from human. That is not the hard problem.

People watch video instead of reading text or listening to a clip because it is more engaging. The question is not what makes a video look real. It is what makes it engaging. Conversation flow is subtle and decisive. Natural gesture — using your hands to articulate a point — is still hard. Emotion is the third piece: the same line in a different tone means something else. Those are the areas HeyGen is still attacking, and they are the breaking point for high-end external advertising. Labenz calls the reframe profound. A year or 18 months earlier, early video models had a psychedelic, wavy superstimulus quality, a bit like the trippy inputs that light up a single neuron in a vision model. Everyone is racing toward realism and forgetting that cartoons are popular. Lifelike video is one path to engagement, not the only form factor that can win.

Lazarte comes in from games. Realism is not the most important thing; engagement is. Humans already choose video when they have a choice. At Wildlife Studios, user acquisition moved from static ads to video ads, and the company was doing about 100 million video views a month. Conversion rate was a life-or-death KPI. In his experience the videos that converted did not win on higher fidelity. They won because the team could test a bunch of variants. The keys to testing are getting the price down and the iteration speed up. That is the powerful thing HeyGen does for businesses: cheaper, faster clips mean you can try a lot more stuff. What excited him as an investor is product velocity, not a finished aesthetic. Streaming Avatar — a digital clone that talks to you in real time — is the example. ChatGPT made talking to a bot feel possible, but typing is suboptimal; GPT-4o added voice; the next step is a video call. He says HeyGen has the best product in the world for that, and nobody else has the technology. What companies will do with it is still unknown. Virtual friends and pets are guesses. Cartoon avatars are already in the mix. They are just getting started.

Labenz is queasy about the near-term personalization. An AI video from his boss that only inserts his name would bum him out. In past large-brand projects he has seen, teams love the idea, then get shy: top hundred names, then “hi friend,” maybe a utility-bill figure, almost nothing that actually needs a model. Xu says looking at personalization only at the name variable is still a pretty narrow view. Video can be generated on the fly and the content itself can change — the message, the story, the angle. For 20 years, two people watching a video had to watch the same one, as if everyone sat in the same theater. They do not have to. A clip with 100 million views gets optimized for “what most people would probably like,” and every iteration sacrifices someone. Serve each arriving viewer a different version and you can search for a global maximum instead of a mass-mode compromise. Two technologies have to exist together: high personalization, and generation on the fly. Streaming Avatar is only the starting point. The end state is streaming the entire video, generation included, not just playback.

Lazarte is practical about the ick. A boss video that only personalizes a name will not make you feel good. A video that intakes what you actually care about and delivers that content is a different product. The applications have to be tested until the ones that feel like they are in the user’s favor show up. On the company itself: 40 to 50 people at $35 million ARR is enough to cover the team, so Labenz asks where the new capital goes. Xu says they never saw themselves as an avatar company. Avatars — the A-roll, the human spokesperson — were the first unsolved piece. The next act is B-roll: background, music, transitions, animation, so they can generate the entire video end to end. Quality and engagement on A-roll keep rising; a lot of investment goes into B-roll generation, model training, and product. Improving quality is not one trick. More data covers more edge cases. Video generation is also not only a math problem: how do you even modularize conversation flow so a model can be trained on it? The mix is more data, more compute, more model breakthroughs, and more product innovation.

Asked whether the studio timeline will grow into Adobe territory or a Sora-like foundation model, Xu rejects both as the frame. He has a very aggressive, strong point of view: the timeline editor will be gone in five years. Today’s video world was defined by that editor 20 or 30 years ago because the camera was expensive. You filmed 20 or 30 times and laid the takes on a timeline to pick the best one. If footage can be generated on the fly, you do not need the timeline, and the timeline is a brutal learning curve for the 99% who are not editors. There has to be a new paradigm. For business video, the need is quality control and consistency, not a general world model. HeyGen’s bet is a composition engine that understands what the brand needs, then orchestrates avatar, voice, music, and B-roll. Labenz asks if the future is sitting in front of a screen saying “replace that background with a green park.” Putting a person in a different scene is possible today. The failure is lighting: drop him on a beach and the light feels wrong; solve the static case and gesture still has to carry dynamic light. Xu’s sharper turn is that lighting itself is a camera-era dimension. In a no-camera world the dimensions that matter become “more engaging,” “more dynamic,” a more fluent conversation flow. Labenz has watched the same simplification at Waymark, where a tangle of captioning and aesthetic models collapsed into asking one large multimodal model which images to use — fewer tools, better results. Xu agrees the stack will probably narrow to a couple of major models behind the orchestration layer, not a thousand specialist gadgets forever.

The last stretch is why Labenz wanted the episode. He was a GPT-4 red-teamer and has since watched cloning products stand up with no guardrails at all. He has cloned Donald Trump, Joe Biden, and Taylor Swift’s voices on many platforms, including some that let you place a call to an arbitrary number and solicit a donation. It still blows his mind that people ship that. HeyGen, by contrast, has the best integrated security measures he has seen, and it has done so while building a real business. Lazarte, looking at several avatar teams, assumed the technology would exist and customers would use it. The risk of using it badly was huge. What stood out was that becoming the most trusted brand was central to HeyGen’s strategy, not a compliance afterthought.

Xu says they have never treated trust and safety as something that slows them down. It is a critical piece of the business and part of the product. Every new experience gets a session on how to roll it out safely. They serve some of the largest companies in the world; without this stack the deals would not close. Creation: no political or celebrity figures on the self-serve path — a celebrity who has a legitimate use case has to contact the operations team. Every avatar or digital twin requires first-party consent, in video. There is randomly generated dynamic-code verification plus human review in the back. Every video created or localized goes through a combination of an AI model and a human reviewer against policy: no hate speech, harassment, fraud, or misinformation. Celebrity matching uses vendor databases plus HeyGen’s own team constantly updating coverage.

Labenz went through the product with the goal of breaking it and never did, at least not in the flagrant ways. He got a little offensive speech here and there. Uploading a Trump or Biden video and trying to mint a public-figure avatar failed. It felt like defense in depth: he systematically found several checks and finally landed on human review saying no. He has been looking for the third-party celebrity-database vendors because so much of the rest of the cloning market is wide open to abuse. Developers often tell him they do not know how they are supposed to detect every celebrity voice. If the answer is “pay for a database and use it,” he wants to popularize that. His closing read of the hour is the same as Xu’s: do not look at today’s avatar and think that is the end of the line. They will get better, faster, real-time, more contextual. The paradigm of what a video even is may get blown open. The future is a little weirder, and maybe cooler, than people first dare to guess. Study HeyGen’s product if you need a safety reference. On his personal power ranking of guardrails, it came out on top.

Xu and Lazarte say they have covered the important things. Labenz’s own landing is practical. For operators: drop existing product videos into translation for one target market; the cost is GPU time, orders of magnitude under a voice-over team, and you can scale after a small proof. For performance marketers: stop polishing one hero spot and generate twenty script variants, spending budget on tests the way Lazarte’s game studio did. For anyone new to the product: walk the path HeyGen designed — stock avatar first, clone yourself only after the content works.

For people building cloning products, he wants the rest of the market to copy the consent flow, the dynamic verification, and the vendor celebrity lists he is still hunting down. Ordinary users can spend half an hour making a twin of themselves even if they do not need it this week; it is the fastest way to feel how close the technology already is. Avatars today are a starting point. They will be faster, more live, and more aware of context, and video as a concept may be sitting on the edge of being rewritten.