Joshua Xu and the camera he wanted to put away

Joshua Xu made his own face a demonstration. The HeyGen co-founder, who describes himself as a person with little enthusiasm for public exposure, put an AI version of himself into onboarding videos, product demos and social marketing. His avatar could do the talking. It was a neat arrangement for an engineer who had spent years working on cameras and had discovered that he would rather not perform for one.

The choice carries the small comedy of a founder solving his own problem in public. To explain why someone might want an artificial presenter, Xu became an artificial presenter. The sales pitch and the evidence arrived wearing the same face. Behind that decision is a serious question about video: when did having an idea become a job that also required lighting, a lens and a tolerable opinion of your recorded voice?

Xu’s career sits inside that question. He helped build software for a world increasingly comfortable taking pictures of itself. He then co-founded a company for people who wanted another way to take part. The distinction matters to understanding him. His interest in synthetic video grew from close familiarity with the real apparatus, including the person expected to stand in front of it.

An engineer on the wrong side of the lens

After graduating from Carnegie Mellon in 2014, Xu joined Snapchat. His work began in advertising, including ranking and recommendation systems. He spent roughly six and a half years at the company, moving in his final years toward the AI camera. There, software could alter an image rather than merely capture it. Generative technology opened up a possibility that continued to interest him after he left.

In his telling, that work revealed two sides of the same device. The camera was becoming more capable. Testing camera software still required him to take selfies or record videos, an activity he found awkward. An engineer could make the tool cleverer while remaining an unwilling subject. The distance between those experiences supplied a product question: could software generate the footage and spare its author repeated performances?

“At HeyGen, we believe that the camera is replaceable.”

Joshua Xu

That is an unusual ambition for someone whose career developed around a camera company. It also has an engineer’s economy. When a step causes persistent trouble, examine whether the step belongs in the process. Xu was considering the recording itself. A person might have the lesson, the script or the explanation ready long before they had the time and inclination to deliver it into a lens.

His position does not require the camera to disappear from everyone else’s life. It offers a way around a particular obligation. The aspiring presenter need not acquire a new personality along with a microphone. Xu’s own reluctance gives that argument its human scale. He knew the task from the inside, both as the builder of the software and as the person expected to test it.

The university connection

Xu’s education was in Carnegie Mellon’s Robotic Systems Development program, where he specialized in computer vision. The university also brought him into contact with Wayne Liang, who studied human-computer interaction. After months of discussion, the two founded HeyGen in December 2020. Engineering and design gave the partnership complementary starting points: how to make the technology work, and how to make working with it feel manageable.

There is a useful continuity here. Computer vision studies what can be understood or done with images. Interface design asks what a person can do with a system. A video product lives at their meeting point. Convincing footage has limited value if making it requires a production ritual as demanding as the one it was supposed to simplify.

For Xu, the university connection extended beyond the co-founder. He has credited Carnegie Mellon with helping bring early team members into HeyGen. The personal network became part of the working company. That is a less theatrical founding scene than the solitary inventor arriving at a complete idea. The company emerged through conversations, complementary skills and people whose work the founders had reason to trust.

A five-dollar audition

The first customer in Xu’s early account paid $5. The order came through Fiverr, where he tested demand for spokesperson videos and manually delivered the results. It is a wonderfully modest beginning for an enterprise concerned with replacing cameras. Before the subscription business could grow, one buyer had to decide that a particular video was worth a small payment.

The experiment had an awkward detail, too. Xu says the original listing did not disclose the use of AI avatars; the founders later changed it to do so, and reported that feedback remained positive. That sequence belongs in the story alongside the successful order. It marks the difference between establishing demand for a deliverable and understanding how customers respond to the method used to make it.

A five-dollar sale cannot establish the prospects of an entire company. It can establish that somebody has a task worth paying to complete. For a technical founder, that is valuable friction. Admiration for a model is cheap and easy to offer. A buyer asks for something specific, expects it to arrive and can decide whether to return.

Xu’s early growth account dates the product launch to July 29, 2022, and says it took 178 days to reach $1 million in annual recurring revenue. The chronology puts the paying customer before the headline metric. He had to learn how people wanted to use the system, then build a product through which more of them could do it.

His face, his product

Xu’s avatar made that learning process unusually visible. A founder demonstrating a spreadsheet can remain beside the product. A founder demonstrating a digital likeness becomes part of what the customer examines. The face, the timing and the delivery all invite judgment. An avatar explaining the service is also quietly asking whether the viewer would be willing to use an avatar of their own.

In July 2025, Xu described continuing to work closely with the team on Avatar IV updates and talking directly with customers as the company grew. The engineering background remained part of his account of being CEO. He wanted to understand how the tool was used and what needed improving. A title could expand his responsibilities without removing him from the product.

That connection helps explain why likeness is such a demanding product category. A small imperfection can become conspicuous when it occurs in a face. Viewers bring years of experience reading human expressions to every clip. The builder is competing with expectations formed before the viewer has heard of the company.

Xu also has an appetite for studying businesses beyond his own. Asked for a favorite follow, he named the Acquired podcast and said he never misses an episode. It is an agreeable detail: the engineer trying to change video spends some of his attention listening to people discuss how companies came to be. The habit fits a career that has required learning the business around the technology.

Permission belongs in the picture

A likeness raises a question that a generic animation can avoid: who agreed to this? The closer a synthetic presenter comes to a recognizable person, the more consequential that question becomes. A product that can carry someone’s face and voice into new footage also needs a credible way to keep that person involved in the decision.

“Earning user trust is more important than short-term growth.”

Joshua Xu

Xu has described verification steps as part of earning that trust. They introduce friction into a service whose attraction is the removal of friction elsewhere. It is a sensible complication. The customer wants less work in producing a video and more confidence about control of a likeness. Those preferences can coexist, even when a signup chart would prefer fewer steps.

His stated approach makes trust a product concern rather than a pleasant word surrounding the product. For a business customer, a believable avatar is useful only if using it is acceptable. The permission, the presenter and the purpose arrive in the same purchasing decision. Xu’s own avatar gives this issue a particularly direct form: he is also one of the people whose identity the system carries.

The arithmetic of a quiet idea

In June 2024, HeyGen announced a $60 million Series A led by Benchmark. By June 25, 2026, the company reported passing $200 million in annual recurring revenue, having doubled that measure in eight months. It also reported more than 30 million users across 196 countries. These are company figures, describing a business built around an idea Xu initially understood at a personal level.

Xu also told an interviewer that HeyGen had remained cash flow break-even during 2026 to that point. The account described a company attentive to team size and to owning its technology stack. Growth becomes more interesting when paired with those operating choices. The business must support the work of improving the product, not merely produce a larger announcement about itself.

The original problem survives the multiplication. More users do not change the basic wish to communicate without arranging another shoot. A course instructor, a small business owner or a colleague explaining a process can recognize that wish without caring about a financing round. The company’s scale and the customer’s task belong to different units of measurement.

After the camera, the editing desk

Xu has described AI-assisted editing as a next step: a person explains the desired result and software helps produce the edited video. It follows the same line of inquiry as the camera idea. Generating footage removes one task. Turning footage into a finished communication introduces others. The ambition is to make more of that process accessible through a description of what the user wants.

There is still an author in this picture. Someone chooses what to say, what to keep and whether the result represents them. Xu’s project changes the labor between intention and publication. Its usefulness will depend on how much control survives that shorter journey.

For now, the memorable image is a founder whose digital face explains a service built partly from his dislike of filming his actual one. The joke is affectionate, and the problem is familiar. Joshua Xu wants people with something to say to spend less of their day preparing to say it. Even a camera enthusiast can appreciate getting the afternoon back.